Update README

perf: Optimize vector viz search performance

CHANGELOG.md

@@ -1,3 +1,22 @@
+## v0.46.1 (2025-11-22)
+
+### Perf
+
+- Optimize vector viz search performance
+
+## v0.46.0 (2025-11-22)
+
+### Feat
+
+- Add Smithery CLI deployment support
+- Implement ADR-016 Smithery stateless deployment mode
+
+### Fix
+
+- **smithery**: Add JSON Schema metadata to mcp-config endpoint
+- **smithery**: Use container runtime pattern for config discovery
+- Add Smithery lifespan and auth mode detection
+
 ## v0.45.0 (2025-11-22)
 
 ### Feat

README.md

@@ -1,11 +1,12 @@
+```markdown
 <p align="center">
   <img src="astrolabe.svg" alt="Nextcloud MCP Server" width="128" height="128">
 </p>
 
 # Nextcloud MCP Server
 
-[![Docker Image](https://img.shields.io/badge/docker-ghcr.io/cbcoutinho/nextcloud--mcp--server-blue)](https://github.com/cbcoutinho/nextcloud-mcp-server/pkgs/container/nextcloud-mcp-server)
 [![smithery badge](https://smithery.ai/badge/@cbcoutinho/nextcloud-mcp-server)](https://smithery.ai/server/@cbcoutinho/nextcloud-mcp-server)
+[![Docker Image](https://img.shields.io/badge/docker-ghcr.io/cbcoutinho/nextcloud--mcp--server-blue)](https://github.com/cbcoutinho/nextcloud-mcp-server/pkgs/container/nextcloud-mcp-server)
 
 **A production-ready MCP server that connects AI assistants to your Nextcloud instance.**
 
@@ -223,3 +224,4 @@ This project is licensed under the AGPL-3.0 License. See [LICENSE](./LICENSE) fo
 - [Model Context Protocol](https://github.com/modelcontextprotocol)
 - [MCP Python SDK](https://github.com/modelcontextprotocol/python-sdk)
 - [Nextcloud](https://nextcloud.com/)
+```

charts/nextcloud-mcp-server/Chart.yaml

@@ -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: 0.45.0
-appVersion: "0.45.0"
+version: 0.46.1
+appVersion: "0.46.1"
 keywords:
   - nextcloud
   - mcp

nextcloud_mcp_server/auth/viz_routes.py

@@ -218,71 +218,41 @@ async def vector_visualization_search(request: Request) -> JSONResponse:
                 }
             )
 
-        # Fetch vectors for specific matching chunks from Qdrant
+        # Fetch vectors for specific matching chunks from Qdrant using batch retrieve
         vector_fetch_start = time.perf_counter()
         qdrant_client = await get_qdrant_client()
 
-        # Build filters for each specific chunk
-        from qdrant_client.models import FieldCondition, Filter, MatchValue
-
         chunk_vectors_map = {}  # Map (doc_id, chunk_start, chunk_end) -> vector
 
-        # Fetch vectors in batches by filtering on chunk-specific fields
-        for result in search_results:
-            chunk_start = result.chunk_start_offset
-            chunk_end = result.chunk_end_offset
+        # Collect point IDs from search results for batch retrieval
+        # point_id is the Qdrant internal ID returned by search algorithms
+        point_ids = [r.point_id for r in search_results if r.point_id]
 
-            # Build filter for this specific chunk
-            must_conditions = [
-                get_placeholder_filter(),  # Always exclude placeholders from user-facing queries
-                FieldCondition(
-                    key="doc_id",
-                    match=MatchValue(value=result.id),
-                ),
-                FieldCondition(
-                    key="user_id",
-                    match=MatchValue(value=username),
-                ),
-            ]
-
-            # Add chunk position filters if available
-            if chunk_start is not None:
-                must_conditions.append(
-                    FieldCondition(
-                        key="chunk_start_offset",
-                        match=MatchValue(value=chunk_start),
-                    )
-                )
-            if chunk_end is not None:
-                must_conditions.append(
-                    FieldCondition(
-                        key="chunk_end_offset",
-                        match=MatchValue(value=chunk_end),
-                    )
-                )
-
-            # Fetch this specific chunk vector
-            points_response = await qdrant_client.scroll(
+        if point_ids:
+            # Single batch retrieve call instead of N sequential scroll calls
+            # This is ~50x faster for 50 results (1 HTTP request vs 50)
+            points_response = await qdrant_client.retrieve(
                 collection_name=settings.get_collection_name(),
-                scroll_filter=Filter(must=must_conditions),
-                limit=1,  # Only need the first match
+                ids=point_ids,
                 with_vectors=["dense"],
-                with_payload=False,
+                with_payload=["doc_id", "chunk_start_offset", "chunk_end_offset"],
             )
 
