fix: Update DCR token_type tests for OIDC app changes

The Nextcloud OIDC app has updated token_type parameter values: - Changed from "Bearer" → "opaque" for opaque tokens - Changed from "JWT" → "jwt" for JWT tokens Updated test_dcr_token_type.py to use lowercase token_type values: - token_type="jwt" for JWT-formatted tokens - token_type="opaque" for opaque/bearer tokens This fixes test failures where tests were using the old "Bearer" and "JWT" (uppercase) values which are no longer recognized by the OIDC app. Fixes test: test_dcr_respects_bearer_token_type 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
docs: Add Ollama embeddings capacity analysis and investigation
2025-10-31 22:30:58 +01:00 · 2025-10-31 03:07:44 +01:00 · 2025-10-31 02:59:53 +01:00 · 2025-10-31 02:59:44 +01:00 · 2025-10-31 02:59:35 +01:00 · 2025-10-30 20:23:52 +01:00
20 changed files with 3373 additions and 49 deletions
@@ -14,16 +14,109 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@08c6903cd8c0fde910a37f88322edcfb5dd907a8 # v5
        with:
          fetch-depth: 0

+
      - name: Configure Git
        run: |
          git config user.name "$GITHUB_ACTOR"
          git config user.email "$GITHUB_ACTOR@users.noreply.github.com"

      - name: Run chart-releaser
-        uses: helm/chart-releaser-action@v1.7.0
+        uses: helm/chart-releaser-action@cae68fefc6b5f367a0275617c9f83181ba54714f # v1.7.0
        env:
          CR_TOKEN: "${{ secrets.GITHUB_TOKEN }}"
+
+      - name: Update gh-pages with Chart README and Index
+        run: |
+          # Get the repository name
+          REPO_NAME="${GITHUB_REPOSITORY##*/}"
+          REPO_OWNER="${GITHUB_REPOSITORY%/*}"
+
+          # Switch to gh-pages branch
+          git fetch origin gh-pages
+          git checkout gh-pages
+
+          # Copy Chart README to root
+          git checkout ${GITHUB_REF#refs/tags/} -- charts/nextcloud-mcp-server/README.md
+          mv charts/nextcloud-mcp-server/README.md README.md || true
+          rm -rf charts 2>/dev/null || true
+
+          # Create index.html with installation instructions
+          cat > index.html <<'EOF'
+          <!DOCTYPE html>
+          <html lang="en">
+          <head>
+              <meta charset="UTF-8">
+              <meta name="viewport" content="width=device-width, initial-scale=1.0">
+              <title>Nextcloud MCP Server Helm Chart</title>
+              <style>
+                  body {
+                      font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, "Helvetica Neue", Arial, sans-serif;
+                      max-width: 800px;
+                      margin: 50px auto;
+                      padding: 20px;
+                      line-height: 1.6;
+                  }
+                  code {
+                      background: #f4f4f4;
+                      padding: 2px 6px;
+                      border-radius: 3px;
+                      font-family: "Monaco", "Courier New", monospace;
+                  }
+                  pre {
+                      background: #f4f4f4;
+                      padding: 15px;
+                      border-radius: 5px;
+                      overflow-x: auto;
+                  }
+                  h1, h2 { color: #0082c9; }
+                  a { color: #0082c9; text-decoration: none; }
+                  a:hover { text-decoration: underline; }
+              </style>
+          </head>
+          <body>
+              <h1>Nextcloud MCP Server Helm Chart</h1>
+
+              <p>A Helm chart for deploying the Nextcloud MCP (Model Context Protocol) Server on Kubernetes, enabling AI assistants to interact with your Nextcloud instance.</p>
+
+              <h2>Installation</h2>
+
+              <p>Add the Helm repository:</p>
+              <pre><code>helm repo add nextcloud-mcp https://REPO_OWNER.github.io/REPO_NAME/
+          helm repo update</code></pre>
+
+              <p>Install the chart:</p>
+              <pre><code>helm install nextcloud-mcp nextcloud-mcp/nextcloud-mcp-server \
+            --set nextcloud.host=https://cloud.example.com \
+            --set auth.basic.username=myuser \
+            --set auth.basic.password=mypassword</code></pre>
+
+              <h2>Documentation</h2>
+
+              <ul>
+                  <li><a href="README.md">Chart README</a> - Full documentation for the Helm chart</li>
+                  <li><a href="https://github.com/REPO_OWNER/REPO_NAME">GitHub Repository</a> - Source code and issues</li>
+                  <li><a href="index.yaml">Helm Repository Index</a> - Chart metadata</li>
+              </ul>
+
+              <h2>Quick Start</h2>
+
+              <p>See the <a href="README.md">full documentation</a> for detailed configuration options, examples, and troubleshooting guides.</p>
+
+              <hr>
+              <p><small>Generated by <a href="https://github.com/helm/chart-releaser">chart-releaser</a></small></p>
+          </body>
+          </html>
+          EOF
+
+          # Replace placeholders
+          sed -i "s/REPO_OWNER/$REPO_OWNER/g" index.html
+          sed -i "s/REPO_NAME/$REPO_NAME/g" index.html
+
+          # Commit changes
+          git add README.md index.html
+          git commit -m "Update README and index from chart release" || echo "No changes to commit"
+          git push origin gh-pages
@@ -20,7 +20,7 @@ jobs:
      - name: Checkout
        uses: actions/checkout@08c6903cd8c0fde910a37f88322edcfb5dd907a8 # v5
      - name: Install uv
-        uses: astral-sh/setup-uv@2ddd2b9cb38ad8efd50337e8ab201519a34c9f24 # v7
+        uses: astral-sh/setup-uv@85856786d1ce8acfbcc2f13a5f3fbd6b938f9f41 # v7.1.2
      - name: Install Python 3.11
        run: uv python install 3.11
      - name: Build
@@ -11,7 +11,7 @@ jobs:
    steps:
      - uses: actions/checkout@08c6903cd8c0fde910a37f88322edcfb5dd907a8 # v5.0.0
      - name: Install the latest version of uv
-        uses: astral-sh/setup-uv@2ddd2b9cb38ad8efd50337e8ab201519a34c9f24 # v7.1.1
+        uses: astral-sh/setup-uv@85856786d1ce8acfbcc2f13a5f3fbd6b938f9f41 # v7.1.2
      - name: Check format
        run: |
          uv run --frozen ruff format --diff
@@ -52,7 +52,7 @@ jobs:
          up-flags: "--build"

      - name: Install the latest version of uv
-        uses: astral-sh/setup-uv@2ddd2b9cb38ad8efd50337e8ab201519a34c9f24 # v7.1.1
+        uses: astral-sh/setup-uv@85856786d1ce8acfbcc2f13a5f3fbd6b938f9f41 # v7.1.2

      - name: Install Playwright dependencies
        run: |
@@ -1,3 +1,53 @@
+## v0.22.7 (2025-10-29)
+
+### Fix
+
+- **helm**: Remove image tag overide
+
+## v0.22.6 (2025-10-29)
+
+### Fix
+
+- **helm**: Update helm chart with extraArgs
+
+## v0.22.5 (2025-10-29)
+
+### Fix
+
+- Update helm chart variables
+
+## v0.22.4 (2025-10-29)
+
+### Fix
+
+- **helm**: Update helm version with release
+- **helm**: Update helm version with release
+
+## v0.22.3 (2025-10-29)
+
+### Fix
+
+- **helm**: Update helm version with release
+
+## v0.22.2 (2025-10-29)
+
+### Fix
+
+- **helm**: Update helm version with release
+
+## v0.22.1 (2025-10-29)
+
+### Fix
+
+- Trigger release
+
+## v0.22.0 (2025-10-29)
+
+### Feat
+
+- **server**: Add /live & /health endpoints
+- Initialize helm chart
+
 ## v0.21.0 (2025-10-25)

 ### Feat
@@ -1,4 +1,4 @@
-FROM ghcr.io/astral-sh/uv:0.9.5-python3.11-alpine@sha256:64ecec379ff82bea84b8a80c0b374f6392bcd54aa52f8c63c12f510f9c0b214d
+FROM ghcr.io/astral-sh/uv:0.9.6-python3.11-alpine@sha256:b2a366adae7002a23dbba79791baac4e607ee5af5d45039d072d30115c505666

 # Install git (required for caldav dependency from git)
 RUN apk add --no-cache git
@@ -72,9 +72,17 @@ uv sync

 # Or using Docker
 docker pull ghcr.io/cbcoutinho/nextcloud-mcp-server:latest
+
+# Or deploy to Kubernetes with Helm
+helm repo add nextcloud-mcp https://cbcoutinho.github.io/nextcloud-mcp-server
+helm repo update
+helm install nextcloud-mcp nextcloud-mcp/nextcloud-mcp-server \
+  --set nextcloud.host=https://cloud.example.com \
+  --set auth.basic.username=myuser \
+  --set auth.basic.password=mypassword
 ```

-See [Installation Guide](docs/installation.md) for detailed instructions.
+See [Installation Guide](docs/installation.md) for detailed instructions, or [Helm Chart README](charts/nextcloud-mcp-server/README.md) for Kubernetes deployment.

 ### 2. Configure

@@ -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: 0.1.0
-appVersion: "0.21.0"
+version: 0.22.7
+appVersion: "0.22.7"
 keywords:
  - nextcloud
  - mcp
@@ -114,14 +114,35 @@ ingress:
 | `auth.oauth.persistence.enabled` | Enable persistent storage for OAuth | `true` |
 | `auth.oauth.persistence.size` | Size of OAuth storage PVC | `100Mi` |

+#### MCP Server Configuration
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `mcp.transport` | Transport mode | `streamable-http` |
+| `mcp.port` | Server port (used by both auth modes) | `8000` |
+| `mcp.extraArgs` | Additional command-line arguments | `[]` |
+
+The `extraArgs` parameter allows you to pass additional command-line arguments to the MCP server. This is useful for enabling debug logging, enabling specific apps, or other runtime configuration.
+
+**Example:**
+```yaml
+mcp:
+  extraArgs:
+    - "--log-level"
+    - "debug"
+    - "--enable-app"
+    - "notes"
+```
+
 #### Image Configuration

 | Parameter | Description | Default |
 |-----------|-------------|---------|
 | `image.repository` | Container image repository | `ghcr.io/cbcoutinho/nextcloud-mcp-server` |
-| `image.tag` | Container image tag | `""` (uses chart appVersion) |
 | `image.pullPolicy` | Image pull policy | `IfNotPresent` |

+**Note:** Image tag is automatically set to the chart's `appVersion` and cannot be overridden.
