Merge remote-tracking branch 'origin/master' into rag-evaluation
This commit is contained in:
Chris Coutinho
@@ -1,3 +1,9 @@
|
||||
## v0.35.0 (2025-11-15)
|
||||
|
||||
### Feat
|
||||
|
||||
- Enable SSE transport for mcp service and update test fixtures
|
||||
|
||||
## v0.34.2 (2025-11-13)
|
||||
|
||||
### Fix
|
||||
|
||||
@@ -2,8 +2,8 @@ apiVersion: v2
|
||||
name: nextcloud-mcp-server
|
||||
description: A Helm chart for Nextcloud MCP Server - enables AI assistants to interact with Nextcloud
|
||||
type: application
|
||||
version: 0.34.2
|
||||
appVersion: "0.34.2"
|
||||
version: 0.35.0
|
||||
appVersion: "0.35.0"
|
||||
keywords:
|
||||
- nextcloud
|
||||
- mcp
|
||||
|
||||
+2
-2
@@ -34,7 +34,7 @@ services:
|
||||
- ./app-hooks:/docker-entrypoint-hooks.d:ro
|
||||
# Mount OIDC development directory outside /var/www/html to avoid rsync conflicts
|
||||
# The post-installation hook will register /opt/apps as an additional app directory
|
||||
- ./third_party:/opt/apps:ro
|
||||
#- ./third_party:/opt/apps:ro
|
||||
environment:
|
||||
- NEXTCLOUD_TRUSTED_DOMAINS=app
|
||||
- NEXTCLOUD_ADMIN_USER=admin
|
||||
@@ -69,7 +69,6 @@ services:
|
||||
|
||||
mcp:
|
||||
build: .
|
||||
command: ["--transport", "streamable-http"]
|
||||
restart: always
|
||||
depends_on:
|
||||
app:
|
||||
@@ -82,6 +81,7 @@ services:
|
||||
- NEXTCLOUD_HOST=http://app:80
|
||||
- NEXTCLOUD_USERNAME=admin
|
||||
- NEXTCLOUD_PASSWORD=admin
|
||||
- NEXTCLOUD_PUBLIC_ISSUER_URL=http://localhost:8080
|
||||
|
||||
# Vector sync configuration (ADR-007)
|
||||
- VECTOR_SYNC_ENABLED=true
|
||||
|
||||
@@ -0,0 +1,619 @@
|
||||
# ADR-012: Unified Multi-Algorithm Search with Client-Configurable Weighting
|
||||
|
||||
## Status
|
||||
Proposed
|
||||
|
||||
## Context
|
||||
|
||||
### Current State
|
||||
|
||||
The Nextcloud MCP server currently provides semantic search via vector similarity (Qdrant), as designed in ADR-003 and implemented through ADR-007. However, users and MCP clients have limited control over search behavior:
|
||||
|
||||
1. **Single algorithm only**: Only pure vector similarity search is available
|
||||
2. **No algorithm selection**: MCP clients cannot choose between semantic, keyword, or fuzzy approaches
|
||||
3. **No weighting control**: Clients cannot adjust the balance between different search methods
|
||||
4. **Disconnected implementations**: Viz pane uses different search algorithms than MCP tools
|
||||
5. **Limited flexibility**: No way to optimize search for different use cases (exact match vs. conceptual similarity)
|
||||
|
||||
### User Needs
|
||||
|
||||
Different search scenarios require different algorithms:
|
||||
|
||||
- **Exact match queries**: "Find note titled 'Q1 Budget'" → keyword search preferred
|
||||
- **Conceptual queries**: "What are my goals for next quarter?" → semantic search preferred
|
||||
- **Typo-tolerant queries**: "Find note about kuberntes" → fuzzy search needed
|
||||
- **Balanced queries**: "Find documentation about API endpoints" → hybrid search optimal
|
||||
|
||||
Additionally, users need a **testing interface** (viz pane) to:
|
||||
- Experiment with different search algorithms on their own documents
|
||||
- Visualize search results and algorithm behavior
|
||||
- Tune weights for optimal results
|
||||
- Understand which algorithm works best for their queries
|
||||
|
||||
### Technical Requirements
|
||||
|
||||
1. **Unified interface**: Single MCP tool supporting multiple algorithms
|
||||
2. **Client control**: MCP clients specify algorithm and weights via tool parameters
|
||||
3. **Backward compatibility**: Existing `nc_semantic_search()` behavior preserved
|
||||
4. **Shared implementation**: Viz pane and MCP tools use identical search algorithms
|
||||
5. **User accessibility**: Viz pane available to all logged-in users with vector sync enabled
|
||||
6. **Performance**: Minimal overhead for algorithm selection
|
||||
|
||||
## Decision
|
||||
|
||||
We will implement a **unified multi-algorithm search architecture** with the following components:
|
||||
|
||||
### Architecture Diagram
|
||||
|
||||
```
|
||||
┌─────────────────────────────────────────────────────────────────────────────┐
|
||||
│ MCP Client / User Browser │
|
||||
│ │
|
||||
│ ┌──────────────────────────┐ ┌──────────────────────────────────┐ │
|
||||
│ │ MCP Tool Call │ │ Viz Pane (Browser UI) │ │
|
||||
│ │ │ │ │ │
|
||||
│ │ nc_semantic_search( │ │ - Algorithm selector dropdown │ │
|
||||
│ │ query="kubernetes", │ │ - Weight adjustment sliders │ │
|
||||
│ │ algorithm="hybrid", │ │ - Interactive 2D scatter plot │ │
|
||||
│ │ semantic_weight=0.5, │ │ - Side-by-side comparison │ │
|
||||
│ │ keyword_weight=0.3, │ │ - Real-time search testing │ │
|
||||
│ │ fuzzy_weight=0.2 │ │ │ │
|
||||
│ │ ) │ │ │ │
|
||||
│ └───────────┬──────────────┘ └────────────┬─────────────────────┘ │
|
||||
└──────────────┼─────────────────────────────────────┼────────────────────────┘
|
||||
│ │
|
||||
│ MCP Protocol │ HTTPS (htmx)
|
||||
│ │
|
||||
┌──────────────▼──────────────────────────────────────▼────────────────────────┐
|
||||
│ MCP Server (/app endpoint) │
|
||||
│ │
|
||||
│ ┌─────────────────────────────────────────────────────────────────────────┐ │
|
||||
│ │ Unified Search Interface (server/semantic.py) │ │
|
||||
│ │ │ │
|
||||
│ │ @mcp.tool() nc_semantic_search(algorithm, weights...) │ │
|
||||
│ │ ├─ Validate parameters (weights sum ≤1.0) │ │
|
||||
│ │ ├─ Dispatch to algorithm selector │ │
|
||||
│ │ └─ Return ranked SearchResponse │ │
|
||||
│ └────────────────────────────┬────────────────────────────────────────────┘ │
|
||||
│ │ │
|
||||
│ ┌────────────────────────────▼────────────────────────────────────────────┐ │
|
||||
│ │ Algorithm Dispatcher (search/algorithms.py) │ │
|
||||
│ │ │ │
|
||||
│ │ if algorithm == "semantic": → semantic.py │ │
|
||||
│ │ if algorithm == "keyword": → keyword.py │ │
|
||||
│ │ if algorithm == "fuzzy": → fuzzy.py │ │
|
||||
│ │ if algorithm == "hybrid": → hybrid.py (RRF fusion) │ │
|
||||
│ └─────────────────────────────────────────────────────────────────────────┘ │
|
||||
│ │
|
||||
│ ┌──────────────────┐ ┌──────────────────┐ ┌──────────────────┐ │
|
||||
│ │ semantic.py │ │ keyword.py │ │ fuzzy.py │ │
|
||||
│ │ │ │ │ │ │ │
|
||||
│ │ • Query Qdrant │ │ • Token matching │ │ • Char overlap │ │
|
||||
│ │ • Cosine dist │ │ • Title weight │ │ • 70% threshold │ │
|
||||
│ │ • Score ≥0.7 │ │ • ADR-001 logic │ │ • Simple impl │ │
|
||||
│ └────────┬─────────┘ └────────┬─────────┘ └────────┬─────────┘ │
|
||||
│ │ │ │ │
|
||||
│ └─────────────────────┼──────────────────────┘ │
|
||||
│ │ │
|
||||
│ ┌──────────────────────────────▼──────────────────────────────────────────┐ │
|
||||
│ │ hybrid.py (Reciprocal Rank Fusion) │ │
|
||||
│ │ │ │
|
||||
│ │ 1. Run algorithms in parallel (semantic, keyword, fuzzy) │ │
|
||||
│ │ 2. Collect ranked results from each │ │
|
||||
│ │ 3. Apply RRF formula: score = weight / (k + rank) │ │
|
||||
│ │ 4. Combine scores across algorithms │ │
|
||||
│ │ 5. Re-rank by combined score │ │
|
||||
│ └─────────────────────────────────────────────────────────────────────────┘ │
|
||||
└───────────────────────────────────┬───────────────────────────────────────────┘
|
||||
│
|
||||
┌───────────────┴───────────────┐
|
||||
│ │
|
||||
┌──────────▼──────────┐ ┌─────────▼────────────┐
|
||||
│ Qdrant Vector DB │ │ Nextcloud APIs │
|
||||
│ │ │ │
|
||||
│ • Vector search │ │ • Access verification│
|
||||
│ • user_id filter │ │ • Full metadata fetch│
|
||||
│ • Score threshold │ │ • Permission checks │
|
||||
│ • 768-dim embeddings│ │ │
|
||||
└─────────────────────┘ └──────────────────────┘
|
||||
```
|
||||
|
||||
### Data Flow
|
||||
|
||||
#### MCP Tool Request
|
||||
```
|
||||
1. Client calls nc_semantic_search(query, algorithm="hybrid", weights...)
|
||||
2. Server validates parameters (weights sum ≤1.0)
|
||||
3. Dispatcher routes to hybrid.py
|
||||
4. Hybrid search runs semantic, keyword, fuzzy in parallel
|
||||
5. RRF combines results with weighted scores
|
||||
6. Access verification via Nextcloud API
|
||||
7. Return ranked SearchResponse to client
|
||||
```
|
||||
|
||||
#### Viz Pane Request (Server-Side Processing)
|
||||
```
|
||||
1. User navigates to /app (Vector Visualization tab)
|
||||
2. Browser loads vector-viz fragment via htmx
|
||||
3. User enters query and adjusts algorithm/weights
|
||||
4. htmx sends request to /app/vector-viz endpoint
|
||||
5. Server executes search via search/algorithms.py:
|
||||
- Filters by user_id (multi-tenant security)
|
||||
- Applies selected algorithm (semantic/keyword/fuzzy/hybrid)
|
||||
- Filters by document type (notes/files/calendar/contacts)
|
||||
- Retrieves matching results + metadata
|
||||
6. Server performs PCA reduction (768-dim → 2D):
|
||||
- Converts matching results to 2D coordinates
|
||||
- Only sends coordinates + metadata (not full vectors)
|
||||
- Dramatically reduces bandwidth (e.g., 768 floats → 2 floats per doc)
|
||||
7. Server returns JSON: {results: [...], coordinates_2d: [...], stats: {...}}
|
||||
8. Browser receives lightweight response
|
||||
9. Plotly.js renders interactive scatter plot
|
||||
10. Matching results highlighted (blue), non-matches grayed (40% opacity)
|
||||
```
|
||||
|
||||
**Performance Benefits of Server-Side Processing**:
|
||||
- **Bandwidth reduction**: ~384x less data (2 floats vs 768 floats per document)
|
||||
- **Client efficiency**: Browser only handles visualization, not computation
|
||||
- **Scalability**: Can visualize 10,000+ documents without client-side lag
|
||||
- **Security**: Raw vectors never leave server
|
||||
- **Consistency**: Same search logic as MCP tool (no drift)
|
||||
|
||||
### 1. Core Search Algorithms
|
||||
|
||||
Four search algorithms will be available:
|
||||
|
||||
#### a) Semantic Search (Vector Similarity)
|
||||
- **Method**: Cosine distance in 768-dimensional embedding space
|
||||
- **Implementation**: Qdrant `query_points` with user_id filtering
|
||||
- **Use case**: Conceptual queries, finding related content
|
||||
- **Current status**: Implemented in `nextcloud_mcp_server/server/semantic.py`
|
||||
|
||||
#### b) Keyword Search (Token-Based)
|
||||
- **Method**: Token matching with weighted scoring (from ADR-001)
|
||||
- **Implementation**: Title matches weighted 3x higher than content
|
||||
- **Use case**: Exact phrase matching, known titles
|
||||
- **Current status**: Designed in ADR-001, not implemented
|
||||
|
||||
#### c) Fuzzy Search (Character Overlap)
|
||||
- **Method**: Simple character-based similarity (70% threshold)
|
||||
- **Implementation**: Character set comparison (current viz pane approach)
|
||||
- **Use case**: Typo tolerance, approximate matching
|
||||
- **Current status**: Implemented in viz pane only
|
||||
|
||||
#### d) Hybrid Search (Multi-Algorithm Fusion)
|
||||
- **Method**: Reciprocal Rank Fusion (RRF) from ADR-003
|
||||
- **Implementation**: Parallel execution + score combination
|
||||
- **Use case**: Balanced queries, general-purpose search
|
||||
- **Current status**: Designed in ADR-003, not implemented
|
||||
|
||||
### 2. Unified MCP Tool Interface
|
||||
|
||||
```python
|
||||
@mcp.tool()
|
||||
@require_scopes("semantic:read")
|
||||
async def nc_semantic_search(
|
||||
query: str,
|
||||
ctx: Context,
|
||||
limit: int = 10,
|
||||
score_threshold: float = 0.7,
|
||||
algorithm: Literal["semantic", "keyword", "fuzzy", "hybrid"] = "hybrid",
|
||||
semantic_weight: float = 0.5,
|
||||
keyword_weight: float = 0.3,
|
||||
fuzzy_weight: float = 0.2,
|
||||
) -> SearchResponse:
|
||||
"""
|
||||
Search Nextcloud content using configurable algorithms.