-            points = points_response[0]
-            if points:
-                # Extract dense vector
-                point = points[0]
+            # Build chunk_vectors_map from batch response
+            for point in points_response:
                 if point.vector is not None:
-                    # If named vectors (dict), extract "dense"
+                    # Extract dense vector (handle both named and unnamed vectors)
                     if isinstance(point.vector, dict):
                         vector = point.vector.get("dense")
                     else:
                         vector = point.vector
 
-                    chunk_key = (result.id, chunk_start, chunk_end)
-                    chunk_vectors_map[chunk_key] = vector
+                    if vector is not None and point.payload:
+                        doc_id = point.payload.get("doc_id")
+                        chunk_start = point.payload.get("chunk_start_offset")
+                        chunk_end = point.payload.get("chunk_end_offset")
+                        chunk_key = (doc_id, chunk_start, chunk_end)
+                        chunk_vectors_map[chunk_key] = vector
 
         vector_fetch_duration = time.perf_counter() - vector_fetch_start
 
@@ -341,16 +311,23 @@ async def vector_visualization_search(request: Request) -> JSONResponse:
 
         chunk_vectors = np.array(chunk_vectors)
 
-        # Generate query embedding for visualization
+        # Reuse query embedding from search algorithm (avoids redundant embedding call)
         query_embed_start = time.perf_counter()
-        from nextcloud_mcp_server.embedding.service import get_embedding_service
+        if search_algo.query_embedding is not None:
+            query_embedding = search_algo.query_embedding
+            logger.info(
+                f"Reusing query embedding from search algorithm "
+                f"(dimension={len(query_embedding)})"
+            )
+        else:
+            # Fallback: generate embedding if not available from search
+            from nextcloud_mcp_server.embedding.service import get_embedding_service
 
-        embedding_service = get_embedding_service()
-        query_embedding = await embedding_service.embed(query)
+            embedding_service = get_embedding_service()
+            query_embedding = await embedding_service.embed(query)
+            logger.info(f"Generated query embedding (dimension={len(query_embedding)})")
         query_embed_duration = time.perf_counter() - query_embed_start
 
-        logger.info(f"Generated query embedding (dimension={len(query_embedding)})")
-
         # Combine query vector with chunk vectors for PCA
         # Query will be the last point in the array
         all_vectors = np.vstack([chunk_vectors, np.array([query_embedding])])
@@ -380,9 +357,19 @@ async def vector_visualization_search(request: Request) -> JSONResponse:
         )
 
         # Apply PCA dimensionality reduction (768-dim → 3D) on normalized vectors
+        # Run in thread pool to avoid blocking the event loop (CPU-bound)
         pca_start = time.perf_counter()
-        pca = PCA(n_components=3)
-        coords_3d = pca.fit_transform(all_vectors_normalized)
+
+        def _compute_pca(vectors: np.ndarray) -> tuple[np.ndarray, PCA]:
+            pca = PCA(n_components=3)
+            coords = pca.fit_transform(vectors)
+            return coords, pca
+
+        import anyio
+
+        coords_3d, pca = await anyio.to_thread.run_sync(  # type: ignore[attr-defined]
+            lambda: _compute_pca(all_vectors_normalized)
+        )
         pca_duration = time.perf_counter() - pca_start
 
         # After fit, these attributes are guaranteed to be set

nextcloud_mcp_server/embedding/bm25_provider.py

@@ -37,7 +37,9 @@ class BM25SparseEmbeddingProvider:
 
     def encode(self, text: str) -> dict[str, Any]:
         """
-        Generate BM25 sparse embedding for a single text.
+        Generate BM25 sparse embedding for a single text (synchronous).
+
+        Note: For async contexts, prefer encode_async() to avoid blocking the event loop.
 