+
 #### Resources

 | Parameter | Description | Default |
@@ -137,7 +158,6 @@ ingress:
 |-----------|-------------|---------|
 | `service.type` | Service type | `ClusterIP` |
 | `service.port` | Service port | `8000` |
-| `service.oauthPort` | OAuth service port | `8001` |

 #### Ingress

@@ -371,7 +391,6 @@ helm upgrade nextcloud-mcp ./helm/nextcloud-mcp-server -f custom-values.yaml

 ```bash
 helm upgrade nextcloud-mcp ./helm/nextcloud-mcp-server \
-  --set image.tag=0.21.0 \
  --set resources.limits.memory=1Gi
 ```

@@ -95,21 +95,17 @@ Create the name of the PVC to use for OAuth storage
 {{- end }}

 {{/*
-Return the appropriate MCP server port based on auth mode
+Return the MCP server port
 */}}
 {{- define "nextcloud-mcp-server.port" -}}
-{{- if eq .Values.auth.mode "oauth" }}
-{{- .Values.auth.oauth.port }}
-{{- else }}
 {{- .Values.mcp.port }}
 {{- end }}
-{{- end }}

 {{/*
-Return the image tag
+Return the image tag (always uses chart appVersion)
 */}}
 {{- define "nextcloud-mcp-server.imageTag" -}}
-{{- .Values.image.tag | default .Chart.AppVersion }}
+{{- .Chart.AppVersion }}
 {{- end }}

 {{/*
@@ -46,11 +46,12 @@ spec:
            - "{{ .Values.mcp.transport }}"
            {{- if eq .Values.auth.mode "oauth" }}
            - "--oauth"
-            - "--port"
-            - "{{ .Values.auth.oauth.port }}"
            - "--oauth-token-type"
            - "{{ .Values.auth.oauth.tokenType }}"
            {{- end }}
+            {{- with .Values.mcp.extraArgs }}
+            {{- toYaml . | nindent 12 }}
+            {{- end }}
          ports:
            - name: http
              containerPort: {{ include "nextcloud-mcp-server.port" . }}
@@ -8,8 +8,7 @@ replicaCount: 1
 image:
  repository: ghcr.io/cbcoutinho/nextcloud-mcp-server
  pullPolicy: IfNotPresent
-  # Overrides the image tag whose default is the chart appVersion.
-  tag: ""
+  # Image tag is automatically set to chart appVersion

 imagePullSecrets: []
 nameOverride: ""
@@ -61,8 +60,6 @@ auth:

  # OAuth2/OIDC settings (experimental)
  oauth:
-    # Port for OAuth MCP server (default: 8001)
-    port: 8001
    # OAuth token type: "jwt" or "opaque"
    tokenType: "jwt"
    # Pre-registered OAuth client ID (optional, ignored if existingSecret is set)
@@ -97,8 +94,11 @@ auth:
 mcp:
  # Transport mode (default: streamable-http for SSE)
  transport: "streamable-http"
-  # Port for basic auth mode
+  # Port for MCP server (both basic auth and OAuth modes)
  port: 8000
+  # Additional command-line arguments to pass to nextcloud-mcp-server
+  # Example: ["--log-level", "debug", "--enable-app", "notes"]
+  extraArgs: []

 # Document processing configuration (optional)
 documentProcessing:
@@ -171,8 +171,6 @@ securityContext:
 service:
  type: ClusterIP
  port: 8000
-  # For OAuth mode, you may want to expose both ports
-  oauthPort: 8001
  annotations: {}

 ingress:
@@ -21,7 +21,7 @@ services:
    restart: always

  app:
-    image: docker.io/library/nextcloud:32.0.1@sha256:42a36b4711191273a9cf8cebfd35602909eb1bee461b7076d4d5a57f7ec2b81e
+    image: docker.io/library/nextcloud:32.0.1@sha256:1e4eae55eebe094cae6f9e7b6e0b4bccf4a4fe7b7e6f6f8f57010994b3b2ee42
    restart: always
    ports:
      - 0.0.0.0:8080:80
@@ -45,7 +45,7 @@ services:
      - REDIS_HOST=redis

  recipes:
-    image: docker.io/library/nginx:alpine@sha256:61e01287e546aac28a3f56839c136b31f590273f3b41187a36f46f6a03bbfe22
+    image: docker.io/library/nginx:alpine@sha256:b3c656d55d7ad751196f21b7fd2e8d4da9cb430e32f646adcf92441b72f82b14
    restart: always
    volumes:
      - ./tests/fixtures/test_recipe.html:/usr/share/nginx/html/test_recipe.html:ro
@@ -0,0 +1,795 @@
+# ADR-002: Vector Database Background Sync Authentication
+
+## Status
+Proposed
+
+## Context
+
+To enable semantic search capabilities, the MCP server needs to index user content (notes, files, calendar events) into a vector database. This requires a background sync worker that:
+
+1. **Runs independently** of user requests (periodic or continuous operation)
+2. **Accesses multiple users' content** to build a comprehensive search index
+3. **Respects user permissions** - only index content users have access to
+4. **Operates in OAuth mode** - where the MCP server doesn't have traditional admin credentials
+
+### Current OAuth Architecture
+
+The MCP server currently operates in two authentication modes:
+
+1. **BasicAuth Mode**: Uses username/password credentials (typically admin account)
+2. **OAuth Mode**: Single OAuth client, multiple user tokens
+   - Users authenticate via OAuth flow
+   - Each request includes user's access token
+   - Server creates per-request `NextcloudClient` with user's bearer token
+   - No tokens are stored server-side
+
+### The Challenge
+
+Background workers need long-lived authentication to:
+- Index content continuously/periodically
+- Process multiple users' data in batch operations
+- Operate when users are not actively making requests
+
+However, in OAuth mode:
+- User access tokens are ephemeral (exist only during request)
+- MCP server doesn't store user credentials
+- Admin credentials defeat the purpose of OAuth
+
+We need an OAuth-native solution that maintains security while enabling background operations.
+
+## Decision
+
+We will implement a **tiered authentication strategy** that leverages OAuth standards with graceful fallback:
+
+### Primary Strategy: OAuth-Based Authentication
+
+**Tier 1: Offline Access with Refresh Tokens** (Preferred)
+- Request `offline_access` scope during OAuth client registration
+- Receive and securely store user refresh tokens
+- Background worker exchanges refresh tokens for access tokens as needed
+- Respects per-user permissions and provides full audit trail
+
+**Tier 2: Token Exchange (RFC 8693)** (If supported)
+- Service account exchanges its token for user-scoped tokens on-demand
+- No token storage required
+- Only available if OIDC provider implements RFC 8693
+
+### Fallback Strategy: Admin Credentials
+
+**Tier 3: Admin BasicAuth** (Development/Simple Deployments)
+- Dedicated sync account with read-only permissions
+- Clear documentation of security implications
+- Recommended only for trusted environments
+
+### Key Architectural Principles
+
+1. **Capability Detection**: Automatically detect which OAuth methods are supported
+2. **Dual-Phase Authorization**:
+   - Sync worker indexes with service credentials
+   - User requests verify access with user's OAuth token
+3. **Defense in Depth**: Vector database is search accelerator, not security boundary
+4. **Separation of Concerns**: Sync credentials ≠ Request credentials
+
+## Implementation Details
+
+### 1. Offline Access Flow (Tier 1)
+
+#### 1.1 Client Registration
+```python
+# During OAuth client registration
+client_metadata = {
+    "client_name": "Nextcloud MCP Server",
+    "redirect_uris": ["http://localhost:8000/oauth/callback"],
+    "grant_types": ["authorization_code", "refresh_token"],
+    "scope": "openid profile email offline_access notes:read files:read ...",
+    "token_type": "Bearer"  # or "jwt"
+}
+```
+
+#### 1.2 Token Storage
+```python
+# Encrypted token storage
+class RefreshTokenStorage:
+    """Securely store and manage user refresh tokens"""
+
+    def __init__(self, db_path: str, encryption_key: bytes):
+        self.db = Database(db_path)
+        self.cipher = Fernet(encryption_key)
+
+    async def store_refresh_token(
+        self,
+        user_id: str,
+        refresh_token: str,
+        expires_at: int | None = None
+    ):
+        """Store encrypted refresh token for user"""
+        encrypted_token = self.cipher.encrypt(refresh_token.encode())
+        await self.db.execute(
+            "INSERT OR REPLACE INTO refresh_tokens VALUES (?, ?, ?, ?)",
+            (user_id, encrypted_token, expires_at, int(time.time()))
+        )
+
+    async def get_refresh_token(self, user_id: str) -> str | None:
+        """Retrieve and decrypt refresh token"""
+        row = await self.db.fetch_one(
+            "SELECT encrypted_token FROM refresh_tokens WHERE user_id = ?",
+            (user_id,)
+        )
+        if row:
+            return self.cipher.decrypt(row[0]).decode()
+        return None
+```
+
+#### 1.3 Token Refresh Flow
+```python
+async def get_user_access_token(user_id: str) -> str:
+    """Exchange refresh token for fresh access token"""
+
+    # Retrieve stored refresh token
+    refresh_token = await token_storage.get_refresh_token(user_id)
+    if not refresh_token:
+        raise ValueError(f"No refresh token for user {user_id}")
+
+    # Exchange for access token
+    async with httpx.AsyncClient() as client:
+        response = await client.post(
+            token_endpoint,
+            data={
+                "grant_type": "refresh_token",
+                "refresh_token": refresh_token
+            },
+            auth=(client_id, client_secret)
+        )
+        response.raise_for_status()
+        token_data = response.json()
+
+        # Store new refresh token if rotated
+        if "refresh_token" in token_data:
+            await token_storage.store_refresh_token(
+                user_id,
+                token_data["refresh_token"],
+                token_data.get("refresh_expires_in")
+            )
+
+        return token_data["access_token"]
+```
+
+#### 1.4 Capturing Refresh Tokens
+
+**Challenge**: MCP protocol doesn't expose refresh tokens to server
+
+**Solution**: Intercept OAuth callback
+```python
+# Add route to MCP server
+@app.route("/oauth/callback")
+async def oauth_callback(request):
+    """Capture OAuth callback and store refresh token"""
+
+    code = request.query_params.get("code")
+    state = request.query_params.get("state")
+
+    # Exchange authorization code for tokens
+    token_response = await exchange_authorization_code(code)
+
+    # Extract user info
+    userinfo = await get_userinfo(token_response["access_token"])
+    user_id = userinfo["sub"]
+
+    # Store refresh token (if present)
+    if "refresh_token" in token_response:
+        await token_storage.store_refresh_token(
+            user_id,
+            token_response["refresh_token"],
+            expires_at=token_response.get("refresh_expires_in")
+        )
+        logger.info(f"Stored refresh token for user: {user_id}")
+
+    # Continue MCP OAuth flow
+    return redirect_to_mcp_client(state, token_response)
+```
+
+### 2. Token Exchange Flow (Tier 2)
+
+#### 2.1 Capability Detection
+```python
+async def check_token_exchange_support(discovery_url: str) -> bool:
+    """Check if OIDC provider supports RFC 8693 token exchange"""
+
+    async with httpx.AsyncClient() as client:
+        response = await client.get(discovery_url)
+        discovery = response.json()
+
+        # Check for token exchange grant type
+        grant_types = discovery.get("grant_types_supported", [])
+        return "urn:ietf:params:oauth:grant-type:token-exchange" in grant_types
+```
+
+#### 2.2 Token Exchange Implementation
+```python
+async def exchange_for_user_token(
+    service_token: str,
+    user_id: str,
+    scopes: list[str]
+) -> str:
+    """Exchange service token for user-scoped token"""
+
+    async with httpx.AsyncClient() as client:
+        response = await client.post(
+            token_endpoint,
+            data={
+                "grant_type": "urn:ietf:params:oauth:grant-type:token-exchange",
+                "subject_token": service_token,
+                "subject_token_type": "urn:ietf:params:oauth:token-type:access_token",
+                "requested_token_type": "urn:ietf:params:oauth:token-type:access_token",
+                "resource": f"user:{user_id}",
+                "scope": " ".join(scopes)
+            },
+            auth=(client_id, client_secret)
+        )
+
+        if response.status_code != 200:
+            logger.warning(f"Token exchange failed: {response.status_code}")
+            raise TokenExchangeNotSupportedError()
+
+        return response.json()["access_token"]
+```
+
+#### 2.3 Service Account Token
+```python
+async def get_service_token() -> str:
+    """Get token for MCP server's service account"""
+
+    async with httpx.AsyncClient() as client:
+        response = await client.post(
+            token_endpoint,
+            data={
+                "grant_type": "client_credentials",
+                "scope": "notes:read files:read calendar:read"
+            },
+            auth=(client_id, client_secret)
+        )
+        response.raise_for_status()
+        return response.json()["access_token"]
+```
+
+### 3. Sync Worker with Tiered Authentication
+
+```python
+# nextcloud_mcp_server/sync_worker.py
+class VectorSyncWorker:
+    """Background worker for indexing content into vector database"""
+
+    def __init__(self):
+        self.auth_method = None
+        self.token_storage = None
+        self.vector_service = None
+
+    async def initialize(self):
+        """Detect and configure authentication method"""
+
+        # Try Tier 1: Offline Access
+        if os.getenv("ENABLE_OFFLINE_ACCESS") == "true":
+            try:
+                encryption_key = os.getenv("TOKEN_ENCRYPTION_KEY")
+                self.token_storage = RefreshTokenStorage(
+                    db_path="tokens.db",
+                    encryption_key=base64.b64decode(encryption_key)
+                )
+                self.auth_method = "offline_access"
+                logger.info("✓ Using offline_access authentication")
+                return
+            except Exception as e:
+                logger.warning(f"Offline access unavailable: {e}")
+
+        # Try Tier 2: Token Exchange
+        try:
+            if await check_token_exchange_support(discovery_url):
+                self.auth_method = "token_exchange"
+                logger.info("✓ Using token exchange authentication (RFC 8693)")
+                return
+        except Exception as e:
+            logger.warning(f"Token exchange unavailable: {e}")
+
+        # Fallback: Admin Credentials
+        if os.getenv("NEXTCLOUD_USERNAME") and os.getenv("NEXTCLOUD_PASSWORD"):
+            self.auth_method = "admin_basic"
+            logger.warning(
+                "⚠ Using admin BasicAuth authentication. "
+                "Consider enabling offline_access for production."