|
||||
|
||||
Args:
|
||||
query: Natural language search query
|
||||
ctx: MCP context for authentication
|
||||
limit: Maximum results to return
|
||||
score_threshold: Minimum similarity score (semantic/hybrid only)
|
||||
algorithm: Search algorithm to use
|
||||
semantic_weight: Weight for semantic results (hybrid only, default: 0.5)
|
||||
keyword_weight: Weight for keyword results (hybrid only, default: 0.3)
|
||||
fuzzy_weight: Weight for fuzzy results (hybrid only, default: 0.2)
|
||||
|
||||
Returns:
|
||||
Ranked search results with scores and excerpts
|
||||
"""
|
||||
```
|
||||
|
||||
**Key decisions**:
|
||||
- **Single tool name**: Keep `nc_semantic_search` for backward compatibility
|
||||
- **Algorithm parameter**: Explicit selection via enum
|
||||
- **Weight parameters**: Client-configurable, only apply to hybrid mode
|
||||
- **Validation**: Weights must sum to ≤1.0, enforced server-side
|
||||
- **Defaults**: Hybrid mode with balanced weights (semantic 50%, keyword 30%, fuzzy 20%)
|
||||
|
||||
### 3. Shared Algorithm Implementation
|
||||
|
||||
Extract search algorithms into reusable module:
|
||||
|
||||
```
|
||||
nextcloud_mcp_server/
|
||||
├── search/
|
||||
│ ├── __init__.py
|
||||
│ ├── algorithms.py # Core search implementations
|
||||
│ ├── semantic.py # Vector similarity search
|
||||
│ ├── keyword.py # Token-based search (ADR-001)
|
||||
│ ├── fuzzy.py # Character overlap search
|
||||
│ └── hybrid.py # RRF fusion (ADR-003)
|
||||
└── server/
|
||||
└── semantic.py # MCP tool wrapper
|
||||
```
|
||||
|
||||
**Benefits**:
|
||||
- Viz pane and MCP tools share identical implementations
|
||||
- Testable in isolation
|
||||
- Easy to add new algorithms (e.g., BM25, neural reranking)
|
||||
- Clear separation of concerns
|
||||
|
||||
### 4. Viz Pane Integration
|
||||
|
||||
Update viz pane (`nextcloud_mcp_server/auth/userinfo_routes.py`) to:
|
||||
|
||||
1. **Use shared algorithms**: Import from `search/algorithms.py`
|
||||
2. **Server-side filtering**: All search and filtering operations happen server-side
|
||||
- Query execution via shared search backend
|
||||
- Document type filtering (notes, files, calendar, contacts)
|
||||
- User ID filtering for multi-tenant security
|
||||
- Only matching results + metadata sent to client
|
||||
- Reduces bandwidth and improves performance
|
||||
3. **PCA reduction**: Server performs dimensionality reduction (768-dim → 2D)
|
||||
- Only 2D coordinates sent to browser for visualization
|
||||
- Dramatically reduces data transfer vs sending full vectors
|
||||
- Enables visualization of large document collections
|
||||
4. **User accessibility**: Available to all users with vector sync enabled
|
||||
5. **Security**: Filter results by `user_id` (only show user's own documents)
|
||||
6. **Interactive testing**: Allow users to:
|
||||
- Select algorithm type
|
||||
- Adjust weights (hybrid mode)
|
||||
- Compare results across algorithms
|
||||
- Visualize result distribution in 2D space
|
||||
|
||||
#### Viz Pane UI Components
|
||||
|
||||
```
|
||||
┌────────────────────────────────────────────────────────────────────────┐
|
||||
│ Vector Visualization [Status] │
|
||||
├────────────────────────────────────────────────────────────────────────┤
|
||||
│ │
|
||||
│ ┌──────────────────────────────────────────────────────────────────┐ │
|
||||
│ │ Search Configuration │ │
|
||||
│ │ │ │
|
||||
│ │ Query: [_______________________________________________] [Search]│ │
|
||||
│ │ │ │
|
||||
│ │ Algorithm: [Hybrid ▼] [Semantic] [Keyword] [Fuzzy] │ │
|
||||
│ │ │ │
|
||||
│ │ Weights (Hybrid Mode): │ │
|
||||
│ │ Semantic: [========50========] 0.5 │ │
|
||||
│ │ Keyword: [======30====== ] 0.3 │ │
|
||||
│ │ Fuzzy: [====20==== ] 0.2 │ │
|
||||
│ │ │ │
|
||||
│ │ Document Types: ☑ Notes ☑ Files ☑ Calendar ☑ Contacts │ │
|
||||
│ └──────────────────────────────────────────────────────────────────┘ │
|
||||
│ │
|
||||
│ ┌──────────────────────────────────────────────────────────────────┐ │
|
||||
│ │ Vector Space Visualization (PCA 2D Projection) │ │
|
||||
│ │ │ │
|
||||
│ │ ▲ │ │
|
||||
│ │ PC2 │ ● ● ● 🔵 Matching results (full opacity) │ │
|
||||
│ │ │ ● ● ● ⚪ Non-matching results (40% opacity) │ │
|
||||
│ │ │ 🔵 ● ● │ │
|
||||
│ │ │ ● 🔵 ● Hover: Show document title + excerpt │ │
|
||||
│ │ │ ● ● 🔵 ● Click: Open document in Nextcloud │ │
|
||||
│ │ ────┼──●─🔵──●─●────► PC1 │ │
|
||||
│ │ │ ● ● ● │ │
|
||||
│ │ │ 🔵 ● ● Explained Variance: │ │
|
||||
│ │ │ ● ● ● PC1: 23.4% | PC2: 18.7% │ │
|
||||
│ │ │ ● ● │ │
|
||||
│ │ │ │
|
||||
│ └──────────────────────────────────────────────────────────────────┘ │
|
||||
│ │
|
||||
│ ┌──────────────────────────────────────────────────────────────────┐ │
|
||||
│ │ Search Results (12 matching documents) │ │
|
||||
│ │ │ │
|
||||
│ │ 🔵 Kubernetes Setup Guide Score: 0.87 │ │
|
||||
│ │ "...configure kubectl to connect to cluster..." │ │
|
||||
│ │ [Open in Nextcloud] │ │
|
||||
│ │ │ │
|
||||
│ │ 🔵 Container Orchestration Notes Score: 0.82 │ │
|
||||
│ │ "...deployment strategies for kubernetes..." │ │
|
||||
│ │ [Open in Nextcloud] │ │
|
||||
│ │ │ │
|
||||
│ │ 🔵 K8s Troubleshooting Score: 0.79 │ │
|
||||
│ │ "...common kuberntes errors and solutions..." │ │
|
||||
│ │ [Open in Nextcloud] │ │
|
||||
│ │ │ │
|
||||
│ │ [Show More Results...] │ │
|
||||
│ └──────────────────────────────────────────────────────────────────┘ │
|
||||
│ │
|
||||
│ ┌──────────────────────────────────────────────────────────────────┐ │
|
||||
│ │ Algorithm Performance Comparison │ │
|
||||
│ │ │ │
|
||||
│ │ Algorithm │ Results │ Avg Score │ Time (ms) │ Precision │ │
|
||||
│ │ ─────────────┼─────────┼───────────┼───────────┼─────────── │ │
|
||||
│ │ Semantic │ 45 │ 0.78 │ 145ms │ ████░ 0.82 │ │
|
||||
│ │ Keyword │ 23 │ 0.91 │ 42ms │ ███░░ 0.67 │ │
|
||||
│ │ Fuzzy │ 67 │ 0.72 │ 89ms │ ██░░░ 0.45 │ │
|
||||
│ │ Hybrid (RRF) │ 52 │ 0.84 │ 198ms │ █████ 0.89 │ │
|
||||
│ └──────────────────────────────────────────────────────────────────┘ │
|
||||
└────────────────────────────────────────────────────────────────────────┘
|
||||
```
|
||||
|
||||
**Key UI Features**:
|
||||
|
||||
1. **Search Input**: Real-time query testing with instant visualization
|
||||
2. **Algorithm Selector**: Dropdown + quick-select buttons
|
||||
3. **Weight Sliders**: Visual adjustment with live preview (hybrid mode only)
|
||||
4. **Document Type Filters**: Checkboxes for notes, files, calendar, contacts
|
||||
5. **2D Scatter Plot**: Interactive Plotly.js visualization
|
||||
- Blue dots = matching documents (full opacity)
|
||||
- Gray dots = non-matching documents (40% opacity)
|
||||
- Hover = show title + excerpt tooltip
|
||||
- Click = open document in Nextcloud
|
||||
- Zoom/pan controls for exploration
|
||||
6. **Results Panel**: Ranked list with scores and excerpts
|
||||
7. **Performance Table**: Compare algorithm speed and accuracy
|
||||
8. **Explained Variance**: Show how much information PCA preserves
|
||||
|
||||
**Technology Stack**:
|
||||
- **Frontend**: htmx for dynamic loading, Alpine.js for reactivity
|
||||
- **Visualization**: Plotly.js for interactive scatter plots
|
||||
- **Styling**: Tailwind CSS (consistent with existing /app UI)
|
||||
- **Backend**: Shared `search/algorithms.py` implementation
|
||||
|
||||
### 5. Reciprocal Rank Fusion (RRF) for Hybrid Search
|
||||
|
||||
Following ADR-003's design:
|
||||
|
||||
```python
|
||||
def reciprocal_rank_fusion(
|
||||
results: dict[str, list[SearchResult]],
|
||||
weights: dict[str, float],
|
||||
k: int = 60
|
||||
) -> list[SearchResult]:
|
||||
"""
|
||||
Combine multiple ranked result lists using RRF.
|
||||
|
||||
Args:
|
||||
results: Dict of algorithm_name -> ranked results
|
||||
weights: Dict of algorithm_name -> weight (0-1)
|
||||
k: RRF constant (default: 60, standard value)
|
||||
|
||||
Returns:
|
||||
Combined and re-ranked results
|
||||
"""
|
||||
scores = defaultdict(float)
|
||||
|
||||
for algo_name, algo_results in results.items():
|
||||
weight = weights.get(algo_name, 0.0)
|
||||
for rank, result in enumerate(algo_results, start=1):
|
||||
# RRF formula: 1 / (k + rank)
|
||||
rrf_score = weight / (k + rank)
|
||||
scores[result.doc_id] += rrf_score
|
||||
|
||||
# Sort by combined score, return top results
|
||||
return sorted(scores.items(), key=lambda x: x[1], reverse=True)
|
||||
```
|
||||
|
||||
**RRF properties**:
|
||||
- **Rank-based**: Uses position, not raw scores (handles score scale differences)
|
||||
- **Proven effective**: Standard approach in information retrieval
|
||||
- **Configurable**: `k` parameter controls rank decay (default: 60)
|
||||
- **Weight support**: Allows algorithm-specific importance
|
||||
|
||||
## Implementation Plan
|
||||
|
||||
### Phase 1: Extract and Unify Algorithms (Week 1)
|
||||
|
||||
1. Create `nextcloud_mcp_server/search/` module
|
||||
2. Implement `algorithms.py` with base interface
|
||||
3. Extract semantic search logic from `server/semantic.py`
|
||||
4. Implement keyword search from ADR-001 design
|
||||
5. Extract fuzzy search from viz pane
|
||||
6. Implement RRF hybrid search from ADR-003
|
||||
7. Add comprehensive unit tests for each algorithm
|
||||
|
||||
### Phase 2: Update MCP Tool (Week 1-2)
|
||||
|
||||
1. Add `algorithm` parameter to `nc_semantic_search()`
|
||||
2. Add weight parameters (`semantic_weight`, etc.)
|
||||
3. Implement algorithm dispatcher
|
||||
4. Add parameter validation (weights sum ≤1.0)
|
||||
5. Update response model to include algorithm metadata
|
||||
6. Maintain backward compatibility (default: hybrid)
|
||||
7. Add integration tests for all algorithm modes
|
||||
|
||||
### Phase 3: Update Viz Pane (Week 2)
|
||||
|
||||
**Critical: All processing must happen server-side**
|
||||
|
||||
1. **Remove client-side search filtering**
|
||||
- Delete JavaScript-based keyword/fuzzy matching
|
||||
- Remove client-side document type filtering
|
||||
- No search logic in browser
|
||||
2. **Implement server-side endpoint** (`/app/vector-viz`)
|
||||
- Accept query, algorithm, weights, doc_type filters
|
||||
- Execute search via `search/algorithms.py`
|
||||
- Filter results by user_id (security)
|
||||
- Perform PCA reduction (768-dim → 2D)
|
||||
- Return JSON with 2D coordinates + metadata only
|
||||
3. **Update frontend**
|
||||
- htmx form submission to `/app/vector-viz`
|
||||
- Algorithm selector dropdown
|
||||
- Weight adjustment sliders (htmx updates on change)
|
||||
- Document type checkboxes
|
||||
- Plotly.js visualization of server response
|
||||
4. **Performance optimization**
|
||||
- Limit results to user's documents only
|
||||
- Cache PCA transformation (invalidate on new vectors)
|
||||
- Stream large result sets if needed
|
||||
- Add loading indicators for server processing
|
||||
|
||||
### Phase 4: Documentation and Testing (Week 2-3)
|
||||
|
||||
1. Update MCP tool documentation
|
||||
2. Add algorithm selection guide
|
||||
3. Document weight tuning recommendations
|
||||
4. Add end-to-end tests (MCP + viz pane)
|
||||
5. Performance benchmarks for each algorithm
|
||||
6. Update CLAUDE.md with search patterns
|
||||
|
||||
## Consequences
|
||||
|
||||
### Positive
|
||||
|
||||
1. **Flexibility**: MCP clients can optimize search for their use case
|
||||
2. **Unified implementation**: Single source of truth for search algorithms
|
||||
3. **User empowerment**: Viz pane enables query testing and tuning
|
||||
4. **Backward compatible**: Existing semantic search behavior preserved
|
||||
5. **Extensible**: Easy to add new algorithms (BM25, neural reranking)
|
||||
6. **Testable**: Each algorithm can be unit tested independently
|
||||
7. **Standards-based**: RRF is proven in production systems
|
||||
|
||||
### Negative
|
||||
|
||||
1. **Complexity**: More parameters for clients to understand
|
||||
2. **API surface**: Larger tool signature (8 parameters)
|
||||
3. **Performance**: Hybrid search requires multiple queries
|
||||
4. **Validation overhead**: Weight validation adds processing
|
||||
5. **Documentation burden**: Need to explain when to use each algorithm
|
||||
|
||||
### Neutral
|
||||
|
||||
1. **Weight defaults**: May need tuning based on user feedback
|
||||
2. **Algorithm performance**: Will vary by content type and query
|
||||
3. **Viz pane adoption**: Unknown if users will utilize testing interface
|
||||
|
||||
## Alternatives Considered
|
||||
|
||||
### Alternative 1: Separate Tools Per Algorithm
|
||||
|
||||
```python
|
||||
@mcp.tool()
|
||||
async def nc_semantic_search(query: str, ctx: Context, ...) -> SearchResponse:
|
||||
"""Pure vector similarity search."""
|
||||
|
||||
@mcp.tool()
|
||||
async def nc_keyword_search(query: str, ctx: Context, ...) -> SearchResponse:
|
||||
"""Pure keyword matching."""
|
||||
|
||||
@mcp.tool()
|
||||
async def nc_hybrid_search(query: str, ctx: Context, weights: dict, ...) -> SearchResponse:
|
||||
"""Hybrid search with weights."""
|
||||
```
|
||||
|
||||
**Rejected because**:
|
||||
- API proliferation (3+ tools instead of 1)
|
||||
- Harder to discover capabilities
|
||||
- Backward compatibility issues
|
||||
- DRY violation (repeated parameters)
|
||||
|
||||
### Alternative 2: Server-Wide Configuration Only
|
||||
|
||||
```python
|
||||
# .env configuration
|
||||
SEARCH_ALGORITHM=hybrid
|
||||
SEMANTIC_WEIGHT=0.5
|
||||
KEYWORD_WEIGHT=0.3
|
||||
FUZZY_WEIGHT=0.2
|
||||
```
|
||||
|
||||
**Rejected because**:
|
||||
- No per-query flexibility
|
||||
- MCP clients cannot optimize for different tasks
|
||||
- Requires server restart for changes
|
||||
- User's requirement: "expose a way for users to override the default weights"
|
||||
|
||||
### Alternative 3: Production-Grade Fuzzy (Levenshtein/RapidFuzz)
|
||||
|
||||
**Rejected because**:
|
||||
- Adds external dependency
|
||||
- Simple character overlap performs adequately
|
||||
- Can always upgrade later if needed
|
||||
- User's preference: "Keep simple character overlap"
|
||||
|
||||
## Related ADRs
|
||||
|
||||
- **ADR-001**: Enhanced Note Search (keyword algorithm design)
|
||||
- **ADR-003**: Vector Database and Semantic Search (hybrid search + RRF design)
|
||||
- **ADR-007**: Background Vector Sync (semantic search implementation)
|
||||
- **ADR-008**: MCP Sampling for RAG (uses semantic search results)
|
||||
- **ADR-009**: Semantic Search OAuth Scope (security model)
|
||||
- **ADR-011**: Improving Semantic Search Quality (mentions future "ADR-013" for hybrid search)
|
||||
|
||||
**This ADR supersedes**:
|
||||
- ADR-011's placeholder for "ADR-013: Hybrid Search"
|
||||
|
||||
**This ADR implements**:
|
||||
- ADR-003's hybrid search design (previously unimplemented)
|
||||
- ADR-001's keyword search design (previously unimplemented)
|
||||
|
||||
## References
|
||||
|
||||
- **Reciprocal Rank Fusion**: Cormack, G. V., Clarke, C. L., & Buettcher, S. (2009). "Reciprocal rank fusion outperforms condorcet and individual rank learning methods." SIGIR '09.
|
||||
- **Vector Search**: Malkov, Y. A., & Yashunin, D. A. (2018). "Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs." TPAMI.
|
||||
- **Hybrid Search Best Practices**: Qdrant documentation on hybrid search patterns
|
||||
- **MCP Protocol**: Model Context Protocol specification for tool design
|
||||
|
||||
## Implementation Notes
|
||||
|
||||
### Weight Validation
|
||||
|
||||
```python
|
||||
def validate_weights(
|
||||
semantic_weight: float,
|
||||
keyword_weight: float,
|
||||
fuzzy_weight: float
|
||||
) -> None:
|
||||
"""Validate hybrid search weights."""
|
||||
if semantic_weight < 0 or keyword_weight < 0 or fuzzy_weight < 0:
|
||||
raise ValueError("Weights must be non-negative")
|
||||
|
||||
total = semantic_weight + keyword_weight + fuzzy_weight
|
||||
if total > 1.0:
|
||||
raise ValueError(f"Weights sum to {total:.2f}, must be ≤1.0")
|
||||
|
||||
if total == 0.0:
|
||||
raise ValueError("At least one weight must be > 0")
|
||||
```
|
||||
|
||||
### Backward Compatibility
|
||||
|
||||
The default behavior (`algorithm="hybrid"` with balanced weights) provides better results than current pure semantic search, while maintaining the same tool name and signature structure. Existing clients will automatically benefit from hybrid search without code changes.
|
||||
|
||||
### Performance Considerations
|
||||
|
||||
- **Semantic search**: ~50-200ms (vector DB query)
|
||||
- **Keyword search**: ~10-50ms (in-memory token matching)
|
||||
- **Fuzzy search**: ~20-100ms (character comparison)
|
||||
- **Hybrid search**: ~100-300ms (parallel execution + fusion)
|
||||
|
||||
Parallel execution of algorithms minimizes hybrid search latency.
|
||||
|
||||
### Security Model
|
||||
|
||||
All algorithms respect the same security boundaries:
|
||||
1. **User filtering**: Qdrant queries filter by `user_id`
|
||||
2. **Access verification**: Results verified via Nextcloud API
|
||||
3. **OAuth scope**: `semantic:read` required for all algorithms
|
||||
4. **Viz pane**: Shows only current user's documents
|
||||
|
||||
## Success Metrics
|
||||
|
||||
1. **Adoption**: % of MCP clients using algorithm parameter
|
||||
2. **Performance**: Search latency percentiles (p50, p95, p99)
|
||||
3. **Quality**: User satisfaction with result relevance
|
||||
4. **Viz pane usage**: % of users accessing testing interface
|
||||
5. **Weight distribution**: Most common weight configurations
|
||||
|
||||
## Future Enhancements
|
||||
|
||||
1. **Additional algorithms**: BM25, TF-IDF, neural reranking
|
||||
2. **Auto-tuning**: Learn optimal weights per user
|
||||
3. **Query analysis**: Automatic algorithm selection based on query
|
||||
4. **Cross-app search**: Extend beyond notes to calendar, files, etc.
