feat(semantic_search): support local OpenAI-compatible embedding servers (#22)

Adds a LocalEmbedder that targets any OpenAI-compatible embedding
endpoint (Ollama, llama.cpp, vLLM, LM Studio, ...). Zero new
Python dependencies — reuses the existing openai SDK with a
custom base_url. Defaults to Ollama at http://localhost:11434/v1
with nomic-embed-text @ 768 dims; override via env vars for other
servers/models (e.g. bge-m3 @ 1024 dims for better Turkish).

Refactors the shared encode/similarity logic into a private base
class so OpenRouterEmbedder and LocalEmbedder don't duplicate ~50
lines. OpenRouter keeps its ranking headers; local sends none.

Adds get_embedder() factory selecting the provider based on
EMBEDDING_PROVIDER (local) or OPENROUTER_API_KEY presence, and
is_semantic_search_available() that returns True for either path.
mcp_server_main now uses these so the semantic_search tool is
exposed when only a local server is configured.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.env.example

@@ -74,18 +74,31 @@ JWT_SECRET_KEY=your_jwt_secret_key_here
 # SEMANTIC SEARCH SETTINGS (Optional)
 # =============================================================================
 
-# OpenRouter API Key for semantic search functionality
+# Embedding provider for the semantic_search tool.
+# Pick exactly one of: OpenRouter (hosted) or Local (your own server).
+
+# --- Option A: OpenRouter (hosted, default) -----------------------------------
 # Get your API key from: https://openrouter.ai/keys
-# If not set, semantic search tool will be disabled
+# If neither this nor EMBEDDING_PROVIDER=local is set, semantic search is off.
 OPENROUTER_API_KEY=sk-or-v1-your_openrouter_api_key_here
 
-# Optional: override the embedding model and dimension.
+# Optional: override the OpenRouter embedding model and dimension.
 # Defaults: google/gemini-embedding-001 at 3072 dims (paid on OpenRouter).
-# Pick any embedding model from https://openrouter.ai/models?modality=embedding
+# Pick any model from https://openrouter.ai/models?modality=embedding
 # and set the dimension to that model's output size — they must match.
 # OPENROUTER_EMBEDDING_MODEL=google/gemini-embedding-001
 # OPENROUTER_EMBEDDING_DIMENSION=3072
 
+# --- Option B: Local OpenAI-compatible server (Ollama / llama.cpp / vLLM) -----
+# Uncomment to use your own server instead of OpenRouter (no API key required).
+# Defaults target Ollama with nomic-embed-text. For Turkish, bge-m3 (1024 dims)
+# tends to work better — pull it with: `ollama pull bge-m3`
+# EMBEDDING_PROVIDER=local
+# LOCAL_EMBEDDING_BASE_URL=http://localhost:11434/v1
+# LOCAL_EMBEDDING_MODEL=nomic-embed-text
+# LOCAL_EMBEDDING_DIMENSION=768
+# LOCAL_EMBEDDING_API_KEY=                          # most local servers ignore this
+
 # =============================================================================
 # USAGE INSTRUCTIONS
 # =============================================================================

mcp_server_main.py

@@ -265,17 +265,18 @@ from bedesten_mcp_module.models import (
 )
 from bedesten_mcp_module.enums import BirimAdiEnum
 
-# Semantic Search Module Imports (conditional based on OPENROUTER_API_KEY)
+# Semantic Search Module Imports (enabled if any embedding provider is configured)
-from semantic_search.embedder import is_openrouter_available
+from semantic_search.embedder import is_semantic_search_available, is_local_embedding_configured
-SEMANTIC_SEARCH_AVAILABLE = is_openrouter_available()
+SEMANTIC_SEARCH_AVAILABLE = is_semantic_search_available()
 
 if SEMANTIC_SEARCH_AVAILABLE:
-    from semantic_search.embedder import OpenRouterEmbedder
+    from semantic_search.embedder import get_embedder
     from semantic_search.vector_store import VectorStore
     from semantic_search.processor import DocumentProcessor
-    logger.info("Semantic search enabled (OPENROUTER_API_KEY found)")
+    provider = "local" if is_local_embedding_configured() else "openrouter"
+    logger.info(f"Semantic search enabled (provider={provider})")
 else:
-    logger.info("Semantic search disabled (OPENROUTER_API_KEY not set)")
+    logger.info("Semantic search disabled (no embedding provider configured)")
 
 from danistay_mcp_module.client import DanistayApiClient
 from emsal_mcp_module.client import EmsalApiClient
@@ -1279,8 +1280,9 @@ YANLIŞ KULLANIM:
         logger.info(f"Semantic search tool called with initial_keyword: {initial_keyword}, query: {query}")
 