+            )
+            return
+
+        raise RuntimeError("No authentication method available for sync worker")
+
+    async def get_user_client(self, user_id: str) -> NextcloudClient:
+        """Get authenticated client for user based on auth method"""
+
+        if self.auth_method == "offline_access":
+            # Exchange refresh token for access token
+            access_token = await get_user_access_token(user_id)
+            return NextcloudClient.from_token(
+                base_url=nextcloud_host,
+                token=access_token,
+                username=user_id
+            )
+
+        elif self.auth_method == "token_exchange":
+            # Get service token and exchange for user token
+            service_token = await get_service_token()
+            user_token = await exchange_for_user_token(
+                service_token,
+                user_id,
+                scopes=["notes:read", "files:read"]
+            )
+            return NextcloudClient.from_token(
+                base_url=nextcloud_host,
+                token=user_token,
+                username=user_id
+            )
+
+        elif self.auth_method == "admin_basic":
+            # Use admin credentials (fallback)
+            return NextcloudClient.from_env()
+
+        raise RuntimeError(f"Unknown auth method: {self.auth_method}")
+
+    async def sync_user_content(self, user_id: str):
+        """Index a user's content into vector database"""
+
+        try:
+            # Get authenticated client for this user
+            client = await self.get_user_client(user_id)
+
+            # Sync notes
+            notes = await client.notes.list_notes()
+            for note in notes:
+                embedding = await self.vector_service.embed(note.content)
+                await self.vector_service.upsert(
+                    collection="nextcloud_content",
+                    id=f"note_{note.id}",
+                    vector=embedding,
+                    metadata={
+                        "user_id": user_id,
+                        "content_type": "note",
+                        "note_id": note.id,
+                        "title": note.title,
+                        "category": note.category
+                    }
+                )
+
+            logger.info(f"Synced {len(notes)} notes for user: {user_id}")
+
+        except Exception as e:
+            logger.error(f"Failed to sync user {user_id}: {e}")
+
+    async def run(self):
+        """Main sync loop"""
+
+        await self.initialize()
+
+        while True:
+            try:
+                # Get list of users to sync
+                if self.auth_method == "admin_basic":
+                    # Admin can list all users
+                    admin_client = NextcloudClient.from_env()
+                    users = await admin_client.users.list_users()
+                    user_ids = [u.id for u in users]
+                else:
+                    # OAuth methods: only sync users with stored tokens
+                    user_ids = await self.token_storage.get_all_user_ids()
+
+                logger.info(f"Syncing content for {len(user_ids)} users")
+
+                for user_id in user_ids:
+                    await self.sync_user_content(user_id)
+
+                logger.info("Sync complete, sleeping...")
+                await asyncio.sleep(300)  # 5 minutes
+
+            except Exception as e:
+                logger.error(f"Sync failed: {e}")
+                await asyncio.sleep(60)  # Retry after 1 minute
+```
+
+### 4. User Request Verification (Dual-Phase Authorization)
+
+```python
+@mcp.tool()
+@require_scopes("notes:read")
+async def nc_notes_semantic_search(
+    query: str,
+    ctx: Context,
+    limit: int = 10
+) -> SemanticSearchResponse:
+    """Semantic search with permission verification"""
+
+    # Get user's OAuth client (uses their access token from request)
+    user_client = get_client(ctx)
+    username = user_client.username
+
+    # Phase 1: Vector search (fast, may include false positives)
+    embedding = await vector_service.embed(query)
+    candidate_results = await qdrant.search(
+        collection_name="nextcloud_content",
+        query_vector=embedding,
+        query_filter={
+            "must": [
+                {
+                    "should": [
+                        {"key": "user_id", "match": {"value": username}},
+                        {"key": "shared_with", "match": {"any": [username]}}
+                    ]
+                },
+                {"key": "content_type", "match": {"value": "note"}}
+            ]
+        },
+        limit=limit * 2  # Get extra candidates
+    )
+
+    # Phase 2: Verify access via Nextcloud API (authoritative)
+    verified_results = []
+    for candidate in candidate_results:
+        note_id = candidate.payload["note_id"]
+        try:
+            # This uses user's OAuth token - will fail if no access
+            note = await user_client.notes.get_note(note_id)
+            verified_results.append({
+                "note": note,
+                "score": candidate.score
+            })
+            if len(verified_results) >= limit:
+                break
+        except HTTPStatusError as e:
+            if e.response.status_code == 403:
+                # User doesn't have access - skip silently
+                logger.debug(f"Filtered out note {note_id} for {username}")
+                continue
+            raise
+
+    return SemanticSearchResponse(results=verified_results)
+```
+
+### 5. Security Implementation
+
+#### 5.1 Token Encryption
+```python
+# Generate encryption key (store securely)
+from cryptography.fernet import Fernet
+
+# On first setup
+encryption_key = Fernet.generate_key()
+# Store in environment or secrets manager
+# NEVER commit to source control
+
+# In production
+encryption_key = os.getenv("TOKEN_ENCRYPTION_KEY")  # Base64-encoded Fernet key
+```
+
+#### 5.2 Token Rotation
+```python
+async def rotate_refresh_token(user_id: str):
+    """Handle refresh token rotation"""
+
+    old_refresh_token = await token_storage.get_refresh_token(user_id)
+
+    # Exchange for new tokens
+    response = await exchange_refresh_token(old_refresh_token)
+
+    if "refresh_token" in response:
+        # Store new refresh token
+        await token_storage.store_refresh_token(
+            user_id,
+            response["refresh_token"],
+            expires_at=response.get("refresh_expires_in")
+        )
+
+        # Securely delete old token
+        await token_storage.delete_refresh_token(user_id, old_refresh_token)
+```
+
+#### 5.3 Audit Logging
+```python
+async def audit_log(
+    event: str,
+    user_id: str,
+    resource_type: str,
+    resource_id: str,
+    auth_method: str
+):
+    """Log sync operations for audit trail"""
+
+    await audit_db.execute(
+        "INSERT INTO audit_logs VALUES (?, ?, ?, ?, ?, ?, ?)",
+        (
+            int(time.time()),
+            event,  # "index_note", "index_file"
+            user_id,
+            resource_type,
+            resource_id,
+            auth_method,
+            socket.gethostname()
+        )
+    )
+```
+
+### 6. Configuration
+
+#### 6.1 Environment Variables
+```bash
+# Tier 1: Offline Access
+ENABLE_OFFLINE_ACCESS=true
+TOKEN_ENCRYPTION_KEY=<base64-encoded-fernet-key>
+TOKEN_STORAGE_DB=/app/data/tokens.db
+
+# Tier 2: Token Exchange (auto-detected)
+# No configuration needed - detected via OIDC discovery
+
+# Tier 3: Admin Fallback
+NEXTCLOUD_USERNAME=sync-bot
+NEXTCLOUD_PASSWORD=<secure-password>
+
+# Vector Database
+QDRANT_URL=http://qdrant:6333
+QDRANT_API_KEY=<api-key>
+
+# Sync Configuration
+SYNC_INTERVAL_SECONDS=300
+SYNC_BATCH_SIZE=100
+```
+
+#### 6.2 Docker Compose
+```yaml
+services:
+  mcp-sync:
+    build: .