|
||||
5. **Feedback loop**: Use click-through rate to improve weights
|
||||
@@ -1477,6 +1477,10 @@ def get_app(transport: str = "sse", enabled_apps: list[str] | None = None):
|
||||
user_info_html,
|
||||
vector_sync_status_fragment,
|
||||
)
|
||||
from nextcloud_mcp_server.auth.viz_routes import (
|
||||
vector_visualization_html,
|
||||
vector_visualization_search,
|
||||
)
|
||||
from nextcloud_mcp_server.auth.webhook_routes import (
|
||||
disable_webhook_preset,
|
||||
enable_webhook_preset,
|
||||
@@ -1496,6 +1500,15 @@ def get_app(transport: str = "sse", enabled_apps: list[str] | None = None):
|
||||
vector_sync_status_fragment,
|
||||
methods=["GET"],
|
||||
), # /app/vector-sync/status
|
||||
# Vector visualization routes
|
||||
Route(
|
||||
"/vector-viz", vector_visualization_html, methods=["GET"]
|
||||
), # /app/vector-viz
|
||||
Route(
|
||||
"/vector-viz/search",
|
||||
vector_visualization_search,
|
||||
methods=["GET"],
|
||||
), # /app/vector-viz/search
|
||||
# Webhook management routes (admin-only)
|
||||
Route("/webhooks", webhook_management_pane, methods=["GET"]), # /app/webhooks
|
||||
Route(
|
||||
|
||||
@@ -489,6 +489,16 @@ async def user_info_html(request: Request) -> HTMLResponse:
|
||||
str(request.url_for("oauth_logout")) if oauth_ctx else "/oauth/logout"
|
||||
)
|
||||
|
||||
# Get Nextcloud host for generating links to apps (used by viz tab)
|
||||
# Use public issuer URL if available (for browser-accessible links),
|
||||
# otherwise fall back to NEXTCLOUD_HOST from settings
|
||||
from nextcloud_mcp_server.config import get_settings
|
||||
|
||||
settings = get_settings()
|
||||
nextcloud_host_for_links = (
|
||||
os.getenv("NEXTCLOUD_PUBLIC_ISSUER_URL") or settings.nextcloud_host
|
||||
)
|
||||
|
||||
# Build host info HTML (BasicAuth only)
|
||||
host_info_html = ""
|
||||
if auth_mode == "basic":
|
||||
@@ -658,6 +668,115 @@ async def user_info_html(request: Request) -> HTMLResponse:
|
||||
<!-- Alpine.js for tab state management -->
|
||||
<script defer src="https://cdn.jsdelivr.net/npm/alpinejs@3.x.x/dist/cdn.min.js"></script>
|
||||
|
||||
<!-- Plotly.js for vector visualization -->
|
||||
<script src="https://cdn.plot.ly/plotly-2.27.0.min.js"></script>
|
||||
|
||||
<!-- Vector visualization app (Alpine.js component) -->
|
||||
<script>
|
||||
function vizApp() {{
|
||||
return {{
|
||||
query: '',
|
||||
algorithm: 'hybrid',
|
||||
showAdvanced: false,
|
||||
docTypes: [''], // Default to "All Types"
|
||||
limit: 50,
|
||||
scoreThreshold: 0.7,
|
||||
semanticWeight: 0.5,
|
||||
keywordWeight: 0.3,
|
||||
fuzzyWeight: 0.2,
|
||||
loading: false,
|
||||
results: [],
|
||||
|
||||
async executeSearch() {{
|
||||
this.loading = true;
|
||||
this.results = [];
|
||||
|
||||
try {{
|
||||
const params = new URLSearchParams({{
|
||||
query: this.query,
|
||||
algorithm: this.algorithm,
|
||||
limit: this.limit,
|
||||
score_threshold: this.scoreThreshold,
|
||||
semantic_weight: this.semanticWeight,
|
||||
keyword_weight: this.keywordWeight,
|
||||
fuzzy_weight: this.fuzzyWeight,
|
||||
}});
|
||||
|
||||
// Add doc_types parameter (filter out empty string for "All Types")
|
||||
const selectedTypes = this.docTypes.filter(t => t !== '');
|
||||
if (selectedTypes.length > 0) {{
|
||||
params.append('doc_types', selectedTypes.join(','));
|
||||
}}
|
||||
|
||||
const response = await fetch(`/app/vector-viz/search?${{params}}`);
|
||||
const data = await response.json();
|
||||
|
||||
if (data.success) {{
|
||||
this.results = data.results;
|
||||
this.renderPlot(data.coordinates_2d, data.results);
|
||||
}} else {{
|
||||
alert('Search failed: ' + data.error);
|
||||
}}
|
||||
}} catch (error) {{
|
||||
alert('Error: ' + error.message);
|
||||
}} finally {{
|
||||
this.loading = false;
|
||||
}}
|
||||
}},
|
||||
|
||||
renderPlot(coordinates, results) {{
|
||||
const trace = {{
|
||||
x: coordinates.map(c => c[0]),
|
||||
y: coordinates.map(c => c[1]),
|
||||
mode: 'markers',
|
||||
type: 'scatter',
|
||||
text: results.map(r => `${{r.title}}<br>Score: ${{r.score.toFixed(3)}}`),
|
||||
marker: {{
|
||||
size: 8,
|
||||
color: results.map(r => r.score),
|
||||
colorscale: 'Viridis',
|
||||
showscale: true,
|
||||
colorbar: {{ title: 'Score' }},
|
||||
cmin: 0,
|
||||
cmax: 1
|
||||
}}
|
||||
}};
|
||||
|
||||
const layout = {{
|
||||
title: `Vector Space (PCA 2D) - ${{results.length}} results`,
|
||||
xaxis: {{ title: 'PC1' }},
|
||||
yaxis: {{ title: 'PC2' }},
|
||||
hovermode: 'closest',
|
||||
height: 600
|
||||
}};
|
||||
|
||||
Plotly.newPlot('viz-plot', [trace], layout);
|
||||
}},
|
||||
|
||||
getNextcloudUrl(result) {{
|
||||
// Generate Nextcloud URL based on document type
|
||||
// Use the actual Nextcloud host (port 8080), not the MCP server
|
||||
const baseUrl = '{nextcloud_host_for_links}';
|
||||
|
||||
switch (result.doc_type) {{
|
||||
case 'note':
|
||||
return `${{baseUrl}}/apps/notes/note/${{result.id}}`;
|
||||
case 'file':
|
||||
return `${{baseUrl}}/apps/files/?fileId=${{result.id}}`;
|
||||
case 'calendar':
|
||||
return `${{baseUrl}}/apps/calendar`;
|
||||
case 'contact':
|
||||
return `${{baseUrl}}/apps/contacts`;
|
||||
case 'deck':
|
||||
return `${{baseUrl}}/apps/deck`;
|
||||
default:
|
||||
return `${{baseUrl}}`;
|
||||
}}
|
||||
}}
|
||||
}}
|
||||
}}
|
||||
</script>
|
||||
|
||||
<style>
|
||||
body {{
|
||||
font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, "Helvetica Neue", Arial, sans-serif;
|
||||
@@ -846,6 +965,18 @@ async def user_info_html(request: Request) -> HTMLResponse:
|
||||
Vector Sync
|
||||
</button>
|
||||
'''
|
||||
}
|
||||
{
|
||||
""
|
||||
if not show_vector_sync_tab
|
||||
else '''
|
||||
<button
|
||||
class="tab"
|
||||
:class="activeTab === 'vector-viz' ? 'active' : ''"
|
||||
@click="activeTab = 'vector-viz'">
|
||||
Vector Viz
|
||||
</button>
|
||||
'''
|
||||
}
|
||||
{
|
||||
""
|
||||
@@ -881,6 +1012,19 @@ async def user_info_html(request: Request) -> HTMLResponse:
|
||||
|
||||
{
|
||||
""
|
||||
if not show_vector_sync_tab
|
||||
else '''
|
||||
<!-- Vector Viz Tab -->
|
||||
<div class="tab-pane" x-show="activeTab === 'vector-viz'" x-transition.opacity.duration.150ms>
|
||||
<div hx-get="/app/vector-viz" hx-trigger="load" hx-swap="outerHTML">
|
||||
<p style="color: #999;">Loading vector visualization...</p>
|
||||
</div>
|
||||
</div>
|
||||
'''
|
||||
}
|
||||
|
||||
{
|
||||
""
|
||||
if not show_webhooks_tab
|
||||
else f'''
|
||||
<!-- Webhooks Tab (admin-only, loaded dynamically) -->
|
||||
|
||||
@@ -0,0 +1,610 @@
|
||||
"""Vector visualization routes for testing search algorithms.
|
||||
|
||||
Provides a web UI for users to test different search algorithms on their own
|
||||
indexed documents and visualize results in 2D space using PCA.
|
||||
|
||||
All processing happens server-side following ADR-012:
|
||||
- Search execution via shared search/algorithms.py
|
||||
- PCA dimensionality reduction (768-dim → 2D)
|
||||
- Only 2D coordinates + metadata sent to client
|
||||
- Bandwidth-efficient (2 floats per doc vs 768)
|
||||
"""
|
||||
|
||||
import logging
|
||||
|
||||
import numpy as np
|
||||
from starlette.authentication import requires
|
||||
from starlette.requests import Request
|
||||
from starlette.responses import HTMLResponse, JSONResponse
|
||||
|
||||
from nextcloud_mcp_server.config import get_settings
|
||||
from nextcloud_mcp_server.search import (
|
||||
FuzzySearchAlgorithm,
|
||||
HybridSearchAlgorithm,
|
||||
KeywordSearchAlgorithm,
|
||||
SemanticSearchAlgorithm,
|
||||
)
|
||||
from nextcloud_mcp_server.vector.pca import PCA
|
||||
from nextcloud_mcp_server.vector.qdrant_client import get_qdrant_client
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
@requires("authenticated", redirect="oauth_login")
|
||||
async def vector_visualization_html(request: Request) -> HTMLResponse:
|
||||
"""Vector visualization page with search controls and interactive plot.
|
||||
|
||||
Provides UI for testing search algorithms with real-time visualization.
|
||||
Requires vector sync to be enabled.
|
||||
|
||||
Args:
|
||||
request: Starlette request object
|
||||
|
||||
Returns:
|
||||
HTML page with search interface
|
||||
"""
|
||||
settings = get_settings()
|
||||
|
||||
if not settings.vector_sync_enabled:
|
||||
return HTMLResponse(
|
||||
"""
|
||||
<div>
|
||||
<h2>Vector Visualization</h2>
|
||||
<div style="padding: 20px; background: #fff3cd; border: 1px solid #ffc107; border-radius: 4px;">
|
||||
Vector sync is not enabled. Set VECTOR_SYNC_ENABLED=true to use this feature.
|
||||
</div>
|
||||
</div>
|
||||
"""
|
||||
)
|
||||
|
||||
# Get user info from auth context
|
||||
username = (
|
||||
request.user.display_name
|
||||
if hasattr(request.user, "display_name")
|
||||
else "unknown"
|
||||
)
|
||||
|
||||
html_content = f"""
|
||||
<style>
|
||||
.viz-card {{
|
||||
background: white;
|
||||
border-radius: 8px;
|
||||
padding: 20px;
|
||||
margin-bottom: 20px;
|
||||
box-shadow: 0 2px 4px rgba(0,0,0,0.1);
|
||||
}}
|
||||
.viz-controls {{
|
||||
margin-bottom: 20px;
|
||||
}}
|
||||
.viz-control-row {{
|
||||
display: grid;
|
||||
grid-template-columns: 2fr 1fr auto;
|
||||
gap: 12px;
|
||||
margin-bottom: 12px;
|
||||
align-items: end;
|
||||
}}
|
||||
.viz-control-group {{
|
||||
margin-bottom: 15px;
|
||||
}}
|
||||
.viz-control-group label {{
|
||||
display: block;
|
||||
margin-bottom: 5px;
|
||||
font-weight: 500;
|
||||
color: #333;
|
||||
}}
|
||||
.viz-control-group input[type="text"],
|
||||
.viz-control-group input[type="number"],
|
||||
.viz-control-group select {{
|
||||
width: 100%;
|
||||
padding: 8px 12px;
|
||||
border: 1px solid #ddd;
|
||||
border-radius: 4px;
|
||||
font-size: 14px;
|
||||
}}
|
||||
.viz-control-group input[type="range"] {{
|
||||
width: 100%;
|
||||
}}
|
||||
.viz-control-group select[multiple] {{
|
||||
min-height: 100px;
|
||||
}}
|
||||
.viz-weight-display {{
|
||||
display: inline-block;
|
||||
min-width: 40px;
|
||||
text-align: right;
|
||||
color: #666;
|
||||
}}
|
||||
.viz-btn {{
|
||||
background: #0066cc;
|
||||
color: white;
|
||||
border: none;
|
||||
padding: 10px 20px;
|
||||
border-radius: 4px;
|
||||
cursor: pointer;
|
||||
font-size: 14px;
|
||||
font-weight: 500;
|
||||
}}
|
||||
.viz-btn:hover {{
|
||||
background: #0052a3;
|
||||
}}
|
||||
.viz-btn-secondary {{
|
||||
background: #6c757d;
|
||||
color: white;
|
||||
border: none;
|
||||
padding: 6px 12px;
|
||||
border-radius: 4px;
|
||||
cursor: pointer;
|
||||
font-size: 13px;
|
||||
margin-bottom: 12px;
|
||||
}}
|
||||
.viz-btn-secondary:hover {{
|
||||
background: #5a6268;
|
||||
}}
|
||||
#viz-plot-container {{
|
||||
width: 100%;
|
||||
height: 600px;
|
||||
position: relative;
|
||||
}}
|
||||
#viz-plot {{
|
||||
width: 100%;
|
||||
height: 100%;
|
||||
}}
|
||||
.viz-loading {{
|
||||
text-align: center;
|
||||
padding: 40px;
|
||||
color: #666;
|
||||
}}
|
||||
.viz-loading-overlay {{
|
||||
position: absolute;
|
||||
inset: 0;
|
||||
display: flex;
|
||||
align-items: center;
|
||||
justify-content: center;
|
||||
background: white;
|
||||
color: #666;
|
||||
}}
|
||||
.viz-no-results {{
|
||||
text-align: center;
|
||||
padding: 40px;
|
||||
color: #666;
|
||||
font-style: italic;
|
||||
}}
|
||||
.viz-advanced-section {{
|
||||
margin-top: 16px;
|
||||
padding: 16px;
|
||||
background: #f8f9fa;
|
||||
border-radius: 4px;
|
||||
border: 1px solid #dee2e6;
|
||||
}}
|
||||
.viz-advanced-grid {{
|
||||
display: grid;
|
||||
grid-template-columns: 1fr 1fr;
|
||||
gap: 20px;
|
||||
}}
|
||||
.viz-info-box {{
|
||||
background: #e3f2fd;
|
||||
border-left: 4px solid #2196f3;
|
||||
padding: 12px;
|
||||
margin-bottom: 20px;
|
||||
font-size: 14px;
|
||||
}}
|
||||
</style>
|
||||
|
||||
<div x-data="vizApp()">
|
||||
<div class="viz-card">
|
||||
<h2>Vector Visualization</h2>
|
||||
<div class="viz-info-box">
|
||||
Testing search algorithms on your indexed documents. User: <strong>{username}</strong>
|
||||
</div>
|
||||
|
||||
<form @submit.prevent="executeSearch">
|
||||
<div class="viz-controls">
|
||||
<!-- Main Controls -->
|
||||
<div class="viz-control-group">
|
||||
<label>Search Query</label>
|
||||
<input type="text" x-model="query" placeholder="Enter search query..." required />
|
||||
</div>
|
||||
|
||||
<div class="viz-control-row">
|
||||
<div class="viz-control-group" style="margin-bottom: 0;">
|
||||
<label>Algorithm</label>
|
||||
<select x-model="algorithm">
|
||||
<option value="semantic">Semantic (Vector Similarity)</option>
|
||||
<option value="keyword">Keyword (Token Matching)</option>
|
||||
<option value="fuzzy">Fuzzy (Character Overlap)</option>
|
||||
<option value="hybrid" selected>Hybrid (RRF Fusion)</option>
|
||||
</select>
|
||||
</div>
|
||||
|
||||
<div style="display: flex; align-items: flex-end;">
|
||||
<button type="submit" class="viz-btn" style="width: 100%;">Search & Visualize</button>
|
||||
</div>
|
||||
|
||||
<div style="display: flex; align-items: flex-end;">
|
||||
<button type="button" class="viz-btn-secondary" @click="showAdvanced = !showAdvanced" style="white-space: nowrap;">
|
||||
<span x-text="showAdvanced ? 'Hide Advanced' : 'Advanced'"></span>
|
||||
</button>
|
||||
</div>
|
||||
</div>
|
||||
|
||||
<!-- Advanced Options (Collapsible) -->
|
||||
<div class="viz-advanced-section" x-show="showAdvanced" x-transition.opacity.duration.200ms>
|
||||
<h3 style="margin-top: 0; margin-bottom: 16px; font-size: 16px;">Advanced Options</h3>
|
||||
|
||||
<div class="viz-advanced-grid">
|
||||
<div class="viz-control-group">
|
||||
<label>Document Types</label>
|
||||
<select x-model="docTypes" multiple>
|
||||
<option value="">All Types (cross-app search)</option>
|
||||
<option value="note">Notes</option>
|
||||
<option value="file">Files</option>
|
||||
<option value="calendar">Calendar Events</option>
|
||||
<option value="contact">Contacts</option>
|
||||
<option value="deck">Deck Cards</option>
|
||||
</select>
|
||||
<small style="color: #666; display: block; margin-top: 4px;">
|
||||
Hold Ctrl/Cmd to select multiple
|
||||
</small>
|
||||
</div>
|
||||
|
||||
<div>
|
||||
<div class="viz-control-group">
|
||||
<label>Score Threshold (Semantic/Hybrid)</label>
|
||||
<input type="number" x-model.number="scoreThreshold" min="0" max="1" step="0.1" />
|
||||
</div>
|
||||
|
||||
<div class="viz-control-group">
|
||||
<label>Result Limit</label>
|
||||
<input type="number" x-model.number="limit" min="1" max="100" />
|
||||
</div>
|
||||
</div>
|
||||
</div>
|
||||
|
||||
<!-- Hybrid Weights (only when hybrid selected) -->
|
||||
<div x-show="algorithm === 'hybrid'" style="margin-top: 16px; padding: 12px; background: #e9ecef; border-radius: 4px;">
|
||||
<label style="margin-bottom: 12px; display: block;">Hybrid Algorithm Weights</label>
|
||||
|
||||
<div style="margin-bottom: 8px;">
|
||||
<label style="display: inline-block; width: 100px; font-weight: normal;">Semantic:</label>
|
||||
<input type="range" x-model.number="semanticWeight" min="0" max="1" step="0.1" style="width: 200px; display: inline-block;">
|
||||
<span class="viz-weight-display" x-text="semanticWeight.toFixed(1)"></span>
|
||||
</div>
|
||||
<div style="margin-bottom: 8px;">
|
||||
<label style="display: inline-block; width: 100px; font-weight: normal;">Keyword:</label>
|
||||
<input type="range" x-model.number="keywordWeight" min="0" max="1" step="0.1" style="width: 200px; display: inline-block;">
|
||||
<span class="viz-weight-display" x-text="keywordWeight.toFixed(1)"></span>
|
||||
</div>
|
||||
<div>
|
||||
<label style="display: inline-block; width: 100px; font-weight: normal;">Fuzzy:</label>
|
||||
<input type="range" x-model.number="fuzzyWeight" min="0" max="1" step="0.1" style="width: 200px; display: inline-block;">
|
||||
<span class="viz-weight-display" x-text="fuzzyWeight.toFixed(1)"></span>
|
||||
</div>
|
||||
</div>
|
||||
</div>
|
||||
</div>
|
||||
</form>
|
||||
</div>
|
||||
|
||||
<div class="viz-card">
|
||||
<div id="viz-plot-container">
|
||||
<div x-show="loading" class="viz-loading-overlay" x-transition.opacity.duration.200ms>
|
||||
Executing search and computing PCA projection...