         try:
-            # Initialize components
+            # Initialize components (provider chosen via EMBEDDING_PROVIDER /
-            embedder = OpenRouterEmbedder()
+            # OPENROUTER_API_KEY env vars)
+            embedder = get_embedder()
             vector_store = VectorStore(dimension=embedder.dimension)
             processor = DocumentProcessor(chunk_size=1500, chunk_overlap=300)
 

semantic_search/__init__.py

@@ -1,7 +1,23 @@
 # semantic_search/__init__.py
 
-from .embedder import OpenRouterEmbedder, is_openrouter_available
+from .embedder import (
+    OpenRouterEmbedder,
+    LocalEmbedder,
+    get_embedder,
+    is_openrouter_available,
+    is_local_embedding_configured,
+    is_semantic_search_available,
+)
 from .vector_store import VectorStore
 from .processor import DocumentProcessor
 
-__all__ = ['OpenRouterEmbedder', 'is_openrouter_available', 'VectorStore', 'DocumentProcessor']
+__all__ = [
+    'OpenRouterEmbedder',
+    'LocalEmbedder',
+    'get_embedder',
+    'is_openrouter_available',
+    'is_local_embedding_configured',
+    'is_semantic_search_available',
+    'VectorStore',
+    'DocumentProcessor',
+]

semantic_search/embedder.py

@@ -2,87 +2,70 @@
 
 import logging
 import os
-from typing import List, Optional
+from typing import Dict, List, Optional
 import numpy as np
 
 logger = logging.getLogger(__name__)
 
 
+# OpenRouter defaults (preserve backward compatibility)
 DEFAULT_MODEL = "google/gemini-embedding-001"
 DEFAULT_DIMENSION = 3072
 
+# Local provider defaults — Ollama with nomic-embed-text out of the box.
+# Override via LOCAL_EMBEDDING_BASE_URL / LOCAL_EMBEDDING_MODEL /
+# LOCAL_EMBEDDING_DIMENSION when using a different server or model
+# (e.g. llama.cpp's server, vLLM, LM Studio, or a different Ollama model
+# such as bge-m3 — better for Turkish — at 1024 dimensions).
+LOCAL_DEFAULT_BASE_URL = "http://localhost:11434/v1"
+LOCAL_DEFAULT_MODEL = "nomic-embed-text"
+LOCAL_DEFAULT_DIMENSION = 768
+
 
 def is_openrouter_available() -> bool:
     """Check if OpenRouter API key is available."""
     return bool(os.getenv("OPENROUTER_API_KEY"))
 
 
-class OpenRouterEmbedder:
+def is_local_embedding_configured() -> bool:
+    """Check if the user opted into a local embedding endpoint."""
+    return os.getenv("EMBEDDING_PROVIDER", "").strip().lower() == "local"
+
+
+def is_semantic_search_available() -> bool:
+    """Returns True if any embedding provider is configured."""
+    return is_local_embedding_configured() or is_openrouter_available()
+
+
+def _coerce_dimension(value, env_name: str, default: int) -> int:
+    """Parse a dimension value (int or str) with clear error messages."""
+    if value is None:
+        return default
+    try:
+        parsed = int(value)
+    except (TypeError, ValueError) as e:
+        raise ValueError(
+            f"{env_name} must be an integer, got {value!r}"
+        ) from e
+    if parsed <= 0:
+        raise ValueError(f"Embedding dimension must be positive, got {parsed}")
+    return parsed
+
+
+class _BaseOpenAICompatibleEmbedder:
     """
-    Embedder using OpenRouter's embedding API.
+    Shared encode/similarity logic for embedders backed by the OpenAI Python
-
+    SDK. Subclasses configure ``client``, ``model``, ``dimension``, and
-    The model and dimension are configurable so users can pick any OpenRouter
+    optionally ``_extra_headers`` (e.g. OpenRouter ranking headers).
-    embedding model (e.g. when one becomes paid or when a different model fits
-    the budget better). Configuration precedence: explicit constructor args >
-    environment variables > defaults.
-
-    Environment variables:
-        OPENROUTER_API_KEY (required): OpenRouter credential
-        OPENROUTER_EMBEDDING_MODEL (optional): override the embedding model id
-        OPENROUTER_EMBEDDING_DIMENSION (optional): override the vector size
-
-    Defaults preserve backward compatibility: ``google/gemini-embedding-001``
-    at 3072 dimensions.
     """
 
-    def __init__(self, model: Optional[str] = None, dimension: Optional[int] = None):
+    # Subclasses may override; sent on every embeddings.create call when set.
-        """
+    _extra_headers: Dict[str, str] = {}
-        Initialize OpenRouter Embedder.
 