+    command: ["python", "-m", "nextcloud_mcp_server.sync_worker"]
+    environment:
+      - NEXTCLOUD_HOST=http://app:80
+      - ENABLE_OFFLINE_ACCESS=true
+      - TOKEN_ENCRYPTION_KEY=${TOKEN_ENCRYPTION_KEY}
+      - QDRANT_URL=http://qdrant:6333
+      # OAuth client credentials (for token refresh)
+      - NEXTCLOUD_OIDC_CLIENT_ID=${NEXTCLOUD_OIDC_CLIENT_ID}
+      - NEXTCLOUD_OIDC_CLIENT_SECRET=${NEXTCLOUD_OIDC_CLIENT_SECRET}
+    volumes:
+      - sync-tokens:/app/data
+    depends_on:
+      - app
+      - qdrant
+
+volumes:
+  sync-tokens:  # Persistent storage for encrypted tokens
+```
+
+## Consequences
+
+### Benefits
+
+1. **OAuth-Native Authentication**
+   - Leverages standard OAuth flows (offline_access, token exchange)
+   - No reliance on admin passwords in production
+   - Compatible with enterprise OIDC providers
+
+2. **User-Level Permissions**
+   - Each user's content indexed with their own credentials
+   - Respects sharing, permissions, and access controls
+   - Full audit trail of which user's token was used
+
+3. **Security**
+   - Tokens encrypted at rest
+   - Short-lived access tokens (refreshed as needed)
+   - Token rotation support
+   - Defense in depth with dual-phase authorization
+
+4. **Flexibility**
+   - Automatic capability detection
+   - Graceful degradation through authentication tiers
+   - Works with varying OIDC provider capabilities
+
+5. **Operational**
+   - Background sync independent of user activity
+   - Efficient batch processing
+   - Clear separation of sync vs request credentials
+
+### Limitations
+
+1. **Complexity**
+   - Multiple authentication paths to maintain
+   - Token storage and encryption infrastructure
+   - More moving parts than simple admin auth
+
+2. **User Experience**
+   - `offline_access` scope may require additional consent
+   - Users must authenticate at least once for indexing
+   - New users not automatically indexed
+
+3. **OIDC Provider Dependency**
+   - Token exchange requires RFC 8693 support (rare)
+   - Refresh token rotation varies by provider
+   - Some providers may not support offline_access
+
+4. **Operational Overhead**
+   - Token database maintenance
+   - Monitoring token expiration
+   - Handling revoked tokens gracefully
+
+### Security Considerations
+
+#### Threat Model
+
+**Threat 1: Token Storage Breach**
+- **Mitigation**: Encryption at rest using Fernet
+- **Mitigation**: Secure key management (secrets manager)
+- **Mitigation**: Minimal token lifetime
+- **Detection**: Audit logs for unusual access patterns
+
+**Threat 2: Token Replay**
+- **Mitigation**: Short-lived access tokens (refreshed frequently)
+- **Mitigation**: Token rotation on each refresh
+- **Mitigation**: Revocation support
+
+**Threat 3: Privilege Escalation**
+- **Mitigation**: Dual-phase authorization (vector DB + Nextcloud API)
+- **Mitigation**: Sync worker uses same scopes as user requests
+- **Mitigation**: Per-user token isolation
+
+**Threat 4: Vector Database Poisoning**
+- **Mitigation**: User requests always verify via Nextcloud API
+- **Mitigation**: Vector DB is cache/accelerator, not source of truth
+- **Mitigation**: Sync operations audited per user
+
+#### Security Best Practices
+
+1. **Token Encryption Key Management**
+   ```bash
+   # Generate secure key
+   python -c "from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())"
+
+   # Store in secrets manager (Vault, AWS Secrets Manager, etc.)
+   # Or use environment variable with restricted permissions
+   ```
+
+2. **Token Storage Permissions**
+   ```bash
+   # Restrict database file permissions
+   chmod 600 /app/data/tokens.db
+   chown mcp-server:mcp-server /app/data/tokens.db
+   ```
+
+3. **Token Rotation Schedule**
+   - Refresh access tokens every 5 minutes (or token expiry)
+   - Rotate refresh tokens on each use (if provider supports)
+   - Revoke tokens on user logout/deauthorization
+
+4. **Monitoring and Alerting**
+   - Alert on token refresh failures
+   - Monitor for unusual access patterns
+   - Track token age and rotation
+   - Audit sync operations per user
+
+### Future Enhancements
+
+1. **Token Revocation Handling**
+   - Webhook endpoint for token revocation events
+   - Periodic validation of stored tokens
+   - Graceful handling of revoked tokens
+
+2. **Selective Sync**
+   - Allow users to opt-in/opt-out of indexing
+   - Per-content-type sync preferences
+   - Privacy controls for sensitive content
+
+3. **Multi-Tenant Token Storage**
+   - Separate token databases per tenant
+   - Key rotation per tenant
+   - Tenant isolation
+
+4. **Token Lifecycle Management**
+   - Automatic cleanup of expired tokens
+   - Token usage analytics
+   - Token health dashboard
+
+5. **Alternative OAuth Flows**
+   - Device flow for headless sync
+   - Resource owner password credentials (ROPC) as fallback
+   - SAML assertion grants
+
+## Alternatives Considered
+
+### Alternative 1: Admin BasicAuth Only
+
+**Approach**: Background worker always uses admin credentials
+
+**Pros**:
+- Simple implementation
+- No token storage complexity
+- Works with any authentication backend
+
+**Cons**:
+- Violates principle of least privilege
+- Single powerful credential
+- No per-user audit trail
+- Bypasses OAuth entirely
+
+**Decision**: Rejected for production use; kept as fallback only
+
+### Alternative 2: Client Credentials Grant Only
+
+**Approach**: Service account with broad read permissions
+
+**Pros**:
+- OAuth-native pattern
+- No user token storage
+- Standard OAuth flow
+
+**Cons**:
+- Requires client_credentials support (may not be available)
+- Still needs broad cross-user permissions
+- Not well-suited for multi-user indexing
+
+**Decision**: Rejected; token exchange is better fit for multi-user scenario
+
+### Alternative 3: Per-User Access Token Storage
+
+**Approach**: Store user access tokens (not refresh tokens)
+
+**Pros**:
+- Simpler than refresh token flow
+- No token refresh logic needed
+
+**Cons**:
+- Access tokens are short-lived (1-24 hours)
+- Requires frequent re-authentication
+- Poor user experience
+- Sync gaps when tokens expire
+
+**Decision**: Rejected; refresh tokens provide better UX
+
+### Alternative 4: On-Demand Indexing Only
+
+**Approach**: Index content when user searches (no background worker)
+
+**Pros**:
+- Uses user's request token
+- No background auth needed
+- Simpler architecture
+
+**Cons**:
+- Very slow first search
+- Poor user experience
+- Incomplete index
+- Can't pre-compute embeddings
+
+**Decision**: Rejected; background indexing is essential for semantic search
+
+### Alternative 5: Nextcloud App Tokens
+
+**Approach**: Generate app-specific passwords for each user
+
+**Pros**:
+- Nextcloud-native feature
+- User-controlled revocation
+- Scoped per-application
+
+**Cons**:
+- Requires user interaction to create
+- May not support programmatic creation
+- Still requires secure storage
+- Not standard OAuth
+
+**Decision**: Rejected; not automatable for background worker
+
+## Related Decisions
+
+- ADR-001: Enhanced Note Search (establishes need for vector search)
+- [Future] ADR-003: Vector Database Selection
+- [Future] ADR-004: Embedding Model Strategy
+
+## References
+
+- [RFC 8693: OAuth 2.0 Token Exchange](https://datatracker.ietf.org/doc/html/rfc8693)
+- [RFC 6749: OAuth 2.0 - Refresh Tokens](https://datatracker.ietf.org/doc/html/rfc6749#section-1.5)
+- [OpenID Connect Core - Offline Access](https://openid.net/specs/openid-connect-core-1_0.html#OfflineAccess)
+- [OWASP: OAuth Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/OAuth2_Cheat_Sheet.html)
+- [RFC 8707: Resource Indicators for OAuth 2.0](https://datatracker.ietf.org/doc/html/rfc8707)
@@ -0,0 +1,441 @@
+# Ollama Capacity Analysis: ollama.internal.coutinho.io
+
+**Date**: 2025-10-30
+**Model**: nomic-embed-text:latest
+**Test Location**: From nextcloud-mcp-server host
+
+## Summary
+
+✅ **Ollama instance is operational and performing well**
+- Embedding generation working correctly
+- Reasonable latency for small-medium workloads
+- Good parallelism support
+- Suitable for development and small production deployments
+
+## Test Results
+
+### Model Configuration
+
+```json
+{
+  "model": "nomic-embed-text",
+  "dimensions": 768,
+  "status": "operational"
+}
+```
+
+### Performance Metrics
+
+#### 1. Single Embedding Latency
+
+**Result**: ~553ms per embedding
+- **Total time**: 0.553 seconds
+- **Includes**: Network + processing + model inference
+- **Quality**: Full 768-dimensional vector
+
+**Analysis**:
+- Higher than bare-metal benchmarks (~100ms) due to network latency
+- Acceptable for interactive search queries
+- Within expected range for remote Ollama instance
+
+#### 2. Batch Processing (5 items)
+
+**Result**: ~1.02 seconds for 5 embeddings
+- **Per-item average**: 204ms
+- **Throughput**: ~4.9 embeddings/sec
+- **Batch efficiency**: 2.7x faster than sequential
+
+**Analysis**:
+- Good batching efficiency (2.7x speedup vs 5x theoretical)
+- Optimal for background indexing
+- Network overhead amortized across batch
+
+#### 3. Batch Processing (20 items)
+
+**Result**: ~6.71 seconds for 20 embeddings
+- **Per-item average**: 336ms
+- **Throughput**: ~3.0 embeddings/sec
+- **Batch efficiency**: 1.65x faster than sequential
+
+**Analysis**:
+- Performance degrades slightly with larger batches
+- Still faster than sequential processing
+- Matches reported Ollama behavior (quality issues at batch >16)
+- **Recommendation**: Keep batch size ≤16 for best quality
+
+#### 4. Concurrent Requests (5 parallel)
+
+**Result**: ~1.27 seconds for 5 parallel requests
+- **Effective parallelism**: ~4x speedup (vs 2.77s sequential)
+- **Per-request average**: 254ms
+- **Throughput**: ~3.9 requests/sec
+
+**Analysis**:
+- Excellent parallelism support
+- Server handles concurrent requests efficiently
+- Network and compute overlap effectively
+- Good for multi-user scenarios
+
+## Capacity Planning
+
+### Current Performance Profile
+
+| Metric | Value | Rating |
+|--------|-------|--------|
+| Single embedding latency | 553ms | ⚠️ Moderate |
+| Batch (5) throughput | 4.9/sec | ✅ Good |
+| Batch (20) throughput | 3.0/sec | ⚠️ Moderate |
+| Concurrent throughput | 3.9/sec | ✅ Good |
+| Network latency | ~300-400ms | ⚠️ Significant |
+
+### Bottleneck Analysis
+
+**Primary Bottleneck**: Network latency (~300-400ms per request)
+- Model inference: ~100-200ms (estimated)
+- Network round-trip: ~300-400ms (measured overhead)
+- **Impact**: 60-70% of total latency is network
+
+**Secondary Bottleneck**: CPU/GPU capacity (unknown hardware)
+- Batch performance degrades at >16 items
+- Suggests resource constraints
+- Likely CPU-only (no GPU metrics available)
+
+### Recommended Usage Patterns
+
+#### ✅ **Excellent For:**
+
+**1. Background Indexing**
+- Use batch size of 10-15 items
+- Expected throughput: 3-5 embeddings/sec
+- **10,000 notes**: ~30-55 minutes to index
+- **1,000 notes**: ~3-5 minutes to index
+
+**2. Interactive Search**
+- Single query embedding: ~550ms
+- Acceptable for user-facing search
+- Add 100-200ms for vector search + verification
+- **Total search time**: ~650-750ms (reasonable UX)
+
+**3. Multi-User Development**
+- 5-10 concurrent users: Comfortable
+- Good parallelism support
+- Network latency dominates (shared)
+
+#### ⚠️ **Consider Alternatives For:**
+
+**1. Real-Time Applications**
+- Sub-100ms latency requirements
+- High-frequency queries (>10/sec sustained)
+- Consider: Local embeddings or Infinity
+
+**2. Large-Scale Batch Processing**
+- >100,000 documents to index
+- >10 embeddings/sec sustained
+- Consider: GPU-accelerated TEI
+
+**3. Production with >50 Users**
+- High concurrent load
+- Latency sensitivity
+- Consider: Dedicated embedding service
+
+### Deployment Scenarios
+
+#### Scenario 1: Development Environment
+
+**Profile**:
+- 1-3 developers
+- 1,000-5,000 notes total
+- Occasional searches/indexing
+
+**Verdict**: ✅ **Perfect fit**
+- Initial index: ~5-15 minutes (one-time)
+- Incremental updates: <1 minute
+- Search latency: Acceptable
+- No infrastructure changes needed
+
+**Configuration**:
+```bash
+OLLAMA_URL=https://ollama.internal.coutinho.io
+OLLAMA_MODEL=nomic-embed-text
+VECTOR_SYNC_INTERVAL=600  # 10 minutes
+VECTOR_SYNC_BATCH_SIZE=10
+```
+
+#### Scenario 2: Small Production (10-20 users)
+
+**Profile**:
+- 10-20 active users
+- 10,000-50,000 notes total
+- 50-200 searches/day
+- Nightly incremental indexing
+
+**Verdict**: ✅ **Suitable with optimizations**
+- Initial index: 1-3 hours (run overnight)
+- Incremental: 5-15 minutes/night
+- Search: Acceptable for most users
+- Monitor network latency
+
+**Configuration**:
+```bash
+OLLAMA_URL=https://ollama.internal.coutinho.io
+OLLAMA_MODEL=nomic-embed-text
+VECTOR_SYNC_INTERVAL=86400  # Daily at night
+VECTOR_SYNC_BATCH_SIZE=12  # Conservative for quality
+SEARCH_TIMEOUT_MS=1000  # Account for 550ms latency
+```
+
+**Optimizations**:
+- Run sync during off-hours
+- Cache query embeddings (common searches)
+- Use hybrid search (keyword + semantic)
+
+#### Scenario 3: Medium Production (50-100 users)
+
+**Profile**:
+- 50-100 active users
+- 100,000+ notes
+- 500-1000 searches/day
+- Real-time indexing desired
+
+**Verdict**: ⚠️ **Marginal - monitor closely**
+- Initial index: 5-10 hours
+- Search latency: May feel slow for some users
+- Concurrent load: Approaching limits
+- **Recommendation**: Plan migration to Infinity
+
+**Configuration**:
+```bash
+OLLAMA_URL=https://ollama.internal.coutinho.io
+OLLAMA_MODEL=nomic-embed-text
+VECTOR_SYNC_INTERVAL=3600  # Hourly
+VECTOR_SYNC_BATCH_SIZE=10
+SEMANTIC_WEIGHT=0.5  # Rely more on keyword search
+SEARCH_TIMEOUT_MS=2000  # Generous timeout
+```
+
+**Migration Path**:
+- Start with Ollama
+- Monitor latency metrics
+- When p95 latency >1s, migrate to Infinity
+- Keep Ollama as fallback
+
+#### Scenario 4: Large Production (>100 users)
+
+**Profile**:
+- >100 active users
+- >500,000 notes
+- >1000 searches/day
+- Real-time expectations
+
+**Verdict**: ❌ **Not recommended**
+- Latency too high for scale
+- Throughput insufficient
+- Network becomes bottleneck
+- **Recommendation**: Use Infinity or TEI from start
+
+## Network Latency Optimization
+
+### Current Overhead: ~300-400ms
+
+**If MCP server runs closer to Ollama**:
+```
+Same VPC/network: ~1-5ms (300-400ms savings!)