|
||||
</div>
|
||||
<div id="viz-plot" x-show="!loading" x-transition.opacity.duration.200ms></div>
|
||||
</div>
|
||||
</div>
|
||||
|
||||
<div class="viz-card">
|
||||
<h3>Search Results (<span x-text="loading ? '...' : results.length"></span>)</h3>
|
||||
|
||||
<div x-show="loading" class="viz-loading" x-transition.opacity.duration.200ms>
|
||||
Loading results...
|
||||
</div>
|
||||
|
||||
<div x-show="!loading && results.length === 0" class="viz-no-results" x-transition.opacity.duration.200ms>
|
||||
No results found. Try a different query or adjust your search parameters.
|
||||
</div>
|
||||
|
||||
<template x-if="!loading && results.length > 0">
|
||||
<div x-transition.opacity.duration.200ms>
|
||||
<template x-for="result in results" :key="result.id">
|
||||
<div style="padding: 12px; border-bottom: 1px solid #eee;">
|
||||
<a :href="getNextcloudUrl(result)" target="_blank" style="font-weight: 500; color: #0066cc; text-decoration: none;">
|
||||
<span x-text="result.title"></span>
|
||||
</a>
|
||||
<div style="font-size: 14px; color: #666; margin-top: 4px;" x-text="result.excerpt"></div>
|
||||
<div style="font-size: 12px; color: #999; margin-top: 4px;">
|
||||
Score: <span x-text="result.score.toFixed(3)"></span> |
|
||||
Type: <span x-text="result.doc_type"></span>
|
||||
</div>
|
||||
</div>
|
||||
</template>
|
||||
</div>
|
||||
</template>
|
||||
</div>
|
||||
</div>
|
||||
"""
|
||||
|
||||
return HTMLResponse(content=html_content)
|
||||
|
||||
|
||||
@requires("authenticated", redirect="oauth_login")
|
||||
async def vector_visualization_search(request: Request) -> JSONResponse:
|
||||
"""Execute server-side search and return 2D coordinates + results.
|
||||
|
||||
All processing happens server-side:
|
||||
1. Execute search via shared algorithm module
|
||||
2. Fetch matching vectors from Qdrant
|
||||
3. Apply PCA reduction (768-dim → 2D)
|
||||
4. Return coordinates + metadata only
|
||||
|
||||
Args:
|
||||
request: Starlette request with query parameters
|
||||
|
||||
Returns:
|
||||
JSON response with coordinates_2d and results
|
||||
"""
|
||||
settings = get_settings()
|
||||
|
||||
if not settings.vector_sync_enabled:
|
||||
return JSONResponse(
|
||||
{"success": False, "error": "Vector sync not enabled"},
|
||||
status_code=400,
|
||||
)
|
||||
|
||||
# Get user info from auth context
|
||||
username = (
|
||||
request.user.display_name if hasattr(request.user, "display_name") else None
|
||||
)
|
||||
|
||||
if not username:
|
||||
return JSONResponse(
|
||||
{"success": False, "error": "User not authenticated"},
|
||||
status_code=401,
|
||||
)
|
||||
|
||||
# Parse query parameters
|
||||
query = request.query_params.get("query", "")
|
||||
algorithm = request.query_params.get("algorithm", "hybrid")
|
||||
limit = int(request.query_params.get("limit", "50"))
|
||||
score_threshold = float(request.query_params.get("score_threshold", "0.7"))
|
||||
semantic_weight = float(request.query_params.get("semantic_weight", "0.5"))
|
||||
keyword_weight = float(request.query_params.get("keyword_weight", "0.3"))
|
||||
fuzzy_weight = float(request.query_params.get("fuzzy_weight", "0.2"))
|
||||
|
||||
# Parse doc_types (comma-separated list, None = all types)
|
||||
doc_types_param = request.query_params.get("doc_types", "")
|
||||
doc_types = doc_types_param.split(",") if doc_types_param else None
|
||||
|
||||
logger.info(
|
||||
f"Viz search: user={username}, query='{query}', "
|
||||
f"algorithm={algorithm}, limit={limit}, doc_types={doc_types}"
|
||||
)
|
||||
|
||||
try:
|
||||
# Get authenticated HTTP client from session
|
||||
# In BasicAuth mode: uses username/password from session
|
||||
# In OAuth mode: uses access token from session
|
||||
from nextcloud_mcp_server.auth.userinfo_routes import (
|
||||
_get_authenticated_client_for_userinfo,
|
||||
)
|
||||
from nextcloud_mcp_server.client.notes import NotesClient
|
||||
|
||||
async with await _get_authenticated_client_for_userinfo(request) as http_client:
|
||||
# Create NotesClient directly with authenticated HTTP client
|
||||
notes_client = NotesClient(http_client, username)
|
||||
|
||||
# Wrap in a minimal client object for search algorithms
|
||||
# This conforms to NextcloudClientProtocol but only implements notes
|
||||
class MinimalNextcloudClient:
|
||||
def __init__(self, notes_client, username):
|
||||
self._notes = notes_client
|
||||
self.username = username
|
||||
|
||||
@property
|
||||
def notes(self):
|
||||
return self._notes
|
||||
|
||||
@property
|
||||
def webdav(self):
|
||||
return None
|
||||
|
||||
@property
|
||||
def calendar(self):
|
||||
return None
|
||||
|
||||
@property
|
||||
def contacts(self):
|
||||
return None
|
||||
|
||||
@property
|
||||
def deck(self):
|
||||
return None
|
||||
|
||||
@property
|
||||
def cookbook(self):
|
||||
return None
|
||||
|
||||
@property
|
||||
def tables(self):
|
||||
return None
|
||||
|
||||
nextcloud_client = MinimalNextcloudClient(notes_client, username)
|
||||
|
||||
# Create search algorithm
|
||||
if algorithm == "semantic":
|
||||
search_algo = SemanticSearchAlgorithm(score_threshold=score_threshold)
|
||||
elif algorithm == "keyword":
|
||||
search_algo = KeywordSearchAlgorithm()
|
||||
elif algorithm == "fuzzy":
|
||||
search_algo = FuzzySearchAlgorithm()
|
||||
elif algorithm == "hybrid":
|
||||
search_algo = HybridSearchAlgorithm(
|
||||
semantic_weight=semantic_weight,
|
||||
keyword_weight=keyword_weight,
|
||||
fuzzy_weight=fuzzy_weight,
|
||||
)
|
||||
else:
|
||||
return JSONResponse(
|
||||
{"success": False, "error": f"Unknown algorithm: {algorithm}"},
|
||||
status_code=400,
|
||||
)
|
||||
|
||||
# Execute search (supports cross-app when doc_types=None)
|
||||
# Get unverified results with buffer for filtering
|
||||
all_results = []
|
||||
if doc_types is None or len(doc_types) == 0:
|
||||
# Cross-app search - search all indexed types
|
||||
unverified_results = await search_algo.search(
|
||||
query=query,
|
||||
user_id=username,
|
||||
limit=limit * 2, # Buffer for verification filtering
|
||||
doc_type=None, # Search all types
|
||||
score_threshold=score_threshold,
|
||||
)
|
||||
all_results.extend(unverified_results)
|
||||
else:
|
||||
# Search each document type and combine
|
||||
for doc_type in doc_types:
|
||||
unverified_results = await search_algo.search(
|
||||
query=query,
|
||||
user_id=username,
|
||||
limit=limit * 2, # Buffer for verification filtering
|
||||
doc_type=doc_type,
|
||||
score_threshold=score_threshold,
|
||||
)
|
||||
all_results.extend(unverified_results)
|
||||
# Sort by score before verification
|
||||
all_results.sort(key=lambda r: r.score, reverse=True)
|
||||
|
||||
# Verify access for all results (deduplicates and filters)
|
||||
from nextcloud_mcp_server.search.verification import verify_search_results
|
||||
|
||||
verified_results = await verify_search_results(
|
||||
all_results, nextcloud_client
|
||||
)
|
||||
search_results = verified_results[:limit]
|
||||
|
||||
if not search_results:
|
||||
return JSONResponse(
|
||||
{
|
||||
"success": True,
|
||||
"results": [],
|
||||
"coordinates_2d": [],
|
||||
"message": "No results found",
|
||||
}
|
||||
)
|
||||
|
||||
# Fetch vectors for matching results from Qdrant
|
||||
qdrant_client = await get_qdrant_client()
|
||||
doc_ids = [r.id for r in search_results]
|
||||
|
||||
# Retrieve vectors for the matching documents
|
||||
from qdrant_client.models import FieldCondition, Filter, MatchAny
|
||||
|
||||
points_response = await qdrant_client.scroll(
|
||||
collection_name=settings.get_collection_name(),
|
||||
scroll_filter=Filter(
|
||||
must=[
|
||||
FieldCondition(
|
||||
key="doc_id",
|
||||
match=MatchAny(any=[str(doc_id) for doc_id in doc_ids]),
|
||||
),
|
||||
FieldCondition(
|
||||
key="user_id",
|
||||
match={"value": username},
|
||||
),
|
||||
]
|
||||
),
|
||||
limit=len(doc_ids) * 2, # Account for multiple chunks per doc
|
||||
with_vectors=True,
|
||||
with_payload=["doc_id"], # Need doc_id to map vectors to results
|
||||
)
|
||||
|
||||
points = points_response[0]
|
||||
|
||||
if not points:
|
||||
return JSONResponse(
|
||||
{
|
||||
"success": True,
|
||||
"results": [],
|
||||
"coordinates_2d": [],
|
||||
"message": "No vectors found for results",
|
||||
}
|
||||
)
|
||||
|
||||
# Extract vectors
|
||||
vectors = np.array([p.vector for p in points if p.vector is not None])
|
||||
|
||||
if len(vectors) < 2:
|
||||
# Not enough points for PCA
|
||||
return JSONResponse(
|
||||
{
|
||||
"success": True,
|
||||
"results": [
|
||||
{
|
||||
"id": r.id,
|
||||
"doc_type": r.doc_type,
|
||||
"title": r.title,
|
||||
"excerpt": r.excerpt,
|
||||
"score": r.score,
|
||||
}
|
||||
for r in search_results
|
||||
],
|
||||
"coordinates_2d": [[0, 0]] * len(search_results),
|
||||
"message": "Not enough vectors for PCA",
|
||||
}
|
||||
)
|
||||
|
||||
# Apply PCA dimensionality reduction (768-dim → 2D)
|
||||
pca = PCA(n_components=2)
|
||||
coords_2d = pca.fit_transform(vectors)
|
||||
|
||||
# After fit, these attributes are guaranteed to be set
|
||||
assert pca.explained_variance_ratio_ is not None
|
||||
|
||||
logger.info(
|
||||
f"PCA explained variance: PC1={pca.explained_variance_ratio_[0]:.3f}, "
|
||||
f"PC2={pca.explained_variance_ratio_[1]:.3f}"
|
||||
)
|
||||
|
||||
# Map results to coordinates (use first chunk per document)
|
||||
result_coords = []
|
||||
seen_doc_ids = set()
|
||||
|
||||
for point, coord in zip(points, coords_2d):
|
||||
if point.payload:
|
||||
doc_id = int(point.payload.get("doc_id", 0))
|
||||
if doc_id not in seen_doc_ids and doc_id in doc_ids:
|
||||
seen_doc_ids.add(doc_id)
|
||||
result_coords.append(coord.tolist())
|
||||
|
||||
# Build response
|
||||
response_results = [
|
||||
{
|
||||
"id": r.id,
|
||||
"doc_type": r.doc_type,
|
||||
"title": r.title,
|
||||
"excerpt": r.excerpt,
|
||||
"score": r.score,
|
||||
}
|
||||
for r in search_results
|
||||
]
|
||||
|
||||
return JSONResponse(
|
||||
{
|
||||
"success": True,
|
||||
"results": response_results,
|
||||
"coordinates_2d": result_coords[: len(search_results)],
|
||||
"pca_variance": {
|
||||
"pc1": float(pca.explained_variance_ratio_[0]),
|
||||
"pc2": float(pca.explained_variance_ratio_[1]),
|
||||
},
|
||||
}
|
||||
)
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"Viz search error: {e}", exc_info=True)
|
||||
return JSONResponse(
|
||||
{"success": False, "error": str(e)},
|
||||
status_code=500,
|
||||
)
|
||||
@@ -66,8 +66,12 @@ class ObservabilityMiddleware(BaseHTTPMiddleware):
|
||||
# Record start time
|
||||
start_time = time.time()
|
||||
|
||||
# Skip tracing for health/metrics endpoints to reduce noise
|
||||
should_trace = not (path.startswith("/health/") or path == "/metrics")
|
||||
# Skip tracing for health/metrics/polling endpoints to reduce noise
|
||||
should_trace = not (
|
||||
path.startswith("/health/")
|
||||
or path == "/metrics"
|
||||
or path == "/app/vector-sync/status"
|
||||
)
|
||||
|
||||
try:
|
||||
if should_trace:
|
||||
|
||||
@@ -0,0 +1,33 @@
|
||||
"""Search algorithms module for unified multi-algorithm search.
|
||||
|
||||
This module provides a unified interface for different search algorithms:
|
||||
- Semantic search (vector similarity)
|
||||
- Keyword search (token-based matching)
|
||||
- Fuzzy search (character overlap)
|
||||
- Hybrid search (RRF fusion of multiple algorithms)
|
||||
|
||||
All algorithms share the same interface and can be used interchangeably by both
|
||||
MCP tools and the visualization pane.
|
||||
"""
|
||||
|
||||
from nextcloud_mcp_server.search.algorithms import (
|
||||
NextcloudClientProtocol,
|
||||
SearchAlgorithm,
|
||||
SearchResult,
|
||||
get_indexed_doc_types,
|
||||
)
|
||||
from nextcloud_mcp_server.search.fuzzy import FuzzySearchAlgorithm
|
||||
from nextcloud_mcp_server.search.hybrid import HybridSearchAlgorithm
|
||||
from nextcloud_mcp_server.search.keyword import KeywordSearchAlgorithm
|
||||
from nextcloud_mcp_server.search.semantic import SemanticSearchAlgorithm
|
||||
|
||||
__all__ = [
|
||||
"NextcloudClientProtocol",
|
||||
"SearchAlgorithm",
|
||||
"SearchResult",
|
||||
"get_indexed_doc_types",
|
||||
"SemanticSearchAlgorithm",
|
||||
"KeywordSearchAlgorithm",
|
||||
"FuzzySearchAlgorithm",
|
||||
"HybridSearchAlgorithm",
|
||||
]
|
||||
@@ -0,0 +1,200 @@
|
||||
"""Base interfaces and data structures for search algorithms."""
|
||||
|
||||
from abc import ABC, abstractmethod
|
||||
from dataclasses import dataclass
|
||||
from typing import Any, Protocol, runtime_checkable
|
||||
|
||||
|
||||
@runtime_checkable
|
||||
class NextcloudClientProtocol(Protocol):
|
||||
"""Protocol for Nextcloud client supporting multi-document search.