         Args:
             text: Input text to encode
@@ -53,6 +55,23 @@ class BM25SparseEmbeddingProvider:
             "values": sparse_embedding.values.tolist(),
         }
 
+    async def encode_async(self, text: str) -> dict[str, Any]:
+        """
+        Generate BM25 sparse embedding for a single text (async).
+
+        Runs CPU-bound BM25 encoding in thread pool to avoid blocking the event loop.
+
+        Args:
+            text: Input text to encode
+
+        Returns:
+            Dictionary with 'indices' and 'values' keys for Qdrant sparse vector
+        """
+        import anyio
+
+        # Run CPU-bound BM25 encoding in thread pool
+        return await anyio.to_thread.run_sync(lambda: self.encode(text))  # type: ignore[attr-defined]
+
     async def encode_batch(self, texts: list[str]) -> list[dict[str, Any]]:
         """
         Generate BM25 sparse embeddings for multiple texts (batched).

nextcloud_mcp_server/search/algorithms.py

@@ -140,6 +140,7 @@ class SearchResult:
         page_number: Page number for PDF documents (None for other doc types)
         chunk_index: Zero-based index of this chunk in the document
         total_chunks: Total number of chunks in the document
+        point_id: Qdrant point ID for batch vector retrieval (None if not from Qdrant)
     """
 
     id: int
@@ -153,6 +154,7 @@ class SearchResult:
     page_number: int | None = None
     chunk_index: int = 0
     total_chunks: int = 1
+    point_id: str | None = None
 
     def __post_init__(self):
         """Validate score is non-negative.
@@ -172,8 +174,15 @@ class SearchAlgorithm(ABC):
 
     All search algorithms must implement the search() method with consistent
     interface, allowing them to be used interchangeably.
+
+    Attributes:
+        query_embedding: The query embedding generated during the last search.
+            Available after search() completes for algorithms that use embeddings.
+            Can be reused by callers to avoid redundant embedding generation.
     """
 
+    query_embedding: list[float] | None = None
+
     @abstractmethod
     async def search(
         self,

nextcloud_mcp_server/search/bm25_hybrid.py

@@ -101,11 +101,13 @@ class BM25HybridSearchAlgorithm(SearchAlgorithm):
         # Generate dense embedding for semantic search
         embedding_service = get_embedding_service()
         dense_embedding = await embedding_service.embed(query)
+        # Store for reuse by callers (e.g., viz_routes PCA visualization)
+        self.query_embedding = dense_embedding
         logger.debug(f"Generated dense embedding (dimension={len(dense_embedding)})")
 
         # Generate sparse embedding for BM25 keyword search
         bm25_service = get_bm25_service()
-        sparse_embedding = bm25_service.encode(query)
+        sparse_embedding = await bm25_service.encode_async(query)
         logger.debug(
             f"Generated sparse embedding "
             f"({len(sparse_embedding['indices'])} non-zero terms)"
@@ -218,6 +220,7 @@ class BM25HybridSearchAlgorithm(SearchAlgorithm):
                     page_number=result.payload.get("page_number"),
                     chunk_index=result.payload.get("chunk_index", 0),
                     total_chunks=result.payload.get("total_chunks", 1),
+                    point_id=str(result.id),  # Qdrant point ID for batch retrieval
                 )
             )
 

nextcloud_mcp_server/search/semantic.py

@@ -78,6 +78,8 @@ class SemanticSearchAlgorithm(SearchAlgorithm):
         # Generate embedding for query
         embedding_service = get_embedding_service()
         query_embedding = await embedding_service.embed(query)
+        # Store for reuse by callers (e.g., viz_routes PCA visualization)
+        self.query_embedding = query_embedding
         logger.debug(
             f"Generated embedding for query (dimension={len(query_embedding)})"
         )
@@ -164,6 +166,7 @@ class SemanticSearchAlgorithm(SearchAlgorithm):
                     page_number=result.payload.get("page_number"),
                     chunk_index=result.payload.get("chunk_index", 0),
                     total_chunks=result.payload.get("total_chunks", 1),
+                    point_id=str(result.id),  # Qdrant point ID for batch retrieval
                 )
             )
 

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "nextcloud-mcp-server"
-version = "0.45.0"
+version = "0.46.1"
 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"}

uv.lock

@@ -1936,7 +1936,7 @@ wheels = [
 
 [[package]]
 name = "nextcloud-mcp-server"
-version = "0.45.0"
+version = "0.46.1"
 source = { editable = "." }
 dependencies = [
     { name = "aiosqlite" },