+Same host: <1ms (300-400ms savings!)
+```
+
+### Recommendation
+
+**Option A: Co-locate MCP server with Ollama**
+- Reduces latency from 550ms → 150-200ms
+- 2.5-3x improvement
+- Makes Ollama competitive with cloud APIs
+
+**Option B: Keep separate (current)**
+- Simpler deployment
+- Better security isolation
+- Accept 550ms latency
+
+**Option C: Add Infinity container to MCP server**
+- Best of both worlds
+- Use Infinity for speed (local)
+- Fallback to Ollama if needed
+
+## Capacity Estimates
+
+### Indexing Capacity
+
+**Sustained Throughput**: 3-4 embeddings/sec (conservative)
+
+| Document Count | Index Time | Notes |
+|----------------|------------|-------|
+| 1,000 | 4-5 min | Quick |
+| 5,000 | 20-25 min | Reasonable |
+| 10,000 | 40-50 min | Acceptable |
+| 50,000 | 3.5-4.5 hours | Overnight job |
+| 100,000 | 7-9 hours | Long batch |
+| 500,000 | 35-45 hours | Not recommended |
+
+**Incremental Updates** (10% change daily):
+- 1,000 docs: ~30 sec
+- 10,000 docs: ~5 min
+- 50,000 docs: ~25 min
+
+### Search Capacity
+
+**Query Latency Budget**:
+- Embedding: 550ms
+- Vector search: 50-100ms
+- Permission verification: 50-100ms
+- **Total**: 650-750ms
+
+**Concurrent Users** (assuming 1 search every 5 minutes):
+- 10 users: 2 queries/min → Comfortable
+- 50 users: 10 queries/min → Near limit
+- 100 users: 20 queries/min → Over capacity
+
+**Peak Load** (all users search at once):
+- Parallelism: ~4 concurrent
+- Queue time: Proportional to position
+- 10 simultaneous: ~1.5-2 sec for last user
+- 50 simultaneous: ~7-10 sec for last user
+
+## Recommendations
+
+### Immediate Actions (Development)
+
+1. **✅ Use Ollama as-is**
+   - Current setup is perfect for dev/testing
+   - No changes needed
+   - Start building semantic search
+
+2. **Configuration**:
+   ```bash
+   OLLAMA_URL=https://ollama.internal.coutinho.io
+   OLLAMA_MODEL=nomic-embed-text
+   VECTOR_SYNC_BATCH_SIZE=10
+   ```
+
+3. **Add Monitoring**:
+   ```python
+   # Track these metrics
+   - embedding_latency_seconds (histogram)
+   - embedding_batch_size (gauge)
+   - embedding_errors_total (counter)
+   ```
+
+### Short-Term (Small Production)
+
+1. **Optimize Batching**:
+   - Use batch size 10-12 (quality sweet spot)
+   - Process during off-hours
+   - Implement incremental sync
+
+2. **Add Caching**:
+   ```python
+   # Cache common query embeddings
+   @lru_cache(maxsize=1000)
+   async def embed_with_cache(query: str):
+       return await ollama.embed(query)
+   ```
+
+3. **Monitor Metrics**:
+   - P50, P95, P99 latency
+   - Throughput (embeddings/sec)
+   - Error rates
+
+### Medium-Term (If Scaling Up)
+
+1. **Add Infinity Container** (when >50 users or latency issues):
+   ```yaml
+   services:
+     infinity:
+       image: michaelf34/infinity:latest
+       # Local to MCP server - ~10-20ms latency
+   ```
+
+2. **Implement Tiered Fallback**:
+   ```
+   Infinity (local, fast) → Ollama (remote, slower) → Local model
+   ```
+
+3. **Load Testing**:
+   - Simulate 50-100 concurrent users
+   - Measure actual throughput limits
+   - Identify breaking points
+
+### Long-Term (Enterprise Scale)
+
+1. **Migrate to TEI Cluster** (when >100 users):
+   - GPU-accelerated
+   - Horizontal scaling
+   - <20ms latency
+
+2. **Consider Managed Services**:
+   - Pinecone, Qdrant Cloud
+   - Removes operational burden
+   - Better SLAs
+
+## Testing Recommendations
+
+### Load Testing Script
+
+```bash
+# Test sustained load
+for i in {1..100}; do
+  curl -s https://ollama.internal.coutinho.io/api/embed \
+    -d "{\"model\": \"nomic-embed-text\", \"input\": \"Test $i\"}" &
+
+  # Rate limit: 5 concurrent
+  if [ $(($i % 5)) -eq 0 ]; then
+    wait
+    sleep 1
+  fi
+done
+```
+
+### Metrics to Collect
+
+1. **Latency Distribution**:
+   - P50 (median)
+   - P95 (acceptable)
+   - P99 (outliers)
+
+2. **Throughput**:
+   - Embeddings/second
+   - Peak vs sustained
+
+3. **Error Rates**:
+   - Timeouts
+   - Server errors
+   - Quality issues
+
+## Conclusion
+
+**Your Ollama instance is ready for development and small production use!**
+
+**Current Capacity**:
+- ✅ Development: Unlimited
+- ✅ Small prod (10-20 users, 10k docs): Comfortable
+- ⚠️ Medium prod (50 users, 50k docs): Monitoring needed
+- ❌ Large prod (>100 users): Migrate to Infinity/TEI
+
+**Key Strengths**:
+- Fully operational
+- Good parallelism
+- Acceptable latency for most use cases
+- Easy to integrate
+
+**Key Limitations**:
+- Network latency adds 300-400ms overhead
+- Batch quality issues at >16 items
+- Limited scalability beyond 50 users
+
+**Recommendation**:
+Start using Ollama immediately for development. Add monitoring and plan for Infinity when you approach 50 users or experience latency issues. The abstraction layer in ADR-003 makes migration seamless.
+
+**Next Steps**:
+1. Configure MCP server with Ollama URL
+2. Implement semantic search tools
+3. Add basic monitoring
+4. Test with real workload
+5. Scale up as needed
@@ -0,0 +1,796 @@
+# Ollama Embeddings Investigation
+
+**Date**: 2025-10-30
+**Status**: Recommendation for Integration
+
+## Executive Summary
+
+Ollama provides a **local, self-hosted embedding solution** that is excellent for **development and small-scale deployments** but has **performance limitations** compared to specialized embedding inference engines (TEI, Infinity).
+
+**Recommendation**: Include Ollama as **Tier 2 fallback** in our embedding strategy (after cloud APIs, before local sentence-transformers), prioritizing ease of setup over maximum performance.
+
+## Overview
+
+Ollama is primarily known as a local LLM runner but added embedding model support in version 0.1.26, making it a convenient option for generating vector embeddings without external API dependencies.
+
+### Key Characteristics
+
+- **Local & Self-Hosted**: No external API calls, full privacy
+- **Easy Setup**: Single binary, simple model downloads (`ollama pull nomic-embed-text`)
+- **Unified Platform**: Same tool for both LLMs and embeddings
+- **OpenAI Compatible**: `/v1/embeddings` endpoint for drop-in replacement
+- **Multi-Platform**: Linux, macOS, Windows support
+- **GPU Support**: CUDA, ROCm, Metal acceleration
+
+## API Details
+
+### Endpoint Structure
+
+**New API** (recommended):
+```bash
+POST http://localhost:11434/api/embed
+```
+
+**OpenAI Compatible**:
+```bash
+POST http://localhost:11434/v1/embeddings
+```
+
+**Legacy API** (deprecated):
+```bash
+POST http://localhost:11434/api/embeddings
+```
+
+### Request Format
+
+**Single Text Embedding**:
+```json
+{
+  "model": "nomic-embed-text",
+  "input": "Text to embed"
+}
+```
+
+**Batch Embedding** (since v0.2.0):
+```json
+{
+  "model": "nomic-embed-text",
+  "input": [
+    "First text to embed",
+    "Second text to embed",
+    "Third text to embed"
+  ]
+}
+```
+
+### Response Format
+
+```json
+{
+  "model": "nomic-embed-text",
+  "embeddings": [
+    [0.123, -0.456, 0.789, ...],  // 768 dimensions for nomic-embed-text
+    [0.234, -0.567, 0.890, ...]