|
||||
|
||||
This protocol defines the interface that search algorithms need from a
|
||||
Nextcloud client to access documents across different apps (Notes, Files,
|
||||
Calendar, etc.). The client provides access to app-specific sub-clients
|
||||
that handle the actual API calls.
|
||||
|
||||
Document types (e.g., "note", "file", "calendar") are NOT 1:1 with apps.
|
||||
For example, the Notes app specializes in markdown files, while Files/WebDAV
|
||||
handles multiple file types. The abstraction is at the document type level.
|
||||
|
||||
Search algorithms query Qdrant to determine which document types are actually
|
||||
indexed before attempting to access them, enabling graceful cross-app search.
|
||||
"""
|
||||
|
||||
username: str
|
||||
|
||||
# App-specific clients that search algorithms dispatch to
|
||||
@property
|
||||
def notes(self) -> Any:
|
||||
"""Notes client for accessing note documents."""
|
||||
...
|
||||
|
||||
@property
|
||||
def webdav(self) -> Any:
|
||||
"""WebDAV client for accessing file documents."""
|
||||
...
|
||||
|
||||
@property
|
||||
def calendar(self) -> Any:
|
||||
"""Calendar client for accessing event/task documents."""
|
||||
...
|
||||
|
||||
@property
|
||||
def contacts(self) -> Any:
|
||||
"""Contacts client for accessing contact card documents."""
|
||||
...
|
||||
|
||||
@property
|
||||
def deck(self) -> Any:
|
||||
"""Deck client for accessing deck card documents."""
|
||||
...
|
||||
|
||||
@property
|
||||
def cookbook(self) -> Any:
|
||||
"""Cookbook client for accessing recipe documents."""
|
||||
...
|
||||
|
||||
@property
|
||||
def tables(self) -> Any:
|
||||
"""Tables client for accessing table row documents."""
|
||||
...
|
||||
|
||||
|
||||
async def get_indexed_doc_types(user_id: str) -> set[str]:
|
||||
"""Query Qdrant to get actually-indexed document types for a user.
|
||||
|
||||
This enables search algorithms to check which document types are available
|
||||
before attempting to search/verify them, allowing graceful cross-app search.
|
||||
|
||||
Args:
|
||||
user_id: User ID to filter by
|
||||
|
||||
Returns:
|
||||
Set of document type strings (e.g., {"note", "file", "calendar"})
|
||||
|
||||
Example:
|
||||
>>> types = await get_indexed_doc_types("alice")
|
||||
>>> if "note" in types:
|
||||
... # Search notes
|
||||
"""
|
||||
import logging
|
||||
|
||||
from qdrant_client.models import FieldCondition, Filter, MatchValue
|
||||
|
||||
from nextcloud_mcp_server.config import get_settings
|
||||
from nextcloud_mcp_server.vector.qdrant_client import get_qdrant_client
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
settings = get_settings()
|
||||
|
||||
qdrant_client = await get_qdrant_client()
|
||||
collection = settings.get_collection_name()
|
||||
|
||||
# Use scroll to sample documents and extract doc_types
|
||||
# Note: This could be optimized with a facet/aggregation query if Qdrant adds support
|
||||
try:
|
||||
scroll_results, _next_offset = await qdrant_client.scroll(
|
||||
collection_name=collection,
|
||||
scroll_filter=Filter(
|
||||
must=[FieldCondition(key="user_id", match=MatchValue(value=user_id))]
|
||||
),
|
||||
limit=1000, # Sample size to discover types
|
||||
with_payload=["doc_type"],
|
||||
with_vectors=False, # Don't need vectors for type discovery
|
||||
)
|
||||
|
||||
doc_types = {
|
||||
point.payload.get("doc_type")
|
||||
for point in scroll_results
|
||||
if point.payload.get("doc_type")
|
||||
}
|
||||
|
||||
logger.debug(f"Found indexed document types for user {user_id}: {doc_types}")
|
||||
return doc_types
|
||||
|
||||
except Exception as e:
|
||||
logger.warning(f"Failed to query Qdrant for doc_types: {e}")
|
||||
return set()
|
||||
|
||||
|
||||
@dataclass
|
||||
class SearchResult:
|
||||
"""A single search result with metadata and score.
|
||||
|
||||
Attributes:
|
||||
id: Document ID
|
||||
doc_type: Document type (note, file, calendar, contact, etc.)
|
||||
title: Document title
|
||||
excerpt: Content excerpt showing match context
|
||||
score: Relevance score (0.0-1.0, higher is better)
|
||||
metadata: Additional algorithm-specific metadata
|
||||
"""
|
||||
|
||||
id: int
|
||||
doc_type: str
|
||||
title: str
|
||||
excerpt: str
|
||||
score: float
|
||||
metadata: dict[str, Any] | None = None
|
||||
|
||||
def __post_init__(self):
|
||||
"""Validate score is in valid range."""
|
||||
if not 0.0 <= self.score <= 1.0:
|
||||
raise ValueError(f"Score must be between 0.0 and 1.0, got {self.score}")
|
||||
|
||||
|
||||
class SearchAlgorithm(ABC):
|
||||
"""Abstract base class for search algorithms.
|
||||
|
||||
All search algorithms must implement the search() method with consistent
|
||||
interface, allowing them to be used interchangeably.
|
||||
"""
|
||||
|
||||
@abstractmethod
|
||||
async def search(
|
||||
self,
|
||||
query: str,
|
||||
user_id: str,
|
||||
limit: int = 10,
|
||||
doc_type: str | None = None,
|
||||
**kwargs: Any,
|
||||
) -> list[SearchResult]:
|
||||
"""Execute search with the given parameters.
|
||||
|
||||
Args:
|
||||
query: Search query string
|
||||
user_id: User ID for multi-tenant filtering
|
||||
limit: Maximum number of results to return
|
||||
doc_type: Optional document type filter (note, file, calendar, etc.)
|
||||
**kwargs: Algorithm-specific parameters
|
||||
|
||||
Returns:
|
||||
List of SearchResult objects ranked by relevance
|
||||
|
||||
Raises:
|
||||
McpError: If search fails or configuration is invalid
|
||||
"""
|
||||
pass
|
||||
|
||||
@property
|
||||
@abstractmethod
|
||||
def name(self) -> str:
|
||||
"""Return algorithm name for identification."""
|
||||
pass
|
||||
|
||||
@property
|
||||
def supports_scoring(self) -> bool:
|
||||
"""Whether this algorithm provides meaningful relevance scores.
|
||||
|
||||
Default: True. Override if algorithm doesn't support scoring.
|
||||
"""
|
||||
return True
|
||||
|
||||
@property
|
||||
def requires_vector_db(self) -> bool:
|
||||
"""Whether this algorithm requires vector database.
|
||||
|
||||
Default: False. Override for semantic search.
|
||||
"""
|
||||
return False
|
||||
@@ -0,0 +1,219 @@
|
||||
"""Fuzzy search algorithm using character overlap matching on Qdrant payload."""
|
||||
|
||||
import logging
|
||||
from typing import Any
|
||||
|
||||
from qdrant_client.models import FieldCondition, Filter, MatchValue
|
||||
|
||||
from nextcloud_mcp_server.config import get_settings
|
||||
from nextcloud_mcp_server.search.algorithms import SearchAlgorithm, SearchResult
|
||||
from nextcloud_mcp_server.vector.qdrant_client import get_qdrant_client
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
class FuzzySearchAlgorithm(SearchAlgorithm):
|
||||
"""Fuzzy search using simple character-based similarity.
|
||||
|
||||
Implements character overlap matching with configurable threshold:
|
||||
- Compares character sets between query and text
|
||||
- Requires configurable % character overlap to match (default: 70%)
|
||||
- Tolerant to typos and minor variations
|
||||
"""
|
||||
|
||||
def __init__(self, threshold: float = 0.7):
|
||||
"""Initialize fuzzy search algorithm.
|
||||
|
||||
Args:
|
||||
threshold: Minimum character overlap ratio (0-1, default: 0.7)
|
||||
"""
|
||||
if not 0.0 <= threshold <= 1.0:
|
||||
raise ValueError(f"Threshold must be between 0.0 and 1.0, got {threshold}")
|
||||
self.threshold = threshold
|
||||
|
||||
@property
|
||||
def name(self) -> str:
|
||||
return "fuzzy"
|
||||
|
||||
async def search(
|
||||
self,
|
||||
query: str,
|
||||
user_id: str,
|
||||
limit: int = 10,
|
||||
doc_type: str | None = None,
|
||||
**kwargs: Any,
|
||||
) -> list[SearchResult]:
|
||||
"""Execute fuzzy search using character overlap on Qdrant payload.
|
||||
|
||||
Queries Qdrant for all indexed documents, then scores based on character
|
||||
overlap in title and excerpt fields. Returns unverified results - access
|
||||
verification should be performed separately at the final output stage.
|
||||
|
||||
Args:
|
||||
query: Search query
|
||||
user_id: User ID for filtering
|
||||
limit: Maximum results to return
|
||||
doc_type: Optional document type filter (None = all types)
|
||||
**kwargs: Additional parameters (threshold override)
|
||||
|
||||
Returns:
|
||||
List of unverified SearchResult objects ranked by character overlap score
|
||||
"""
|
||||
settings = get_settings()
|
||||
threshold = kwargs.get("threshold", self.threshold)
|
||||
|
||||
logger.info(
|
||||
f"Fuzzy search: query='{query}', user={user_id}, "
|
||||
f"limit={limit}, threshold={threshold}, doc_type={doc_type}"
|
||||
)
|
||||
|
||||
# Build Qdrant filter
|
||||
filter_conditions = [
|
||||
FieldCondition(key="user_id", match=MatchValue(value=user_id))
|
||||
]
|
||||
if doc_type:
|
||||
filter_conditions.append(
|
||||
FieldCondition(key="doc_type", match=MatchValue(value=doc_type))
|
||||
)
|
||||
|
||||
# Scroll through Qdrant to get all matching documents
|
||||
qdrant_client = await get_qdrant_client()
|
||||
collection = settings.get_collection_name()
|
||||
|
||||
all_points = []
|
||||
offset = None
|
||||
|
||||
# Scroll through all points matching filter
|
||||
while True:
|
||||
scroll_result, next_offset = await qdrant_client.scroll(
|
||||
collection_name=collection,
|
||||
scroll_filter=Filter(must=filter_conditions),
|
||||
limit=100, # Batch size
|
||||
offset=offset,
|
||||
with_payload=["doc_id", "doc_type", "title", "excerpt", "chunk_index"],
|
||||
with_vectors=False, # Don't need vectors
|
||||
)
|
||||
|
||||
all_points.extend(scroll_result)
|
||||
|
||||
if next_offset is None:
|
||||
break
|
||||
offset = next_offset
|
||||
|
||||
logger.debug(f"Retrieved {len(all_points)} points from Qdrant for fuzzy search")
|
||||
|
||||
# Deduplicate by (doc_id, doc_type) - keep first chunk
|
||||
seen_docs = {}
|
||||
for point in all_points:
|
||||
doc_id = int(point.payload["doc_id"])
|
||||
dtype = point.payload.get("doc_type", "note")
|
||||
doc_key = (doc_id, dtype)
|
||||
|
||||
chunk_idx = point.payload.get("chunk_index", 0)
|
||||
if doc_key not in seen_docs or chunk_idx == 0:
|
||||
seen_docs[doc_key] = point
|
||||
|
||||
logger.debug(f"Deduplicated to {len(seen_docs)} unique documents")
|
||||
|
||||
# Score each document based on fuzzy matches
|
||||
scored_results = []
|
||||
query_lower = query.lower()
|
||||
|
||||
for doc_key, point in seen_docs.items():
|
||||
doc_id, dtype = doc_key
|
||||
title = point.payload.get("title", "")
|
||||
excerpt = point.payload.get("excerpt", "")
|
||||
|
||||
# Check title match
|
||||
title_score = self._calculate_char_overlap(query_lower, title.lower())
|
||||
|
||||
# Check excerpt match
|
||||
excerpt_score = self._calculate_char_overlap(query_lower, excerpt.lower())
|
||||
|
||||
# Use best score
|
||||
best_score = max(title_score, excerpt_score)
|
||||
|
||||
if best_score >= threshold:
|
||||
match_location = "title" if title_score >= excerpt_score else "excerpt"
|
||||
scored_results.append(
|
||||
{
|
||||
"doc_id": doc_id,
|
||||
"doc_type": dtype,
|
||||
"title": title,
|
||||
"excerpt": excerpt
|
||||
if excerpt_score >= title_score
|
||||
else f"Title match: {title}",
|
||||
"score": best_score,
|
||||
"match_location": match_location,
|
||||
}
|
||||
)
|
||||
|
||||
# Sort by score (descending) and limit
|
||||
scored_results.sort(key=lambda x: x["score"], reverse=True)
|
||||
top_results = scored_results[:limit]
|
||||
|
||||
# Return unverified results (verification happens at output stage)
|
||||
final_results = []
|
||||
for result in top_results:
|
||||
final_results.append(
|
||||
SearchResult(
|
||||
id=result["doc_id"],
|
||||
doc_type=result["doc_type"],
|
||||
title=result["title"],
|
||||
excerpt=result["excerpt"],
|
||||
score=result["score"],
|
||||
metadata={"match_location": result["match_location"]},
|
||||
)
|
||||
)
|
||||
|
||||
logger.info(f"Fuzzy search returned {len(final_results)} unverified results")
|
||||
if final_results:
|
||||
result_details = [
|
||||
f"{r.doc_type}_{r.id} (score={r.score:.3f}, title='{r.title}')"
|
||||
for r in final_results[:5]
|
||||
]
|
||||
logger.debug(f"Top fuzzy results: {', '.join(result_details)}")
|
||||
|
||||
return final_results
|
||||
|
||||
def _calculate_char_overlap(self, query: str, text: str) -> float:
|
||||
"""Calculate character overlap ratio between query and text.
|
||||
|
||||
Args:
|
||||
query: Query string (normalized)
|
||||
text: Text to compare (normalized)
|
||||
|
||||
Returns:
|
||||
Overlap ratio (0.0-1.0)
|
||||
"""
|
||||
if not query or not text:
|
||||
return 0.0
|
||||
|
||||
# Convert to character sets
|
||||
query_chars = set(query)
|
||||
text_chars = set(text)
|
||||
|
||||
# Calculate overlap
|
||||
overlap = query_chars & text_chars
|
||||
overlap_ratio = len(overlap) / len(query_chars)
|
||||
|
||||
return overlap_ratio
|
||||
|
||||
def _extract_excerpt(self, content: str, max_length: int = 200) -> str:
|
||||
"""Extract excerpt from content.
|
||||
|
||||
Args:
|
||||
content: Full document content
|
||||
max_length: Maximum excerpt length
|
||||
|
||||
Returns:
|
||||
Excerpt string
|
||||
"""
|
||||
if not content:
|
||||
return ""
|
||||
|
||||
excerpt = content[:max_length].strip()
|
||||
if len(content) > max_length:
|
||||
excerpt += "..."
|
||||
|
||||
return excerpt
|
||||
@@ -0,0 +1,237 @@
|
||||
"""Hybrid search algorithm using Reciprocal Rank Fusion (RRF)."""
|
||||
|
||||
import asyncio
|
||||
import logging
|
||||
from collections import defaultdict
|
||||
from typing import Any
|
||||
|
||||
from nextcloud_mcp_server.search.algorithms import SearchAlgorithm, SearchResult
|
||||
from nextcloud_mcp_server.search.fuzzy import FuzzySearchAlgorithm
|
||||
from nextcloud_mcp_server.search.keyword import KeywordSearchAlgorithm
|
||||
from nextcloud_mcp_server.search.semantic import SemanticSearchAlgorithm
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
class HybridSearchAlgorithm(SearchAlgorithm):
|
||||
"""Hybrid search combining multiple algorithms using Reciprocal Rank Fusion.
|
||||
|
||||
Implements RRF from ADR-003 to combine results from:
|
||||
- Semantic search (vector similarity)
|
||||
- Keyword search (token matching)
|
||||
- Fuzzy search (character overlap)
|
||||
|
||||
RRF formula: score = weight / (k + rank)
|
||||
where k=60 (standard value) and rank is 1-indexed position.
|
||||
"""
|
||||
|
||||
DEFAULT_RRF_K = 60 # Standard RRF constant
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
semantic_weight: float = 0.5,
|
||||
keyword_weight: float = 0.3,
|
||||
fuzzy_weight: float = 0.2,
|
||||
rrf_k: int = DEFAULT_RRF_K,
|
||||
):
|
||||
"""Initialize hybrid search with algorithm weights.