-        Args:
+    # Set by subclasses
-            model: OpenRouter embedding model id. Falls back to
+    client = None
-                OPENROUTER_EMBEDDING_MODEL env var, then DEFAULT_MODEL.
+    model: str = ""
-            dimension: Output vector size. Falls back to
+    dimension: int = 0
-                OPENROUTER_EMBEDDING_DIMENSION env var, then DEFAULT_DIMENSION.
-                Must match the chosen model's actual output size — the vector
-                store and similarity math rely on it.
-
-        Raises:
-            ValueError: If OPENROUTER_API_KEY is not set or dimension is invalid
-            ImportError: If openai package is not installed
-        """
-        api_key = os.getenv("OPENROUTER_API_KEY")
-        if not api_key:
-            raise ValueError("OPENROUTER_API_KEY environment variable is not set")
-
-        try:
-            from openai import OpenAI
-        except ImportError:
-            raise ImportError("openai package is required. Install with: pip install openai")
-
-        self.client = OpenAI(
-            base_url="https://openrouter.ai/api/v1",
-            api_key=api_key,
-        )
-        self.model = model or os.getenv("OPENROUTER_EMBEDDING_MODEL") or DEFAULT_MODEL
-
-        dim_value = dimension if dimension is not None else os.getenv("OPENROUTER_EMBEDDING_DIMENSION")
-        if dim_value is None:
-            self.dimension = DEFAULT_DIMENSION
-        else:
-            try:
-                self.dimension = int(dim_value)
-            except (TypeError, ValueError) as e:
-                raise ValueError(
-                    f"OPENROUTER_EMBEDDING_DIMENSION must be an integer, got {dim_value!r}"
-                ) from e
-            if self.dimension <= 0:
-                raise ValueError(f"Embedding dimension must be positive, got {self.dimension}")
-
-        logger.info(
-            f"OpenRouter Embedder initialized with model: {self.model} "
-            f"(dimension={self.dimension})"
-        )
 
     def encode_query(self, query: str, task: str = "search result") -> np.ndarray:
         """
@@ -95,7 +78,6 @@ class OpenRouterEmbedder:
         Returns:
             Numpy array of embeddings (``self.dimension`` elements).
         """
-        # Apply query prompt template
         text = f"task: {task} | query: {query}"
 
         try:
@@ -103,10 +85,7 @@ class OpenRouterEmbedder:
                 model=self.model,
                 input=text,
                 encoding_format="float",
-                extra_headers={
+                extra_headers=self._extra_headers or None,
-                    "HTTP-Referer": "https://yargimcp.com",
-                    "X-Title": "Yargi MCP Server",
-                }
             )
 
             embedding = np.array(response.data[0].embedding, dtype=np.float32)
@@ -125,7 +104,7 @@ class OpenRouterEmbedder:
 
     def encode_documents(self, documents: List[str], titles: Optional[List[str]] = None) -> np.ndarray:
         """
-        Encode multiple documents with batch API call.
+        Encode multiple documents with a batch API call.
 
         Args:
             documents: List of document texts
@@ -137,28 +116,22 @@ class OpenRouterEmbedder:
         if not documents:
             return np.array([])
 
-        # Apply document prompt template
         texts = []
         for i, doc in enumerate(documents):
             title = titles[i] if titles and i < len(titles) else "none"
-            text = f"title: {title} | text: {doc}"
+            texts.append(f"title: {title} | text: {doc}")
-            texts.append(text)
 
         try:
             response = self.client.embeddings.create(
                 model=self.model,
                 input=texts,
                 encoding_format="float",
-                extra_headers={
+                extra_headers=self._extra_headers or None,
-                    "HTTP-Referer": "https://yargimcp.com",
-                    "X-Title": "Yargi MCP Server",
-                }
             )
 
-            # Extract embeddings in order
             embeddings = np.array(
                 [d.embedding for d in sorted(response.data, key=lambda x: x.index)],
-                dtype=np.float32
+                dtype=np.float32,
             )
 
             # L2 normalize each embedding for cosine similarity
@@ -183,11 +156,139 @@ class OpenRouterEmbedder:
         Returns:
             Similarity scores (N,)
         """
-        # Ensure query is 2D for matrix multiplication
         if len(query_embedding.shape) == 1:
             query_embedding = query_embedding.reshape(1, -1)
 