+  ]
+}
+```
+
+### Python Integration
+
+```python
+import ollama
+
+# Single embedding
+response = ollama.embed(
+    model='nomic-embed-text',
+    input='Text to embed'
+)
+embedding = response['embeddings'][0]
+
+# Batch embeddings (more efficient)
+response = ollama.embed(
+    model='nomic-embed-text',
+    input=[
+        'First text',
+        'Second text',
+        'Third text'
+    ]
+)
+embeddings = response['embeddings']
+```
+
+## Available Models
+
+### 1. nomic-embed-text (Recommended)
+
+**Specifications**:
+- **Parameters**: 137M
+- **Dimensions**: 768
+- **Context Length**: 8,192 tokens (2K effective)
+- **Size**: 274MB
+- **Architecture**: BERT-based
+
+**Performance**:
+- Outperforms OpenAI `text-embedding-ada-002` and `text-embedding-3-small`
+- Excellent for long-context tasks
+- Strong general-purpose performance
+
+**Use Cases**:
+- General RAG applications
+- Long document processing
+- Semantic search
+- Document clustering
+
+**Pull Command**:
+```bash
+ollama pull nomic-embed-text
+```
+
+### 2. mxbai-embed-large
+
+**Specifications**:
+- **Parameters**: 334M
+- **Dimensions**: 1,024
+- **Context Length**: 512 tokens
+- **Architecture**: BERT-large optimized
+
+**Performance**:
+- Claims to outperform commercial models
+- Higher precision for complex queries
+- Best quality but slower
+
+**Use Cases**:
+- High-precision semantic search
+- Enterprise knowledge bases
+- Multilingual content
+
+**Pull Command**:
+```bash
+ollama pull mxbai-embed-large
+```
+
+### 3. all-minilm
+
+**Specifications**:
+- **Parameters**: 23M
+- **Dimensions**: 384
+- **Context Length**: 256 tokens
+- **Size**: Smallest footprint
+
+**Performance**:
+- Fastest processing speed
+- Good for sentence-level tasks
+- Limited context window
+
+**Use Cases**:
+- Real-time applications
+- Resource-constrained environments
+- High-throughput scenarios
+- Development/testing
+
+**Pull Command**:
+```bash
+ollama pull all-minilm
+```
+
+## Performance Benchmarks
+
+### Throughput Comparison
+
+| Hardware | Model | Batch Size | Throughput | Notes |
+|----------|-------|------------|------------|-------|
+| RTX 4090 (24GB) | nomic-embed-text | 256 | 12,450 tok/sec | GPU-accelerated |
+| RTX 4090 (24GB) | mxbai-embed-large | 128 | 8,920 tok/sec | GPU-accelerated |
+| Intel i9-13900K (CPU) | nomic-embed-text | 32 | 3,250 tok/sec | CPU-only |
+| Intel i9-13900K (CPU) | mxbai-embed-large | 16 | 2,180 tok/sec | CPU-only |
+
+### Latency Comparison
+
+**Single Request Latency** (RTX 4060):
+- Ollama: ~99ms
+- TEI: ~20ms (5x faster)
+- Infinity: ~30-40ms (2.5-3x faster)
+
+**Batch Processing**:
+- Optimal batch size: 32-64 (model dependent)
+- Performance degrades with batches >16 (quality issues reported)
+- 2x slower than direct sentence-transformers usage
+
+### Engine Comparison
+
+Based on benchmarks from Baseten (2024):
+
+| Engine | Relative Throughput | Notes |
+|--------|---------------------|-------|
+| BEI | 9.0x (baseline) | Fastest (proprietary) |
+| TEI | 4.5x | Open source, Rust-based |
+| Infinity | 3.5x | PyTorch/ONNX optimized |
+| vLLM | 3.0x | General LLM inference |
+| **Ollama** | **1.0x** | Slowest for embeddings |
+
+**Key Insight**: Ollama is **5-9x slower** than specialized embedding engines but trades performance for ease of use and unified platform.
+
+## Integration Implementation
+
+### Python Client Wrapper
+
+```python
+# nextcloud_mcp_server/embeddings/ollama.py
+import httpx
+from typing import List
+
+
+class OllamaEmbedding:
+    """Ollama embedding provider"""
+
+    def __init__(
+        self,
+        base_url: str = "http://localhost:11434",
+        model: str = "nomic-embed-text"
+    ):
+        self.base_url = base_url.rstrip("/")
+        self.model = model
+        self.client = httpx.AsyncClient(timeout=60.0)
+
+        # Model dimension mapping
+        self.dimensions = {
+            "nomic-embed-text": 768,
+            "mxbai-embed-large": 1024,
+            "all-minilm": 384
+        }
+        self.dimension = self.dimensions.get(model, 768)
+
+    async def embed(self, text: str) -> List[float]:
+        """Generate embedding for single text"""
+        response = await self.client.post(
+            f"{self.base_url}/api/embed",
+            json={
+                "model": self.model,
+                "input": text
+            }
+        )
+        response.raise_for_status()
+        data = response.json()
+        return data["embeddings"][0]
+
+    async def embed_batch(
+        self,
+        texts: List[str],
+        batch_size: int = 32
+    ) -> List[List[float]]:
+        """
+        Generate embeddings for multiple texts in batches.
+
+        Note: Ollama has reported quality issues with batch sizes >16.
+        We use batch_size=32 as default but allow configuration.
+        """
+        all_embeddings = []
+
+        # Process in chunks to avoid batch size issues
+        for i in range(0, len(texts), batch_size):
+            batch = texts[i:i + batch_size]
+
+            response = await self.client.post(
+                f"{self.base_url}/api/embed",
+                json={
+                    "model": self.model,
+                    "input": batch
+                }
+            )
+            response.raise_for_status()
+            data = response.json()
+            all_embeddings.extend(data["embeddings"])
+
+        return all_embeddings
+
+    async def check_health(self) -> bool:
+        """Check if Ollama server is running and model is available"""
+        try:
+            # Check if server is up
+            response = await self.client.get(f"{self.base_url}/api/tags")
+            response.raise_for_status()
+
+            # Check if model is pulled
+            models = response.json().get("models", [])
+            model_names = [m["name"] for m in models]
+
+            if self.model not in model_names:
+                raise ValueError(
+                    f"Model '{self.model}' not found. "
+                    f"Run: ollama pull {self.model}"
+                )
+
+            return True
+
+        except Exception as e:
+            raise ConnectionError(f"Ollama health check failed: {e}")
+
+    async def close(self):
+        """Close HTTP client"""
+        await self.client.aclose()
+```
+
+### Auto-Detection in Embedding Service
+
+```python
+# nextcloud_mcp_server/embeddings/service.py
+from typing import Optional
+import os
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class EmbeddingService:
+    """Unified embedding service with automatic provider detection"""
+
+    def __init__(self):
+        self.provider = None
+        self._detect_provider()
+
+    def _detect_provider(self):
+        """Auto-detect available embedding provider"""
+
+        # Tier 1: OpenAI API (best quality)
+        if os.getenv("OPENAI_API_KEY"):
+            from .openai import OpenAIEmbedding
+            self.provider = OpenAIEmbedding(
+                model=os.getenv("OPENAI_EMBEDDING_MODEL", "text-embedding-3-small"),
+                api_key=os.getenv("OPENAI_API_KEY")
+            )
+            logger.info("✓ Using OpenAI embeddings")
+            return
+
+        # Tier 2a: Infinity (optimized self-hosted)
+        if os.getenv("INFINITY_URL"):
+            from .infinity import InfinityEmbedding
+            try:
+                self.provider = InfinityEmbedding(
+                    url=os.getenv("INFINITY_URL"),
+                    model=os.getenv("EMBEDDING_MODEL", "BAAI/bge-small-en-v1.5")
+                )
+                logger.info("✓ Using Infinity embeddings (optimized)")
+                return
+            except Exception as e:
+                logger.warning(f"Infinity unavailable: {e}")
+
+        # Tier 2b: Ollama (easy self-hosted)
+        if os.getenv("OLLAMA_URL"):
+            from .ollama import OllamaEmbedding
+            try:
+                self.provider = OllamaEmbedding(
+                    base_url=os.getenv("OLLAMA_URL", "http://localhost:11434"),
+                    model=os.getenv("OLLAMA_MODEL", "nomic-embed-text")
+                )
+                # Verify Ollama is running and model is available
+                import asyncio
+                asyncio.run(self.provider.check_health())
+                logger.info("✓ Using Ollama embeddings (easy setup)")
+                return
+            except Exception as e:
+                logger.warning(f"Ollama unavailable: {e}")
+
+        # Tier 3: Local model (fallback)
+        logger.warning("No cloud/hosted embeddings available, using local model")
+        from .local import LocalEmbedding
+        self.provider = LocalEmbedding(
+            model=os.getenv("LOCAL_EMBEDDING_MODEL", "all-MiniLM-L6-v2")
+        )
+        logger.info("✓ Using local embeddings (CPU fallback)")
+
+    async def embed(self, text: str):
+        """Generate embedding for text"""
+        return await self.provider.embed(text)
+
+    async def embed_batch(self, texts: list[str]):
+        """Generate embeddings for multiple texts"""
+        return await self.provider.embed_batch(texts)
+
+    @property
+    def dimension(self) -> int:
+        """Get embedding dimension"""
+        return self.provider.dimension
+```
+
+### Docker Compose Configuration
+
+```yaml
+services:
+  # Ollama embedding service
+  ollama:
+    image: ollama/ollama:latest
+    restart: always
+    ports:
+      - 127.0.0.1:11434:11434
+    volumes:
+      - ollama_models:/root/.ollama
+    # Optional: GPU support
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+    # Pull models on startup
+    entrypoint: ["/bin/sh", "-c"]
+    command:
+      - |
+        ollama serve &
+        sleep 5
+        ollama pull nomic-embed-text
+        wait
+
+  # MCP Server with Ollama embeddings
+  mcp:
+    build: .
+    depends_on:
+      - ollama
+    environment:
+      # ... other vars ...
+      - OLLAMA_URL=http://ollama:11434
+      - OLLAMA_MODEL=nomic-embed-text
+
+  # Vector sync worker
+  mcp-vector-sync:
+    build: .
+    command: ["python", "-m", "nextcloud_mcp_server.sync.vector_indexer"]
+    depends_on:
+      - ollama
+      - qdrant
+    environment:
+      # ... other vars ...
+      - OLLAMA_URL=http://ollama:11434
+      - OLLAMA_MODEL=nomic-embed-text
+
+volumes:
+  ollama_models:
+```
+
+## Advantages of Ollama
+
+### 1. **Ease of Setup**
+
+```bash
+# Install Ollama
+curl -fsSL https://ollama.com/install.sh | sh
+
+# Pull embedding model
+ollama pull nomic-embed-text
+
+# Done! API available at localhost:11434
+```
+
+No complex configuration, no Docker registries, no model conversion.