|
||||
|
||||
Args:
|
||||
semantic_weight: Weight for semantic results (default: 0.5)
|
||||
keyword_weight: Weight for keyword results (default: 0.3)
|
||||
fuzzy_weight: Weight for fuzzy results (default: 0.2)
|
||||
rrf_k: RRF constant for rank decay (default: 60)
|
||||
|
||||
Raises:
|
||||
ValueError: If weights are invalid
|
||||
"""
|
||||
# Validate weights
|
||||
if semantic_weight < 0 or keyword_weight < 0 or fuzzy_weight < 0:
|
||||
raise ValueError("Weights must be non-negative")
|
||||
|
||||
total_weight = semantic_weight + keyword_weight + fuzzy_weight
|
||||
if total_weight > 1.0:
|
||||
raise ValueError(f"Weights sum to {total_weight:.2f}, must be ≤1.0")
|
||||
|
||||
if total_weight == 0.0:
|
||||
raise ValueError("At least one weight must be > 0")
|
||||
|
||||
self.semantic_weight = semantic_weight
|
||||
self.keyword_weight = keyword_weight
|
||||
self.fuzzy_weight = fuzzy_weight
|
||||
self.rrf_k = rrf_k
|
||||
self.total_weight = total_weight
|
||||
|
||||
# Initialize sub-algorithms
|
||||
self.semantic = SemanticSearchAlgorithm()
|
||||
self.keyword = KeywordSearchAlgorithm()
|
||||
self.fuzzy = FuzzySearchAlgorithm()
|
||||
|
||||
@property
|
||||
def name(self) -> str:
|
||||
return "hybrid"
|
||||
|
||||
@property
|
||||
def requires_vector_db(self) -> bool:
|
||||
# Requires vector DB if semantic search has non-zero weight
|
||||
return self.semantic_weight > 0
|
||||
|
||||
async def search(
|
||||
self,
|
||||
query: str,
|
||||
user_id: str,
|
||||
limit: int = 10,
|
||||
doc_type: str | None = None,
|
||||
**kwargs: Any,
|
||||
) -> list[SearchResult]:
|
||||
"""Execute hybrid search using RRF to combine algorithms.
|
||||
|
||||
Returns unverified results from combined algorithms. Access verification
|
||||
should be performed separately at the final output stage.
|
||||
|
||||
Args:
|
||||
query: Search query
|
||||
user_id: User ID for filtering
|
||||
limit: Maximum results to return
|
||||
doc_type: Optional document type filter
|
||||
**kwargs: Additional parameters passed to sub-algorithms
|
||||
|
||||
Returns:
|
||||
List of unverified SearchResult objects ranked by RRF combined score
|
||||
"""
|
||||
logger.info(
|
||||
f"Hybrid search: query='{query}', user={user_id}, limit={limit}, "
|
||||
f"weights=(semantic={self.semantic_weight}, keyword={self.keyword_weight}, "
|
||||
f"fuzzy={self.fuzzy_weight})"
|
||||
)
|
||||
|
||||
# Run algorithms in parallel
|
||||
tasks = []
|
||||
algo_names = []
|
||||
|
||||
if self.semantic_weight > 0:
|
||||
tasks.append(
|
||||
self.semantic.search(query, user_id, limit * 2, doc_type, **kwargs)
|
||||
)
|
||||
algo_names.append("semantic")
|
||||
|
||||
if self.keyword_weight > 0:
|
||||
tasks.append(
|
||||
self.keyword.search(query, user_id, limit * 2, doc_type, **kwargs)
|
||||
)
|
||||
algo_names.append("keyword")
|
||||
|
||||
if self.fuzzy_weight > 0:
|
||||
tasks.append(
|
||||
self.fuzzy.search(query, user_id, limit * 2, doc_type, **kwargs)
|
||||
)
|
||||
algo_names.append("fuzzy")
|
||||
|
||||
# Execute searches in parallel
|
||||
results_list = await asyncio.gather(*tasks)
|
||||
|
||||
# Build results dict
|
||||
algo_results = {}
|
||||
for algo_name, results in zip(algo_names, results_list):
|
||||
algo_results[algo_name] = results
|
||||
logger.debug(f"{algo_name} returned {len(results)} results")
|
||||
|
||||
# Combine using RRF
|
||||
combined_results = self._reciprocal_rank_fusion(
|
||||
algo_results,
|
||||
{
|
||||
"semantic": self.semantic_weight,
|
||||
"keyword": self.keyword_weight,
|
||||
"fuzzy": self.fuzzy_weight,
|
||||
},
|
||||
limit,
|
||||
)
|
||||
|
||||
logger.info(f"Hybrid search returned {len(combined_results)} combined results")
|
||||
if combined_results:
|
||||
result_details = [
|
||||
f"{r.doc_type}_{r.id} (score={r.score:.3f}, title='{r.title}')"
|
||||
for r in combined_results[:5]
|
||||
]
|
||||
logger.debug(f"Top hybrid results: {', '.join(result_details)}")
|
||||
|
||||
return combined_results
|
||||
|
||||
def _reciprocal_rank_fusion(
|
||||
self,
|
||||
algo_results: dict[str, list[SearchResult]],
|
||||
weights: dict[str, float],
|
||||
limit: int,
|
||||
) -> list[SearchResult]:
|
||||
"""Combine multiple ranked result lists using RRF.
|
||||
|
||||
Args:
|
||||
algo_results: Dict of algorithm_name -> ranked results
|
||||
weights: Dict of algorithm_name -> weight (0-1)
|
||||
limit: Maximum results to return
|
||||
|
||||
Returns:
|
||||
Combined and re-ranked results
|
||||
"""
|
||||
# Track RRF scores per document
|
||||
rrf_scores: dict[tuple[int, str], float] = defaultdict(float)
|
||||
# Track best result object for each document
|
||||
best_results: dict[tuple[int, str], SearchResult] = {}
|
||||
|
||||
for algo_name, results in algo_results.items():
|
||||
weight = weights.get(algo_name, 0.0)
|
||||
if weight == 0:
|
||||
continue
|
||||
|
||||
for rank, result in enumerate(results, start=1):
|
||||
doc_key = (result.id, result.doc_type)
|
||||
|
||||
# RRF formula: weight / (k + rank)
|
||||
rrf_score = weight / (self.rrf_k + rank)
|
||||
rrf_scores[doc_key] += rrf_score
|
||||
|
||||
# Track best result object (prefer higher original scores)
|
||||
if doc_key not in best_results:
|
||||
best_results[doc_key] = result
|
||||
elif result.score > best_results[doc_key].score:
|
||||
best_results[doc_key] = result
|
||||
|
||||
# Sort by combined RRF score
|
||||
sorted_docs = sorted(
|
||||
rrf_scores.items(),
|
||||
key=lambda x: x[1],
|
||||
reverse=True,
|
||||
)[:limit]
|
||||
|
||||
# Calculate normalization factor to scale RRF scores to 0-1 range
|
||||
# Theoretical max RRF score = total_weight / (rrf_k + 1)
|
||||
# Normalization factor = (rrf_k + 1) / total_weight
|
||||
normalization_factor = (self.rrf_k + 1) / self.total_weight
|
||||
|
||||
# Build final results with normalized RRF scores
|
||||
final_results = []
|
||||
for doc_key, rrf_score in sorted_docs:
|
||||
result = best_results[doc_key]
|
||||
|
||||
# Normalize RRF score to 0-1 range for better user comprehension
|
||||
normalized_score = rrf_score * normalization_factor
|
||||
|
||||
# Create new result with normalized score
|
||||
# Keep original metadata but add RRF details
|
||||
metadata = result.metadata or {}
|
||||
metadata["rrf_score_raw"] = rrf_score # Original RRF score
|
||||
metadata["original_score"] = result.score # Original algorithm score
|
||||
metadata["normalization_factor"] = normalization_factor
|
||||
|
||||
final_results.append(
|
||||
SearchResult(
|
||||
id=result.id,
|
||||
doc_type=result.doc_type,
|
||||
title=result.title,
|
||||
excerpt=result.excerpt,
|
||||
score=normalized_score, # Use normalized score (0-1 range)
|
||||
metadata=metadata,
|
||||
)
|
||||
)
|
||||
|
||||
return final_results
|
||||
@@ -0,0 +1,277 @@
|
||||
"""Keyword search algorithm using token-based matching on Qdrant payload (ADR-001)."""
|
||||
|
||||
import logging
|
||||
from typing import Any
|
||||
|
||||
from qdrant_client.models import FieldCondition, Filter, MatchValue
|
||||
|
||||
from nextcloud_mcp_server.config import get_settings
|
||||
from nextcloud_mcp_server.search.algorithms import SearchAlgorithm, SearchResult
|
||||
from nextcloud_mcp_server.vector.qdrant_client import get_qdrant_client
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
class KeywordSearchAlgorithm(SearchAlgorithm):
|
||||
"""Keyword search using token-based matching with weighted scoring.
|
||||
|
||||
Implements token-based search from ADR-001:
|
||||
- Title matches weighted 3x higher than content matches
|
||||
- Case-insensitive token matching
|
||||
- Relevance scoring based on match frequency and location
|
||||
"""
|
||||
|
||||
# Weighting constants from ADR-001
|
||||
TITLE_WEIGHT = 3.0
|
||||
CONTENT_WEIGHT = 1.0
|
||||
|
||||
@property
|
||||
def name(self) -> str:
|
||||
return "keyword"
|
||||
|
||||
async def search(
|
||||
self,
|
||||
query: str,
|
||||
user_id: str,
|
||||
limit: int = 10,
|
||||
doc_type: str | None = None,
|
||||
**kwargs: Any,
|
||||
) -> list[SearchResult]:
|
||||
"""Execute keyword search using token matching on Qdrant payload.
|
||||
|
||||
Queries Qdrant for all indexed documents, then scores based on token
|
||||
matches in title and excerpt fields. Returns unverified results - access
|
||||
verification should be performed separately at the final output stage.
|
||||
|
||||
Args:
|
||||
query: Search query to tokenize and match
|
||||
user_id: User ID for filtering
|
||||
limit: Maximum results to return
|
||||
doc_type: Optional document type filter (None = all types)
|
||||
**kwargs: Additional parameters (unused)
|
||||
|
||||
Returns:
|
||||
List of unverified SearchResult objects ranked by keyword match score
|
||||
"""
|
||||
settings = get_settings()
|
||||
|
||||
logger.info(
|
||||
f"Keyword search: query='{query}', user={user_id}, "
|
||||
f"limit={limit}, doc_type={doc_type}"
|
||||
)
|
||||
|
||||
# Tokenize query
|
||||
query_tokens = self._process_query(query)
|
||||
logger.debug(f"Query tokens: {query_tokens}")
|
||||
|
||||
# Build Qdrant filter
|
||||
filter_conditions = [
|
||||
FieldCondition(key="user_id", match=MatchValue(value=user_id))
|
||||
]
|
||||
if doc_type:
|
||||
filter_conditions.append(
|
||||
FieldCondition(key="doc_type", match=MatchValue(value=doc_type))
|
||||
)
|
||||
|
||||
# Scroll through Qdrant to get all matching documents
|
||||
# We need title and excerpt from payload for token matching
|
||||
qdrant_client = await get_qdrant_client()
|
||||
collection = settings.get_collection_name()
|
||||
|
||||
all_points = []
|
||||
offset = None
|
||||
|
||||
# Scroll through all points matching filter
|
||||
while True:
|
||||
scroll_result, next_offset = await qdrant_client.scroll(
|
||||
collection_name=collection,
|
||||
scroll_filter=Filter(must=filter_conditions),
|
||||
limit=100, # Batch size
|
||||
offset=offset,
|
||||
with_payload=[
|
||||
"doc_id",
|
||||
"doc_type",
|
||||
"title",
|
||||
"excerpt",
|
||||
"chunk_index",
|
||||
"total_chunks",
|
||||
],
|
||||
with_vectors=False, # Don't need vectors for keyword search
|
||||
)
|
||||
|
||||
all_points.extend(scroll_result)
|
||||
|
||||
if next_offset is None:
|
||||
break
|
||||
offset = next_offset
|
||||
|
||||
logger.debug(
|
||||
f"Retrieved {len(all_points)} points from Qdrant for keyword search"
|
||||
)
|
||||
|
||||
# Deduplicate by (doc_id, doc_type) - keep best chunk per document
|
||||
seen_docs = {}
|
||||
for point in all_points:
|
||||
doc_id = int(point.payload["doc_id"])
|
||||
dtype = point.payload.get("doc_type", "note")
|
||||
doc_key = (doc_id, dtype)
|
||||
|
||||
# Keep first chunk (chunk_index=0) as it has the most relevant content
|
||||
chunk_idx = point.payload.get("chunk_index", 0)
|
||||
if doc_key not in seen_docs or chunk_idx == 0:
|
||||
seen_docs[doc_key] = point
|
||||
|
||||
logger.debug(f"Deduplicated to {len(seen_docs)} unique documents")
|
||||
|
||||
# Score each document based on keyword matches
|
||||
scored_results = []
|
||||
for doc_key, point in seen_docs.items():
|
||||
doc_id, dtype = doc_key
|
||||
title = point.payload.get("title", "")
|
||||
excerpt = point.payload.get("excerpt", "")
|
||||
|
||||
# Calculate keyword match score
|
||||
score = self._calculate_score(query_tokens, title, excerpt)
|
||||
|
||||
if score > 0: # Only include matches
|
||||
scored_results.append(
|
||||
{
|
||||
"doc_id": doc_id,
|
||||
"doc_type": dtype,
|
||||
"title": title,
|
||||
"excerpt": excerpt,
|
||||
"score": score,
|
||||
}
|
||||
)
|
||||
|
||||
# Sort by score (descending) and limit
|
||||
scored_results.sort(key=lambda x: x["score"], reverse=True)
|
||||
top_results = scored_results[:limit]
|
||||
|
||||
# Return unverified results (verification happens at output stage)
|
||||
final_results = []
|
||||
for result in top_results:
|
||||
final_results.append(
|
||||
SearchResult(
|
||||
id=result["doc_id"],
|
||||
doc_type=result["doc_type"],
|
||||
title=result["title"],
|
||||
excerpt=result["excerpt"],
|
||||
score=result["score"],
|
||||
metadata={},
|
||||
)
|
||||
)
|
||||
|
||||
logger.info(f"Keyword search returned {len(final_results)} unverified results")
|
||||
if final_results:
|
||||
result_details = [
|
||||
f"{r.doc_type}_{r.id} (score={r.score:.3f}, title='{r.title}')"
|
||||
for r in final_results[:5]
|
||||
]
|
||||
logger.debug(f"Top keyword results: {', '.join(result_details)}")
|
||||
|
||||
return final_results
|
||||
|
||||
def _process_query(self, query: str) -> list[str]:
|
||||
"""Tokenize and normalize query.
|
||||
|
||||
Args:
|
||||
query: Raw query string
|
||||
|
||||
Returns:
|
||||
List of normalized tokens
|
||||
"""
|
||||
# Convert to lowercase and split into tokens
|
||||
tokens = query.lower().split()
|
||||
|
||||
# Filter out very short tokens (optional)
|
||||
tokens = [token for token in tokens if len(token) > 1]
|
||||
|
||||
return tokens
|
||||
|
||||
def _calculate_score(
|
||||
self, query_tokens: list[str], title: str, content: str
|
||||
) -> float:
|
||||
"""Calculate relevance score based on token matches.
|
||||
|
||||
Args:
|
||||
query_tokens: List of query tokens
|
||||
title: Document title
|
||||
content: Document content
|
||||
|
||||
Returns:
|
||||
Relevance score (0.0-1.0)
|
||||
"""
|
||||
if not query_tokens:
|
||||
return 0.0
|
||||
|
||||
# Process title and content
|
||||
title_tokens = title.lower().split()
|
||||
content_tokens = content.lower().split()
|
||||
|
||||
score = 0.0
|
||||
|
||||
# Count matches in title
|
||||
title_matches = sum(1 for qt in query_tokens if qt in title_tokens)
|
||||
if query_tokens: # Avoid division by zero
|
||||
title_match_ratio = title_matches / len(query_tokens)
|
||||
score += self.TITLE_WEIGHT * title_match_ratio
|
||||
|
||||
# Count matches in content
|
||||
content_matches = sum(1 for qt in query_tokens if qt in content_tokens)
|
||||
if query_tokens:
|
||||
content_match_ratio = content_matches / len(query_tokens)
|
||||
score += self.CONTENT_WEIGHT * content_match_ratio
|
||||
|
||||
# Normalize score to 0-1 range
|
||||
# Max score would be TITLE_WEIGHT + CONTENT_WEIGHT if all tokens match everywhere
|
||||
max_score = self.TITLE_WEIGHT + self.CONTENT_WEIGHT
|
||||
normalized_score = min(score / max_score, 1.0)
|
||||
|
||||
return normalized_score
|
||||
|
||||
def _extract_excerpt(
|
||||
self, content: str, query_tokens: list[str], max_length: int = 200
|
||||
) -> str:
|
||||
"""Extract excerpt showing match context.
|
||||
|
||||
Args:
|
||||
content: Full document content
|
||||
query_tokens: Query tokens to find
|
||||
max_length: Maximum excerpt length in characters
|
||||
|
||||
Returns:
|
||||
Excerpt string with context around matches
|
||||
"""
|
||||
if not content:
|
||||
return ""
|
||||
|
||||
content_lower = content.lower()
|
||||
|
||||
# Find first occurrence of any query token
|
||||
first_match_pos = -1
|
||||
for token in query_tokens:
|
||||
pos = content_lower.find(token)
|
||||
if pos != -1:
|
||||
if first_match_pos == -1 or pos < first_match_pos:
|
||||
first_match_pos = pos
|
||||
|
||||
if first_match_pos == -1:
|
||||
# No matches found, return beginning
|
||||
return content[:max_length].strip() + (
|
||||
"..." if len(content) > max_length else ""
|
||||
)
|
||||
|
||||
# Extract context around match
|
||||
start = max(0, first_match_pos - max_length // 2)
|
||||
end = min(len(content), first_match_pos + max_length // 2)
|
||||
|
||||
excerpt = content[start:end].strip()
|
||||
|
||||
# Add ellipsis if truncated
|
||||
if start > 0:
|
||||
excerpt = "..." + excerpt
|
||||
if end < len(content):
|
||||
excerpt = excerpt + "..."