+
+### 2. **Privacy & Data Sovereignty**
+
+- All processing happens locally
+- No data leaves your infrastructure
+- No API keys or external dependencies
+- Ideal for sensitive content (medical, legal, financial)
+
+### 3. **Unified Platform**
+
+- Same tool for LLMs and embeddings
+- Consistent API across model types
+- Single point of management
+- Simplified operations
+
+### 4. **Developer Experience**
+
+- Simple API (similar to OpenAI)
+- Good documentation
+- Active community
+- Framework integrations (LangChain, LlamaIndex)
+
+### 5. **Cost**
+
+- Free and open source
+- No per-token API costs
+- Only infrastructure costs (compute)
+
+### 6. **Model Variety**
+
+Growing library of embedding models:
+- nomic-embed-text (general purpose)
+- mxbai-embed-large (high quality)
+- all-minilm (fast)
+- More models added regularly
+
+## Limitations of Ollama
+
+### 1. **Performance**
+
+- **5-9x slower** than specialized engines (TEI, Infinity)
+- Not optimized specifically for embedding inference
+- Batch processing issues at larger batch sizes (>16)
+- Higher latency compared to alternatives
+
+### 2. **Scalability**
+
+- Single-instance deployment (no native clustering)
+- Limited concurrent request handling
+- Not designed for high-throughput production
+- Resource usage per request is higher
+
+### 3. **Batch Processing Issues**
+
+- Quality degradation reported with large batches
+- Optimal batch size: 32-64 (conservative)
+- Less efficient than specialized engines
+- GitHub issues tracking batch problems (#6262)
+
+### 4. **Resource Usage**
+
+- Models stay loaded in memory (VRAM/RAM)
+- Higher memory footprint per model
+- GPU context switching overhead
+- Not as memory-efficient as specialized engines
+
+### 5. **Production Features**
+
+- No built-in load balancing
+- Limited monitoring/metrics
+- No automatic scaling
+- Basic error handling
+
+## Use Case Recommendations
+
+### ✅ **Excellent For:**
+
+1. **Development & Testing**
+   - Quick setup for prototyping
+   - Local development environments
+   - Testing embedding pipelines
+
+2. **Small Deployments**
+   - <10 users
+   - <10,000 documents
+   - Infrequent searches (<100/day)
+   - Hobbyist/personal projects
+
+3. **Privacy-Critical Applications**
+   - Medical/healthcare records
+   - Legal documents
+   - Financial data
+   - Air-gapped environments
+
+4. **Unified LLM Stack**
+   - Projects already using Ollama for LLMs
+   - Simplified operations
+   - Consistent tooling
+
+5. **Educational/Learning**
+   - Teaching RAG concepts
+   - Learning embeddings
+   - Hackathons/workshops
+
+### ⚠️ **Consider Alternatives For:**
+
+1. **Production at Scale**
+   - >100 users
+   - >100,000 documents
+   - High query volume (>1000/day)
+   - Use: TEI or Infinity
+
+2. **Performance-Critical**
+   - Real-time search (<50ms latency)
+   - High-throughput batch processing
+   - Use: TEI with GPU
+
+3. **Enterprise Deployments**
+   - Need for high availability
+   - Load balancing requirements
+   - Advanced monitoring
+   - Use: Managed services or TEI cluster
+
+4. **Large-Scale Indexing**
+   - Millions of documents
+   - Continuous high-volume ingestion
+   - Use: Infinity or commercial solutions
+
+## Integration Strategy
+
+### Recommended Tier Placement
+
+**Update ADR-003 embedding strategy:**
+
+```
+Tier 1: OpenAI API (best quality, requires API key)
+  ↓ fallback
+Tier 2a: Infinity (optimized self-hosted, complex setup)
+  ↓ fallback
+Tier 2b: Ollama (easy self-hosted, moderate performance) ← NEW
+  ↓ fallback
+Tier 3: Local sentence-transformers (CPU fallback, simplest)
+```
+
+### Configuration
+
+```bash
+# Option 1: Use Infinity (if available)
+INFINITY_URL=http://infinity:7997
+EMBEDDING_MODEL=BAAI/bge-small-en-v1.5
+
+# Option 2: Use Ollama (if Infinity unavailable)
+OLLAMA_URL=http://ollama:11434
+OLLAMA_MODEL=nomic-embed-text
+
+# Option 3: Use local model (automatic fallback)
+# No configuration needed
+```
+
+### When to Choose Ollama
+
+**Choose Ollama if**:
+- You're already using Ollama for LLMs
+- You need privacy/data sovereignty
+- You have <10k documents and <100 users
+- Ease of setup is more important than max performance
+- You're in development/testing phase
+
+**Choose Infinity/TEI if**:
+- You need maximum throughput (>1000 embeddings/sec)
+- You have >100k documents
+- Latency is critical (<50ms)
+- You're in production with >100 users
+
+**Choose OpenAI API if**:
+- You're okay with cloud dependencies
+- You need best-in-class quality
+- Cost is not a concern (~$0.02 per 1M tokens)
+
+## Production Deployment Guidance
+
+### Small Production (Ollama Acceptable)
+
+**Profile**:
+- 5-20 users
+- 1,000-10,000 documents
+- 50-200 searches/day
+- <2 sec acceptable latency
+
+**Configuration**:
+```yaml
+ollama:
+  image: ollama/ollama:latest
+  deploy:
+    resources:
+      limits:
+        memory: 4GB
+        cpus: "2.0"
+      reservations:
+        devices:
+          - driver: nvidia  # GPU if available
+            count: 1
+            capabilities: [gpu]
+  environment:
+    - OLLAMA_NUM_PARALLEL=2  # Concurrent requests
+```
+
+**Expected Performance**:
+- Embedding latency: 100-200ms
+- Throughput: 5-10 embeddings/sec
+- Memory: 2-3GB (model loaded)
+
+### Medium Production (Use Infinity/TEI)
+
+**Profile**:
+- 20-200 users
+- 10,000-1M documents
+- 500-5,000 searches/day
+- <500ms acceptable latency
+
+**Recommendation**: Migrate to Infinity or TEI
+```yaml
+infinity:
+  image: michaelf34/infinity:latest
+  # Better throughput and latency
+```
+
+### Large Production (Use Specialized Solution)
+
+**Profile**:
+- >200 users
+- >1M documents
+- >5,000 searches/day
+- <100ms required latency
+
+**Recommendation**: Use TEI cluster or commercial service
+
+## Monitoring Considerations
+
+### Key Metrics to Track
+
+```python
+# Add Ollama-specific metrics
+from prometheus_client import Histogram, Counter, Gauge
+
+ollama_embedding_latency = Histogram(
+    'ollama_embedding_duration_seconds',
+    'Ollama embedding generation time',
+    ['model', 'batch_size']
+)
+
+ollama_batch_size = Gauge(
+    'ollama_batch_size',
+    'Current batch size being processed'
+)
+
+ollama_errors = Counter(
+    'ollama_errors_total',
+    'Ollama embedding errors',
+    ['error_type']
+)
+```
+
+### Health Checks
+
+```python
+async def ollama_health_check():
+    """Check Ollama availability"""
+    try:
+        async with httpx.AsyncClient() as client:
+            # Check server
+            response = await client.get("http://ollama:11434/api/tags")
+            response.raise_for_status()
+
+            # Verify model loaded
+            models = response.json().get("models", [])
+            if "nomic-embed-text" not in [m["name"] for m in models]:
+                return False, "Model not pulled"
+
+            return True, "OK"
+    except Exception as e:
+        return False, str(e)
+```
+
+## Migration Path
+
+### Starting with Ollama
+
+**Phase 1: Development** (Ollama)
+- Use Ollama for initial development
+- Validate embedding pipeline
+- Test search quality
+
+**Phase 2: Growth** (Ollama → Infinity)
+- Monitor performance metrics
+- When >50 users or >10k docs, migrate to Infinity
+- Simple config change, no code changes
+
+**Phase 3: Scale** (Infinity → TEI/Commercial)
+- When >200 users or performance issues
+- Consider TEI cluster or managed services
+
+### Code Compatibility
+
+All embedding providers use the same interface:
+```python
+# Works with Ollama, Infinity, OpenAI, Local
+embedding = await embedding_service.embed(text)
+embeddings = await embedding_service.embed_batch(texts)
+```
+
+**Migration is a configuration change only** - no code rewrite needed.
+
+## Conclusion
+
+**Ollama is a solid choice for:**
+- Early-stage projects
+- Development/testing
+- Privacy-critical applications
+- Small deployments (<10 users, <10k docs)
+- Unified LLM + embedding stack
+
+**But recognize its limitations:**
+- 5-9x slower than specialized engines
+- Not designed for high-throughput production
+- Batch processing can be problematic
+- Limited scalability
+
+**Recommendation**:
+✅ **Include Ollama as Tier 2b** (after Infinity, before local models) in the embedding strategy. It provides a good balance of ease-of-use and privacy for small-to-medium deployments while allowing seamless migration to more performant engines as needs grow.
+
+The key is designing the abstraction layer (as done in ADR-003) so migration between engines requires only configuration changes, not code rewrites.
@@ -1,6 +1,6 @@
 [project]
 name = "nextcloud-mcp-server"
-version = "0.21.0"
+version = "0.22.7"
 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"}
@@ -65,6 +65,13 @@ version_scheme = "pep440"
 version_provider = "uv"
 update_changelog_on_bump = true
 major_version_zero = true
+version_files = [
+    "charts/nextcloud-mcp-server/Chart.yaml:appVersion",
+    "charts/nextcloud-mcp-server/Chart.yaml:version"
+]
+ignored_tag_formats = [
+    "nextcloud-mcp-server-*"
+]

 [tool.ruff.lint]
 extend-select = ["I"]
@@ -3,8 +3,8 @@ Tests for Dynamic Client Registration (DCR) token_type parameter.

 These tests verify that the Nextcloud OIDC server properly honors the token_type
 parameter during client registration, issuing the correct type of access tokens:
- token_type="JWT" → JWT-formatted tokens (RFC 9068)
- token_type="Bearer" → Opaque tokens (standard OAuth2)
+- token_type="jwt" → JWT-formatted tokens (RFC 9068)
+- token_type="opaque" → Opaque tokens (standard OAuth2)

 This is critical for ensuring:
 1. Client choice is respected by the OIDC server
@@ -208,12 +208,14 @@ async def test_dcr_respects_jwt_token_type(
    oauth_callback_server,
 ):
    """
-    Test that DCR honors token_type=JWT and issues JWT-formatted tokens.
+    Test that DCR honors token_type=jwt and issues JWT-formatted tokens.

    This verifies:
-    1. Client registration with token_type="JWT" succeeds
+    1. Client registration with token_type="jwt" succeeds
    2. Tokens obtained via this client are JWT format (base64.base64.signature)
    3. JWT payload contains expected claims (sub, iss, scope, etc.)
+
+    Note: The OIDC app uses lowercase 'jwt' (not 'JWT').
    """
    nextcloud_host = os.getenv("NEXTCLOUD_HOST")
    if not nextcloud_host:
@@ -232,15 +234,15 @@ async def test_dcr_respects_jwt_token_type(
        token_endpoint = oidc_config.get("token_endpoint")
        authorization_endpoint = oidc_config.get("authorization_endpoint")

-    # Register client with token_type="JWT"
-    logger.info("Registering OAuth client with token_type=JWT...")
+    # Register client with token_type="jwt"
+    logger.info("Registering OAuth client with token_type=jwt...")
    client_info = await register_client(
        nextcloud_url=nextcloud_host,
        registration_endpoint=registration_endpoint,
        client_name="DCR Test - JWT Token Type",
        redirect_uris=[callback_url],
        scopes="openid profile email notes:read notes:write",
-        token_type="JWT",
+        token_type="jwt",
    )

    logger.info(f"Registered JWT client: {client_info.client_id[:16]}...")
@@ -278,7 +280,7 @@ async def test_dcr_respects_jwt_token_type(
    assert "notes:write" in scopes, "JWT scope claim missing notes:write"

    logger.info(
-        f"✅ DCR with token_type=JWT works correctly! "
+        f"✅ DCR with token_type=jwt works correctly! "
        f"Token is JWT format with scope claim: {payload['scope']}"
    )

@@ -290,12 +292,14 @@ async def test_dcr_respects_bearer_token_type(
    oauth_callback_server,
 ):
    """
-    Test that DCR honors token_type=Bearer and issues opaque tokens.
+    Test that DCR honors token_type=opaque and issues opaque tokens.

    This verifies:
-    1. Client registration with token_type="Bearer" succeeds
+    1. Client registration with token_type="opaque" succeeds
    2. Tokens obtained via this client are opaque (NOT JWT format)
    3. Opaque tokens are simple strings, not base64-encoded structures
+
+    Note: The OIDC app uses 'opaque' or 'jwt' as token_type values (not 'Bearer').
    """
    nextcloud_host = os.getenv("NEXTCLOUD_HOST")
    if not nextcloud_host:
@@ -314,18 +318,18 @@ async def test_dcr_respects_bearer_token_type(
        token_endpoint = oidc_config.get("token_endpoint")
        authorization_endpoint = oidc_config.get("authorization_endpoint")

-    # Register client with token_type="Bearer" (opaque tokens)
-    logger.info("Registering OAuth client with token_type=Bearer...")
+    # Register client with token_type="opaque" (opaque tokens)
+    logger.info("Registering OAuth client with token_type=opaque...")
    client_info = await register_client(
        nextcloud_url=nextcloud_host,
        registration_endpoint=registration_endpoint,
-        client_name="DCR Test - Bearer Token Type",
+        client_name="DCR Test - Opaque Token Type",
        redirect_uris=[callback_url],
        scopes="openid profile email notes:read notes:write",
-        token_type="Bearer",
+        token_type="opaque",
    )

-    logger.info(f"Registered Bearer client: {client_info.client_id[:16]}...")
+    logger.info(f"Registered Opaque token client: {client_info.client_id[:16]}...")

    # Obtain token via OAuth flow
    access_token = await get_oauth_token_with_client(
@@ -353,7 +357,7 @@ async def test_dcr_respects_bearer_token_type(
        pass

    logger.info(
-        f"✅ DCR with token_type=Bearer works correctly! "
+        f"✅ DCR with token_type=opaque works correctly! "
        f"Token is opaque (not JWT format): {access_token[:30]}..."
    )

@@ -941,7 +941,7 @@ wheels = [

 [[package]]
 name = "nextcloud-mcp-server"
-version = "0.21.0"
+version = "0.22.7"
 source = { editable = "." }
 dependencies = [
    { name = "caldav" },
Author	SHA1	Message	Date
Chris Coutinho	4a3b80cb98	fix: Update DCR token_type tests for OIDC app changes The Nextcloud OIDC app has updated token_type parameter values: - Changed from "Bearer" → "opaque" for opaque tokens - Changed from "JWT" → "jwt" for JWT tokens Updated test_dcr_token_type.py to use lowercase token_type values: - token_type="jwt" for JWT-formatted tokens - token_type="opaque" for opaque/bearer tokens This fixes test failures where tests were using the old "Bearer" and "JWT" (uppercase) values which are no longer recognized by the OIDC app. Fixes test: test_dcr_respects_bearer_token_type 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>	2025-10-31 22:30:58 +01:00
Chris Coutinho	fc3ab8d0ac	docs: Add Ollama embeddings capacity analysis and investigation Documents Ollama embedding service evaluation for ADR-003 semantic search implementation, including performance benchmarks and capacity analysis. ## Documentation ### Ollama Capacity Analysis - Performance metrics for ollama.internal.coutinho.io - Model: nomic-embed-text:latest - Embedding generation benchmarks (single, batch, parallel) - Latency analysis and throughput measurements - Resource usage and capacity recommendations ### Ollama Embeddings Investigation - Evaluation of Ollama for semantic search use case - Comparison with other embedding providers - Integration considerations with ADR-003 architecture - Deployment scenarios and operational requirements ## Key Findings ✅ Ollama instance operational and performing well ✅ Reasonable latency for small-medium workloads ✅ Good parallelism support ✅ Suitable for development and small production deployments ## References - ADR-003: Vector Database Semantic Search - Ollama API documentation - nomic-embed-text model specifications 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>	2025-10-31 03:07:44 +01:00
Chris Coutinho	0f03541486	Merge branch 'master' of github.com:cbcoutinho/nextcloud-mcp-server	2025-10-31 02:59:53 +01:00
Chris Coutinho	ef07b1a6c9	docs: Add ADRs	2025-10-31 02:59:44 +01:00
Chris Coutinho	4f82357f24	ci: update submodule	2025-10-31 02:59:35 +01:00
Chris Coutinho	c4293b6750	Merge pull request #251 from cbcoutinho/renovate/docker.io-library-nginx-alpine chore(deps): update docker.io/library/nginx:alpine docker digest to b3c656d	2025-10-30 20:23:52 +01:00
renovate-bot-cbcoutinho[bot]	72e4eb3d19	chore(deps): update docker.io/library/nginx:alpine docker digest to b3c656d	2025-10-30 17:06:28 +00:00
Chris Coutinho	47dd2df7aa	Merge pull request #250 from cbcoutinho/renovate/ghcr.io-astral-sh-uv-0.x chore(deps): update ghcr.io/astral-sh/uv docker tag to v0.9.6	2025-10-30 12:55:02 +01:00
renovate-bot-cbcoutinho[bot]	9fd2022151	chore(deps): update ghcr.io/astral-sh/uv docker tag to v0.9.6	2025-10-29 23:07:53 +00:00
Chris Coutinho	b99dc52c95	docs: Update README with instructions on helm install	2025-10-29 12:47:20 +01:00
Chris Coutinho	78b27fb5e9	Merge pull request #249 from cbcoutinho/renovate/actions-checkout-5.x chore(deps): update actions/checkout action to v5	2025-10-29 12:42:59 +01:00
renovate-bot-cbcoutinho[bot]	03e39a3f94	chore(deps): update actions/checkout action to v5	2025-10-29 11:28:09 +00:00
github-actions[bot]	5259658458	bump: version 0.22.6 → 0.22.7	2025-10-29 11:18:41 +00:00
Chris Coutinho	e03a3c2e83	fix(helm): Remove image tag overide	2025-10-29 12:18:12 +01:00
Chris Coutinho	94cbd3015d	Merge pull request #248 from cbcoutinho/renovate/pin-dependencies chore(deps): pin dependencies	2025-10-29 12:14:10 +01:00
renovate-bot-cbcoutinho[bot]	49a961cbcc	chore(deps): pin dependencies	2025-10-29 11:06:51 +00:00
github-actions[bot]	e1aca04aff	bump: version 0.22.5 → 0.22.6	2025-10-29 10:57:44 +00:00
Chris Coutinho	3b12e585ca	fix(helm): Update helm chart with extraArgs	2025-10-29 11:57:13 +01:00
github-actions[bot]	e647c87dd8	bump: version 0.22.4 → 0.22.5	2025-10-29 10:54:54 +00:00
Chris Coutinho	cb74157d51	fix: Update helm chart variables	2025-10-29 11:54:26 +01:00
github-actions[bot]	202058bdc8	bump: version 0.22.3 → 0.22.4	2025-10-29 10:44:11 +00:00
Chris Coutinho	c312911538	fix(helm): Update helm version with release	2025-10-29 11:43:30 +01:00
Chris Coutinho	e602684743	fix(helm): Update helm version with release	2025-10-29 11:43:02 +01:00
github-actions[bot]	8221046d8a	bump: version 0.22.2 → 0.22.3	2025-10-29 10:35:58 +00:00
Chris Coutinho	3e45b6ca25	fix(helm): Update helm version with release	2025-10-29 11:34:58 +01:00
github-actions[bot]	9ec7637579	bump: version 0.22.1 → 0.22.2	2025-10-29 10:30:39 +00:00
Chris Coutinho	670188f9e4	fix(helm): Update helm version with release	2025-10-29 11:29:59 +01:00
github-actions[bot]	3878beaf65	bump: version 0.22.0 → 0.22.1	2025-10-29 10:17:08 +00:00
Chris Coutinho	a5a0571bde	fix: Trigger release	2025-10-29 11:16:30 +01:00
github-actions[bot]	0e7e74867f	bump: version 0.21.0 → 0.22.0	2025-10-29 09:32:27 +00:00
Chris Coutinho	a29045cca4	Merge pull request #246 from cbcoutinho/feature/helm-chart Feature/helm chart	2025-10-29 10:32:02 +01:00
Chris Coutinho	3c3646bec2	Merge pull request #247 from cbcoutinho/renovate/docker.io-library-nginx-alpine chore(deps): update docker.io/library/nginx:alpine docker digest to 9dacca6	2025-10-29 09:37:07 +01:00
renovate-bot-cbcoutinho[bot]	dd636e6a08	chore(deps): update docker.io/library/nginx:alpine docker digest to 9dacca6	2025-10-29 05:07:08 +00:00
Chris Coutinho	e0de2e17e9	Merge pull request #245 from cbcoutinho/renovate/docker.io-library-nextcloud-32.0.1 chore(deps): update docker.io/library/nextcloud:32.0.1 docker digest to 1e4eae5	2025-10-28 09:19:39 +01:00
renovate-bot-cbcoutinho[bot]	4fc0cb5a41	chore(deps): update docker.io/library/nextcloud:32.0.1 docker digest to 1e4eae5	2025-10-27 23:10:34 +00:00
Chris Coutinho	ff9cca716b	Merge pull request #243 from cbcoutinho/renovate/astral-sh-setup-uv-digest chore(deps): update astral-sh/setup-uv digest to 8585678	2025-10-26 22:00:45 +01:00
Chris Coutinho	ef4a82e589	Update .github/workflows/release.yml	2025-10-26 22:00:36 +01:00
Chris Coutinho	301c502e57	Merge pull request #244 from cbcoutinho/renovate/astral-sh-setup-uv-7.x chore(deps): update astral-sh/setup-uv action to v7.1.2	2025-10-26 21:59:19 +01:00
renovate-bot-cbcoutinho[bot]	d4d291d6d2	chore(deps): update astral-sh/setup-uv action to v7.1.2	2025-10-26 17:07:33 +00:00
renovate-bot-cbcoutinho[bot]	e4b0ea5093	chore(deps): update astral-sh/setup-uv digest to 8585678	2025-10-26 17:07:29 +00:00