|
||||
|
||||
return excerpt
|
||||
@@ -0,0 +1,166 @@
|
||||
"""Semantic search algorithm using vector similarity (Qdrant)."""
|
||||
|
||||
import logging
|
||||
from typing import Any
|
||||
|
||||
from qdrant_client.models import FieldCondition, Filter, MatchValue
|
||||
|
||||
from nextcloud_mcp_server.config import get_settings
|
||||
from nextcloud_mcp_server.embedding import get_embedding_service
|
||||
from nextcloud_mcp_server.observability.metrics import record_qdrant_operation
|
||||
from nextcloud_mcp_server.search.algorithms import SearchAlgorithm, SearchResult
|
||||
from nextcloud_mcp_server.vector.qdrant_client import get_qdrant_client
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
class SemanticSearchAlgorithm(SearchAlgorithm):
|
||||
"""Semantic search using vector similarity in Qdrant.
|
||||
|
||||
Searches documents by meaning rather than exact keywords using
|
||||
768-dimensional embeddings and cosine distance.
|
||||
"""
|
||||
|
||||
def __init__(self, score_threshold: float = 0.7):
|
||||
"""Initialize semantic search algorithm.
|
||||
|
||||
Args:
|
||||
score_threshold: Minimum similarity score (0-1, default: 0.7)
|
||||
"""
|
||||
self.score_threshold = score_threshold
|
||||
|
||||
@property
|
||||
def name(self) -> str:
|
||||
return "semantic"
|
||||
|
||||
@property
|
||||
def requires_vector_db(self) -> bool:
|
||||
return True
|
||||
|
||||
async def search(
|
||||
self,
|
||||
query: str,
|
||||
user_id: str,
|
||||
limit: int = 10,
|
||||
doc_type: str | None = None,
|
||||
**kwargs: Any,
|
||||
) -> list[SearchResult]:
|
||||
"""Execute semantic search using vector similarity.
|
||||
|
||||
Returns unverified results from Qdrant. Access verification should be
|
||||
performed separately at the final output stage using verify_search_results().
|
||||
|
||||
Args:
|
||||
query: Natural language search query
|
||||
user_id: User ID for filtering
|
||||
limit: Maximum results to return
|
||||
doc_type: Optional document type filter
|
||||
**kwargs: Additional parameters (score_threshold override)
|
||||
|
||||
Returns:
|
||||
List of unverified SearchResult objects ranked by similarity score
|
||||
|
||||
Raises:
|
||||
McpError: If vector sync is not enabled or search fails
|
||||
"""
|
||||
settings = get_settings()
|
||||
score_threshold = kwargs.get("score_threshold", self.score_threshold)
|
||||
|
||||
logger.info(
|
||||
f"Semantic search: query='{query}', user={user_id}, "
|
||||
f"limit={limit}, score_threshold={score_threshold}, doc_type={doc_type}"
|
||||
)
|
||||
|
||||
# Generate embedding for query
|
||||
embedding_service = get_embedding_service()
|
||||
query_embedding = await embedding_service.embed(query)
|
||||
logger.debug(
|
||||
f"Generated embedding for query (dimension={len(query_embedding)})"
|
||||
)
|
||||
|
||||
# Build Qdrant filter
|
||||
filter_conditions = [
|
||||
FieldCondition(
|
||||
key="user_id",
|
||||
match=MatchValue(value=user_id),
|
||||
)
|
||||
]
|
||||
|
||||
# Add doc_type filter if specified
|
||||
if doc_type:
|
||||
filter_conditions.append(
|
||||
FieldCondition(
|
||||
key="doc_type",
|
||||
match=MatchValue(value=doc_type),
|
||||
)
|
||||
)
|
||||
|
||||
# Search Qdrant
|
||||
qdrant_client = await get_qdrant_client()
|
||||
try:
|
||||
search_response = await qdrant_client.query_points(
|
||||
collection_name=settings.get_collection_name(),
|
||||
query=query_embedding,
|
||||
query_filter=Filter(must=filter_conditions),
|
||||
limit=limit * 2, # Get extra for deduplication
|
||||
score_threshold=score_threshold,
|
||||
with_payload=True,
|
||||
with_vectors=False, # Don't return vectors to save bandwidth
|
||||
)
|
||||
record_qdrant_operation("search", "success")
|
||||
except Exception:
|
||||
record_qdrant_operation("search", "error")
|
||||
raise
|
||||
|
||||
logger.info(
|
||||
f"Qdrant returned {len(search_response.points)} results "
|
||||
f"(before deduplication)"
|
||||
)
|
||||
|
||||
if search_response.points:
|
||||
# Log top 3 scores to help with threshold tuning
|
||||
top_scores = [p.score for p in search_response.points[:3]]
|
||||
logger.debug(f"Top 3 similarity scores: {top_scores}")
|
||||
|
||||
# Deduplicate by (doc_id, doc_type) - multiple chunks per document
|
||||
seen_docs = set()
|
||||
results = []
|
||||
|
||||
for result in search_response.points:
|
||||
doc_id = int(result.payload["doc_id"])
|
||||
doc_type = result.payload.get("doc_type", "note")
|
||||
doc_key = (doc_id, doc_type)
|
||||
|
||||
# Skip if we've already seen this document
|
||||
if doc_key in seen_docs:
|
||||
continue
|
||||
|
||||
seen_docs.add(doc_key)
|
||||
|
||||
# Return unverified results (verification happens at output stage)
|
||||
results.append(
|
||||
SearchResult(
|
||||
id=doc_id,
|
||||
doc_type=doc_type,
|
||||
title=result.payload.get("title", "Untitled"),
|
||||
excerpt=result.payload.get("excerpt", ""),
|
||||
score=result.score,
|
||||
metadata={
|
||||
"chunk_index": result.payload.get("chunk_index"),
|
||||
"total_chunks": result.payload.get("total_chunks"),
|
||||
},
|
||||
)
|
||||
)
|
||||
|
||||
if len(results) >= limit:
|
||||
break
|
||||
|
||||
logger.info(f"Returning {len(results)} unverified results after deduplication")
|
||||
if results:
|
||||
result_details = [
|
||||
f"{r.doc_type}_{r.id} (score={r.score:.3f}, title='{r.title}')"
|
||||
for r in results[:5] # Show top 5
|
||||
]
|
||||
logger.debug(f"Top results: {', '.join(result_details)}")
|
||||
|
||||
return results
|
||||
@@ -0,0 +1,122 @@
|
||||
"""Access verification for search results.
|
||||
|
||||
This module provides centralized verification of Nextcloud access permissions
|
||||
for search results. Verification happens at the final output stage (MCP tool/viz endpoint)
|
||||
rather than within individual search algorithms, preventing redundant API calls.
|
||||
|
||||
Key benefits:
|
||||
- Deduplication: Each document verified exactly once (even in hybrid mode)
|
||||
- Parallel execution: All verifications run concurrently via anyio task groups
|
||||
- Separation of concerns: Algorithms handle scoring, this module handles security
|
||||
"""
|
||||
|
||||
import logging
|
||||
from dataclasses import replace
|
||||
from typing import Protocol
|
||||
|
||||
import anyio
|
||||
|
||||
from nextcloud_mcp_server.search.algorithms import SearchResult
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
class NextcloudClientProtocol(Protocol):
|
||||
"""Protocol for Nextcloud client with app-specific access."""
|
||||
|
||||
@property
|
||||
def notes(self):
|
||||
"""Notes client for accessing notes API."""
|
||||
...
|
||||
|
||||
|
||||
async def verify_search_results(
|
||||
results: list[SearchResult],
|
||||
nextcloud_client: NextcloudClientProtocol,
|
||||
) -> list[SearchResult]:
|
||||
"""
|
||||
Verify Nextcloud access for search results.
|
||||
|
||||
Deduplicates by (doc_id, doc_type), verifies in parallel using anyio task groups,
|
||||
and filters out inaccessible documents. Maintains original result ordering.
|
||||
|
||||
Args:
|
||||
results: Unverified search results from Qdrant
|
||||
nextcloud_client: Nextcloud client for access checks
|
||||
|
||||
Returns:
|
||||
Verified and accessible results (same order as input)
|
||||
|
||||
Example:
|
||||
>>> unverified = await search_algo.search(query="test", limit=10)
|
||||
>>> verified = await verify_search_results(unverified, client)
|
||||
>>> # verified contains only documents user can access
|
||||
"""
|
||||
# Deduplicate by (doc_id, doc_type) while preserving order
|
||||
# This is critical for hybrid search where same doc may appear in multiple algorithm results
|
||||
seen = set()
|
||||
unique_results = []
|
||||
for result in results:
|
||||
key = (result.id, result.doc_type)
|
||||
if key not in seen:
|
||||
seen.add(key)
|
||||
unique_results.append(result)
|
||||
|
||||
if not unique_results:
|
||||
return []
|
||||
|
||||
logger.debug(
|
||||
f"Verifying access for {len(unique_results)} unique documents "
|
||||
f"(from {len(results)} total results)"
|
||||
)
|
||||
|
||||
# Verify all unique documents in parallel using anyio task group
|
||||
# Use list to maintain order (index-based storage)
|
||||
verified_results = [None] * len(unique_results)
|
||||
|
||||
async def verify_one(index: int, result: SearchResult):
|
||||
"""
|
||||
Verify a single document and store result at index.
|
||||
|
||||
Args:
|
||||
index: Position in verified_results list
|
||||
result: Search result to verify
|
||||
"""
|
||||
try:
|
||||
if result.doc_type == "note":
|
||||
# Fetch note to verify access and get fresh metadata
|
||||
note = await nextcloud_client.notes.get_note(result.id)
|
||||
# Update metadata with fresh data from Nextcloud
|
||||
updated_metadata = {**(result.metadata or {}), **note}
|
||||
verified_results[index] = replace(result, metadata=updated_metadata)
|
||||
# TODO: Add verification for other doc types (calendar, deck, file, etc.)
|
||||
else:
|
||||
# For now, assume other types are accessible
|
||||
# In production, add proper verification for each type
|
||||
logger.debug(
|
||||
f"No verification implemented for doc_type={result.doc_type}, "
|
||||
"assuming accessible"
|
||||
)
|
||||
verified_results[index] = result
|
||||
|
||||
except Exception as e:
|
||||
# Document is inaccessible (403, 404, or other error)
|
||||
# Log at debug level since this is expected for filtered results
|
||||
logger.debug(f"Document {result.doc_type}/{result.id} not accessible: {e}")
|
||||
verified_results[index] = None
|
||||
|
||||
# Run all verifications in parallel using anyio task group
|
||||
# This provides structured concurrency with automatic cancellation on errors
|
||||
async with anyio.create_task_group() as tg:
|
||||
for idx, result in enumerate(unique_results):
|
||||
tg.start_soon(verify_one, idx, result)
|
||||
|
||||
# Filter out None (inaccessible) and return verified results
|
||||
accessible = [r for r in verified_results if r is not None]
|
||||
|
||||
logger.debug(
|
||||
f"Verification complete: {len(accessible)} accessible, "
|
||||
f"{len(unique_results) - len(accessible)} filtered out"
|
||||
)
|
||||
|
||||
return accessible
|
||||
@@ -1,8 +1,9 @@
|
||||
"""Semantic search MCP tools using vector database."""
|
||||
|
||||
import logging
|
||||
from typing import Literal
|
||||
|
||||
from httpx import HTTPStatusError, RequestError
|
||||
from httpx import RequestError
|
||||
from mcp.server.fastmcp import Context, FastMCP
|
||||
from mcp.shared.exceptions import McpError
|
||||
from mcp.types import (
|
||||
@@ -23,7 +24,12 @@ from nextcloud_mcp_server.models.semantic import (
|
||||
)
|
||||
from nextcloud_mcp_server.observability.metrics import (
|
||||
instrument_tool,
|
||||
record_qdrant_operation,
|
||||
)
|
||||
from nextcloud_mcp_server.search import (
|
||||
FuzzySearchAlgorithm,
|
||||
HybridSearchAlgorithm,
|
||||
KeywordSearchAlgorithm,
|
||||
SemanticSearchAlgorithm,
|
||||
)
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
@@ -36,187 +42,171 @@ def configure_semantic_tools(mcp: FastMCP):
|
||||
@require_scopes("semantic:read")
|
||||
@instrument_tool
|
||||
async def nc_semantic_search(
|
||||
query: str, ctx: Context, limit: int = 10, score_threshold: float = 0.7
|
||||
query: str,
|
||||
ctx: Context,
|
||||
limit: int = 10,
|
||||
doc_types: list[str] | None = None,
|
||||
score_threshold: float = 0.7,
|
||||
algorithm: Literal["semantic", "keyword", "fuzzy", "hybrid"] = "hybrid",
|
||||
semantic_weight: float = 0.5,
|
||||
keyword_weight: float = 0.3,
|
||||
fuzzy_weight: float = 0.2,
|
||||
) -> SemanticSearchResponse:
|
||||
"""
|
||||
Semantic search across all indexed Nextcloud apps using vector embeddings.
|
||||
Search Nextcloud content using configurable algorithms with cross-app support.
|
||||
|
||||
Searches documents by meaning rather than exact keywords across notes, calendar
|
||||
events, deck cards, files, and contacts. Requires vector database synchronization
|
||||
to be enabled (VECTOR_SYNC_ENABLED=true).
|
||||
Supports multiple search algorithms with client-configurable weighting:
|
||||
- semantic: Vector similarity search (requires VECTOR_SYNC_ENABLED=true)
|
||||
- keyword: Token-based matching (title matches weighted 3x)
|
||||
- fuzzy: Character overlap matching (typo-tolerant)
|
||||
- hybrid: Combines all algorithms using Reciprocal Rank Fusion (default)
|
||||
|
||||
Document types are queried from the vector database to determine what's
|
||||
actually indexed. Currently only "note" documents are fully supported.
|
||||
|
||||
Args:
|
||||
query: Natural language search query
|
||||
limit: Maximum number of results to return (default: 10)
|
||||
score_threshold: Minimum similarity score (0-1, default: 0.7)
|
||||
doc_types: Document types to search (e.g., ["note", "file"]). None = search all indexed types (default)
|
||||
score_threshold: Minimum similarity score for semantic/hybrid (0-1, default: 0.7)
|
||||
algorithm: Search algorithm to use (default: "hybrid")
|
||||
semantic_weight: Weight for semantic results in hybrid mode (default: 0.5)
|
||||
keyword_weight: Weight for keyword results in hybrid mode (default: 0.3)
|
||||
fuzzy_weight: Weight for fuzzy results in hybrid mode (default: 0.2)
|
||||
|
||||
Returns:
|
||||
SemanticSearchResponse with matching documents and similarity scores
|
||||
SemanticSearchResponse with matching documents and relevance scores
|
||||
"""
|
||||
from qdrant_client.models import FieldCondition, Filter, MatchValue
|
||||
|
||||
from nextcloud_mcp_server.config import get_settings
|
||||
from nextcloud_mcp_server.embedding import get_embedding_service
|
||||
from nextcloud_mcp_server.vector.qdrant_client import get_qdrant_client
|
||||
|
||||
settings = get_settings()
|
||||
|
||||
# Check if vector sync is enabled
|
||||
if not settings.vector_sync_enabled:
|
||||
raise McpError(
|
||||
ErrorData(
|
||||
code=-1,
|
||||
message="Semantic search is not enabled. Set VECTOR_SYNC_ENABLED=true and ensure vector database is configured.",
|
||||
)
|
||||
)
|
||||
|
||||
client = await get_client(ctx)
|
||||
username = client.username
|
||||
|
||||
logger.info(
|
||||
f"Semantic search: query='{query}', user={username}, "
|
||||
f"Search: query='{query}', user={username}, algorithm={algorithm}, "
|
||||
f"limit={limit}, score_threshold={score_threshold}"
|
||||
)
|
||||
|
||||
try:
|
||||
# Generate embedding for query
|
||||
embedding_service = get_embedding_service()
|
||||
query_embedding = await embedding_service.embed(query)
|
||||
logger.debug(
|
||||
f"Generated embedding for query (dimension={len(query_embedding)})"
|
||||
)
|
||||
|
||||
# Search Qdrant with user filtering
|
||||
# Note: Currently only searching notes (doc_type="note")
|
||||
# Future: Remove doc_type filter to search all apps
|
||||
qdrant_client = await get_qdrant_client()
|
||||
try:
|
||||
search_response = await qdrant_client.query_points(
|
||||
collection_name=settings.get_collection_name(),
|
||||
query=query_embedding,
|
||||
query_filter=Filter(
|
||||
must=[
|
||||
FieldCondition(
|
||||
key="user_id",
|
||||
match=MatchValue(value=username),
|
||||
),
|
||||
FieldCondition(
|
||||
key="doc_type",
|
||||
match=MatchValue(value="note"),
|
||||
),
|
||||
]
|
||||
),
|
||||
limit=limit * 2, # Get extra for filtering
|
||||
score_threshold=score_threshold,
|
||||
with_payload=True,
|
||||
with_vectors=False, # Don't return vectors to save bandwidth
|
||||
)
|
||||
# Record successful search operation
|
||||
record_qdrant_operation("search", "success")
|
||||
except Exception:
|
||||
# Record failed search operation
|
||||
record_qdrant_operation("search", "error")
|
||||
raise
|
||||
|
||||
logger.info(
|
||||
f"Qdrant returned {len(search_response.points)} results "
|
||||
f"(before deduplication and access verification)"
|
||||
)
|
||||
if search_response.points:
|
||||
# Log top 3 scores to help with threshold tuning
|
||||
top_scores = [p.score for p in search_response.points[:3]]
|
||||
logger.debug(f"Top 3 similarity scores: {top_scores}")
|
||||
|
||||
# Deduplicate by document ID (multiple chunks per document)
|
||||
seen_doc_ids = set()
|
||||
results = []
|
||||
|
||||
for result in search_response.points:
|
||||
doc_id = int(result.payload["doc_id"])
|
||||
doc_type = result.payload.get("doc_type", "note")
|
||||
|
||||
# Skip if we've already seen this document
|
||||
if doc_id in seen_doc_ids:
|
||||
continue
|
||||
|
||||
seen_doc_ids.add(doc_id)
|
||||
|
||||
# Verify access via Nextcloud API (dual-phase authorization)
|
||||
# Currently only supports notes, will be extended to other apps
|
||||
if doc_type == "note":
|
||||
try:
|
||||
note = await client.notes.get_note(doc_id)
|
||||
|
||||
results.append(
|
||||
SemanticSearchResult(
|
||||
id=doc_id,
|
||||
doc_type="note",
|
||||
title=result.payload["title"],
|
||||
category=note.get("category", ""),
|
||||
excerpt=result.payload["excerpt"],
|
||||
score=result.score,
|
||||
chunk_index=result.payload["chunk_index"],
|
||||
total_chunks=result.payload["total_chunks"],
|
||||
)
|
||||
# Create appropriate algorithm instance
|
||||
if algorithm == "semantic":
|
||||
if not settings.vector_sync_enabled:
|
||||
raise McpError(
|
||||
ErrorData(
|
||||
code=-1,
|
||||
message="Semantic search requires VECTOR_SYNC_ENABLED=true",
|
||||
)
|
||||
)
|
||||
search_algo = SemanticSearchAlgorithm(score_threshold=score_threshold)
|
||||
elif algorithm == "keyword":
|
||||
search_algo = KeywordSearchAlgorithm()
|
||||
elif algorithm == "fuzzy":
|
||||
search_algo = FuzzySearchAlgorithm()
|
||||
elif algorithm == "hybrid":
|
||||
if semantic_weight > 0 and not settings.vector_sync_enabled:
|
||||
raise McpError(
|
||||
ErrorData(
|
||||
code=-1,
|
||||
message="Hybrid search with semantic component requires VECTOR_SYNC_ENABLED=true",
|
||||
)
|
||||
)
|
||||
search_algo = HybridSearchAlgorithm(
|
||||
semantic_weight=semantic_weight,
|
||||
keyword_weight=keyword_weight,
|
||||
fuzzy_weight=fuzzy_weight,
|
||||
)
|
||||
else:
|
||||
raise McpError(
|
||||
ErrorData(code=-1, message=f"Unknown algorithm: {algorithm}")
|
||||
)
|
||||
|
||||
if len(results) >= limit:
|
||||
break
|
||||
# Execute search across requested document types
|
||||
# If doc_types is None, search all indexed types (cross-app search)
|
||||
# If doc_types is a list, search only those types
|
||||
all_results = []
|
||||
|
||||
except HTTPStatusError as e:
|
||||
if e.response.status_code == 403:
|
||||
# User lost access, skip this document
|
||||
logger.debug(f"Skipping note {doc_id}: access denied (403)")
|
||||
continue
|
||||
elif e.response.status_code == 404:
|
||||
# Document was deleted but not yet removed from vector DB
|
||||
logger.debug(
|
||||
f"Skipping note {doc_id}: not found (404), "
|
||||
f"likely deleted after indexing"
|
||||
)
|
||||
continue
|
||||
else:
|
||||
# Log other errors but continue processing
|
||||
logger.warning(
|
||||
f"Error verifying access to note {doc_id}: {e.response.status_code}"
|
||||
)
|
||||
continue
|
||||
if doc_types is None:
|
||||
# Cross-app search: search all indexed types
|
||||
# Get unverified results from Qdrant
|
||||
unverified_results = await search_algo.search(
|
||||
query=query,
|
||||
user_id=username,
|
||||
limit=limit * 2, # Get extra for access filtering
|
||||
doc_type=None, # Signal to search all types
|
||||
score_threshold=score_threshold,
|
||||
)
|
||||
all_results.extend(unverified_results)
|
||||
else:
|
||||
# Search specific document types
|
||||
# For each requested type, execute search and combine results
|
||||
for dtype in doc_types:
|
||||
unverified_results = await search_algo.search(
|
||||
query=query,
|
||||
user_id=username,
|
||||
limit=limit * 2, # Get extra for combining and filtering
|
||||
doc_type=dtype,
|
||||
score_threshold=score_threshold,
|
||||
)
|
||||
all_results.extend(unverified_results)
|
||||
|
||||
logger.info(
|
||||
f"Returning {len(results)} results after deduplication and access verification"
|
||||
)
|
||||
if results:
|
||||
result_details = [
|
||||
f"note_{r.id} (score={r.score:.3f}, title='{r.title}')"
|
||||
for r in results[:5] # Show top 5
|
||||
]
|
||||
logger.debug(f"Top results: {', '.join(result_details)}")
|
||||
# Sort combined results by score
|
||||
all_results.sort(key=lambda r: r.score, reverse=True)
|
||||
|
||||
# Verify access for all results (deduplicates and filters)
|
||||
from nextcloud_mcp_server.search.verification import verify_search_results
|
||||
|
||||
verified_results = await verify_search_results(all_results, client)
|
||||
search_results = verified_results[:limit] # Final limit after verification
|
||||
|
||||
# Convert SearchResult objects to SemanticSearchResult for response
|
||||
results = []
|
||||
for r in search_results:
|
||||
results.append(
|
||||
SemanticSearchResult(
|
||||
id=r.id,
|
||||
doc_type=r.doc_type,
|
||||
title=r.title,
|
||||
category=r.metadata.get("category", "") if r.metadata else "",
|
||||
excerpt=r.excerpt,
|
||||
score=r.score,
|
||||
chunk_index=r.metadata.get("chunk_index", 0)
|
||||
if r.metadata
|
||||
else 0,
|
||||
total_chunks=r.metadata.get("total_chunks", 1)
|
||||
if r.metadata
|
||||
else 1,
|
||||
)
|
||||
)
|
||||
|
||||
logger.info(f"Returning {len(results)} results from {algorithm} search")
|
||||
|
||||
return SemanticSearchResponse(
|
||||
results=results,
|
||||
query=query,
|
||||
total_found=len(results),
|
||||
search_method="semantic",
|
||||
search_method=algorithm,
|
||||
)
|
||||
|
||||
except ValueError as e:
|
||||
if "No embedding provider configured" in str(e):
|
||||
error_msg = str(e)
|
||||
if "No embedding provider configured" in error_msg:
|
||||
raise McpError(
|
||||
ErrorData(
|
||||
code=-1,
|
||||
message="Embedding service not configured. Set OLLAMA_BASE_URL environment variable.",
|
||||
)
|
||||
)
|
||||
raise McpError(ErrorData(code=-1, message=f"Configuration error: {str(e)}"))
|
||||
raise McpError(
|
||||
ErrorData(code=-1, message=f"Configuration error: {error_msg}")
|
||||
)
|
||||
except RequestError as e:
|
||||
raise McpError(
|
||||
ErrorData(code=-1, message=f"Network error during search: {str(e)}")
|
||||
)
|
||||
except Exception as e:
|
||||
logger.error(f"Semantic search error: {e}", exc_info=True)
|
||||
raise McpError(
|
||||
ErrorData(code=-1, message=f"Semantic search failed: {str(e)}")
|
||||
)
|
||||
logger.error(f"Search error: {e}", exc_info=True)
|
||||
raise McpError(ErrorData(code=-1, message=f"Search failed: {str(e)}"))
|
||||
|
||||
@mcp.tool()
|
||||
@require_scopes("semantic:read")
|
||||
|
||||
@@ -0,0 +1,140 @@
|
||||
"""Custom PCA implementation for dimensionality reduction.
|
||||
|
||||
Implements Principal Component Analysis without scikit-learn dependency.
|
||||
Used for reducing high-dimensional embeddings (768-dim) to 2D for visualization.
|
||||
"""
|
||||
|
||||
import logging
|
||||
|
||||
import numpy as np
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
class PCA:
|
||||
"""Principal Component Analysis for dimensionality reduction.
|
||||
|
||||
Simple implementation that finds principal components via eigendecomposition
|
||||
of the covariance matrix. Suitable for small-to-medium datasets.
|
||||
|
||||
Attributes:
|
||||
n_components: Number of principal components to keep
|
||||
mean_: Mean of training data (set during fit)
|
||||
components_: Principal components (eigenvectors)
|
||||
explained_variance_: Variance explained by each component
|
||||
explained_variance_ratio_: Fraction of total variance explained
|
||||
"""
|
||||
|
||||
def __init__(self, n_components: int = 2):
|
||||
"""Initialize PCA.
|
||||
|
||||
Args:
|
||||
n_components: Number of components to keep (default: 2)
|
||||
"""
|
||||
if n_components < 1:
|
||||
raise ValueError(f"n_components must be >= 1, got {n_components}")
|
||||
|
||||
self.n_components = n_components
|
||||
self.mean_: np.ndarray | None = None
|
||||
self.components_: np.ndarray | None = None
|
||||
self.explained_variance_: np.ndarray | None = None
|
||||
self.explained_variance_ratio_: np.ndarray | None = None
|
||||
|
||||
def fit(self, X: np.ndarray) -> "PCA":
|
||||
"""Fit PCA model to data.
|
||||
|
||||
Args:
|
||||
X: Training data of shape (n_samples, n_features)
|
||||
|
||||
Returns:
|
||||
self (for method chaining)
|
||||
|
||||
Raises:
|
||||
ValueError: If X has fewer features than n_components
|
||||
"""
|
||||
X = np.asarray(X)
|
||||
|
||||
if X.ndim != 2:
|
||||
raise ValueError(f"X must be 2D array, got shape {X.shape}")
|
||||
|
||||
n_samples, n_features = X.shape
|
||||
|
||||
if n_features < self.n_components:
|
||||
raise ValueError(
|
||||
f"n_components={self.n_components} > n_features={n_features}"
|
||||
)
|
||||
|
||||
# Center data
|
||||
self.mean_ = np.mean(X, axis=0)
|
||||
X_centered = X - self.mean_
|
||||
|
||||
# Compute covariance matrix
|
||||
# Use (X^T X) / (n-1) for numerical stability with high-dim data
|
||||
cov = np.cov(X_centered.T)
|
||||
|
||||
# Eigendecomposition
|
||||
eigenvalues, eigenvectors = np.linalg.eigh(cov)
|
||||
|
||||
# Sort by eigenvalue (descending)
|
||||
idx = np.argsort(eigenvalues)[::-1]
|
||||
eigenvalues = eigenvalues[idx]
|
||||
eigenvectors = eigenvectors[:, idx]
|
||||
|
||||
# Keep top n_components
|
||||
self.components_ = eigenvectors[:, : self.n_components].T
|
||||
self.explained_variance_ = eigenvalues[: self.n_components]
|
||||
|
||||
# Calculate explained variance ratio
|
||||
total_variance = np.sum(eigenvalues)
|
||||
if total_variance > 0:
|
||||
self.explained_variance_ratio_ = self.explained_variance_ / total_variance
|
||||
else:
|
||||
self.explained_variance_ratio_ = np.zeros(self.n_components)
|
||||
|
||||
logger.debug(
|
||||
f"PCA fit: {n_samples} samples, {n_features} features → "
|
||||
f"{self.n_components} components, "
|
||||
f"explained variance: {self.explained_variance_ratio_}"
|
||||
)
|
||||
|
||||
return self
|
||||
|
||||
def transform(self, X: np.ndarray) -> np.ndarray:
|
||||
"""Transform data to principal component space.
|
||||
|
||||
Args:
|
||||
X: Data to transform of shape (n_samples, n_features)
|
||||
|
||||
Returns:
|
||||
Transformed data of shape (n_samples, n_components)
|
||||
|
||||
Raises:
|
||||
ValueError: If PCA not fitted yet
|
||||
"""
|
||||
if self.mean_ is None or self.components_ is None:
|
||||
raise ValueError("PCA not fitted yet. Call fit() first.")
|
||||
|
||||
X = np.asarray(X)
|
||||
|
||||
if X.ndim != 2:
|
||||
raise ValueError(f"X must be 2D array, got shape {X.shape}")
|
||||
|
||||
# Center using training mean
|
||||
X_centered = X - self.mean_
|
||||
|
||||
# Project onto principal components
|
||||
X_transformed = np.dot(X_centered, self.components_.T)
|
||||
|
||||
return X_transformed
|
||||
|
||||
def fit_transform(self, X: np.ndarray) -> np.ndarray:
|
||||
"""Fit PCA model and transform data in one step.
|
||||
|
||||
Args:
|
||||
X: Training data of shape (n_samples, n_features)
|
||||
|
||||
Returns:
|
||||
Transformed data of shape (n_samples, n_components)
|
||||
"""
|
||||
self.fit(X)
|
||||
return self.transform(X)
|
||||
+1
-1
@@ -1,6 +1,6 @@
|
||||
[project]
|
||||
name = "nextcloud-mcp-server"
|
||||
version = "0.34.2"
|
||||
version = "0.35.0"
|
||||
description = "Model Context Protocol (MCP) server for Nextcloud integration - enables AI assistants to interact with Nextcloud data"
|
||||
authors = [
|
||||
{name = "Chris Coutinho", email = "chris@coutinho.io"}
|
||||
|
||||
+51
-3
@@ -9,6 +9,7 @@ import pytest
|
||||
from httpx import HTTPStatusError
|
||||
from mcp import ClientSession
|
||||
from mcp.client.session import RequestContext
|
||||
from mcp.client.sse import sse_client
|
||||
from mcp.client.streamable_http import streamablehttp_client
|
||||
from mcp.types import ElicitRequestParams, ElicitResult, ErrorData
|
||||
|
||||
@@ -165,6 +166,51 @@ async def create_mcp_client_session(
|
||||
logger.debug(f"{client_name} client session cleaned up successfully")
|
||||
|
||||
|
||||
async def create_mcp_client_session_sse(
|
||||
url: str,
|
||||
token: str | None = None,
|
||||
client_name: str = "MCP",
|
||||
elicitation_callback: Any = None,
|
||||
) -> AsyncGenerator[ClientSession, Any]:
|
||||
"""
|
||||
Factory function to create an MCP client session using SSE transport.
|
||||
|
||||
Similar to create_mcp_client_session but uses SSE transport instead of streamable-http.
|
||||
Uses native async context managers to ensure correct LIFO cleanup order.
|
||||
|
||||
Args:
|
||||
url: MCP server URL (e.g., "http://localhost:8000/sse")
|
||||
token: Optional OAuth access token for Bearer authentication
|
||||
client_name: Client name for logging (e.g., "Basic MCP (SSE)")
|
||||
elicitation_callback: Optional callback for handling elicitation requests
|
||||
|
||||
Yields:
|
||||
Initialized MCP ClientSession
|
||||
|
||||
Note:
|
||||
SSE transport is being deprecated in favor of streamable-http.
|
||||
This function exists for compatibility testing only.
|
||||
"""
|
||||
logger.info(f"Creating SSE client for {client_name}")
|
||||
|
||||
# Prepare headers with OAuth token if provided
|
||||
headers = {"Authorization": f"Bearer {token}"} if token else None
|
||||
|
||||
# Use native async with - Python ensures LIFO cleanup
|
||||
# Cleanup order will be: ClientSession.__aexit__ -> sse_client.__aexit__
|
||||
# Note: sse_client yields only (read_stream, write_stream), not 3 values like streamablehttp_client
|
||||
async with sse_client(url, headers=headers) as (read_stream, write_stream):
|
||||
async with ClientSession(
|
||||
read_stream, write_stream, elicitation_callback=elicitation_callback
|
||||
) as session:
|
||||
await session.initialize()
|
||||
logger.info(f"{client_name} client session initialized successfully")
|
||||
yield session
|
||||
|
||||
# Cleanup happens automatically in LIFO order - no exception suppression needed
|
||||
logger.debug(f"{client_name} client session cleaned up successfully")
|
||||
|
||||
|
||||
@pytest.fixture(scope="session")
|
||||
async def nc_client(anyio_backend) -> AsyncGenerator[NextcloudClient, Any]:
|
||||
"""
|
||||
@@ -203,12 +249,14 @@ async def nc_client(anyio_backend) -> AsyncGenerator[NextcloudClient, Any]:
|
||||
@pytest.fixture(scope="session")
|
||||
async def nc_mcp_client(anyio_backend) -> AsyncGenerator[ClientSession, Any]:
|
||||
"""
|
||||
Fixture to create an MCP client session for integration tests using streamable-http.
|
||||
Fixture to create an MCP client session for integration tests using SSE transport.
|
||||
|
||||
Uses anyio pytest plugin for proper async fixture handling.
|
||||
|
||||
Note: SSE transport is being deprecated. This fixture uses SSE for compatibility testing.
|
||||
"""
|
||||
async for session in create_mcp_client_session(
|
||||
url="http://localhost:8000/mcp", client_name="Basic MCP"
|
||||
async for session in create_mcp_client_session_sse(
|
||||
url="http://localhost:8000/sse", client_name="Basic MCP (SSE)"
|
||||
):
|
||||
yield session
|
||||
|
||||
|
||||
Vendored
+1
-1
Submodule third_party/oidc updated: e83dabbac1...9616294911
Reference in New Issue
Block a user