-        # Compute cosine similarity (embeddings are already normalized)
+        # Embeddings are already L2-normalized.
         similarities = np.dot(document_embeddings, query_embedding.T).squeeze()
-
         return similarities
+
+
+class OpenRouterEmbedder(_BaseOpenAICompatibleEmbedder):
+    """
+    Embedder using OpenRouter's embedding API.
+
+    The model and dimension are configurable so users can pick any OpenRouter
+    embedding model (e.g. when one becomes paid). Configuration precedence:
+    explicit constructor args > environment variables > defaults.
+
+    Environment variables:
+        OPENROUTER_API_KEY (required): OpenRouter credential
+        OPENROUTER_EMBEDDING_MODEL (optional): override the embedding model id
+        OPENROUTER_EMBEDDING_DIMENSION (optional): override the vector size
+
+    Defaults preserve backward compatibility: ``google/gemini-embedding-001``
+    at 3072 dimensions.
+    """
+
+    _extra_headers = {
+        "HTTP-Referer": "https://yargimcp.com",
+        "X-Title": "Yargi MCP Server",
+    }
+
+    def __init__(self, model: Optional[str] = None, dimension: Optional[int] = None):
+        api_key = os.getenv("OPENROUTER_API_KEY")
+        if not api_key:
+            raise ValueError("OPENROUTER_API_KEY environment variable is not set")
+
+        try:
+            from openai import OpenAI
+        except ImportError:
+            raise ImportError("openai package is required. Install with: pip install openai")
+
+        self.client = OpenAI(
+            base_url="https://openrouter.ai/api/v1",
+            api_key=api_key,
+        )
+        self.model = model or os.getenv("OPENROUTER_EMBEDDING_MODEL") or DEFAULT_MODEL
+        self.dimension = _coerce_dimension(
+            dimension if dimension is not None else os.getenv("OPENROUTER_EMBEDDING_DIMENSION"),
+            "OPENROUTER_EMBEDDING_DIMENSION",
+            DEFAULT_DIMENSION,
+        )
+
+        logger.info(
+            f"OpenRouter Embedder initialized with model: {self.model} "
+            f"(dimension={self.dimension})"
+        )
+
+
+class LocalEmbedder(_BaseOpenAICompatibleEmbedder):
+    """
+    Embedder for a local OpenAI-compatible embedding server — Ollama,
+    llama.cpp, vLLM, LM Studio, etc. Zero new Python dependencies; just
+    point the existing OpenAI SDK at a local base URL.
+
+    Environment variables:
+        EMBEDDING_PROVIDER=local              (selects this provider)
+        LOCAL_EMBEDDING_BASE_URL              (default: http://localhost:11434/v1)
+        LOCAL_EMBEDDING_MODEL                 (default: nomic-embed-text)
+        LOCAL_EMBEDDING_DIMENSION             (default: 768)
+        LOCAL_EMBEDDING_API_KEY               (optional; ignored by most local servers)
+
+    Setup (Ollama):
+        $ ollama serve
+        $ ollama pull nomic-embed-text          # or bge-m3 for better Turkish
+
+    The dimension MUST match the model's actual output size (e.g. 768 for
+    nomic-embed-text, 1024 for bge-m3, 1024 for mxbai-embed-large).
+    """
+
+    def __init__(
+        self,
+        base_url: Optional[str] = None,
+        model: Optional[str] = None,
+        dimension: Optional[int] = None,
+        api_key: Optional[str] = None,
+    ):
+        try:
+            from openai import OpenAI
+        except ImportError:
+            raise ImportError("openai package is required. Install with: pip install openai")
+
+        self.base_url = (
+            base_url
+            or os.getenv("LOCAL_EMBEDDING_BASE_URL")
+            or LOCAL_DEFAULT_BASE_URL
+        )
+        # Most local servers don't validate the key — use a placeholder so
+        # the OpenAI SDK doesn't error on the missing-key check.
+        effective_key = (
+            api_key
+            or os.getenv("LOCAL_EMBEDDING_API_KEY")
+            or "no-key-needed"
+        )
+
+        self.client = OpenAI(base_url=self.base_url, api_key=effective_key)
+        self.model = model or os.getenv("LOCAL_EMBEDDING_MODEL") or LOCAL_DEFAULT_MODEL
+        self.dimension = _coerce_dimension(
+            dimension if dimension is not None else os.getenv("LOCAL_EMBEDDING_DIMENSION"),
+            "LOCAL_EMBEDDING_DIMENSION",
+            LOCAL_DEFAULT_DIMENSION,
+        )
+
+        logger.info(
+            f"Local Embedder initialized: model={self.model} "
+            f"base_url={self.base_url} dimension={self.dimension}"
+        )
+
+
+def get_embedder():
+    """
+    Factory that picks the embedder based on EMBEDDING_PROVIDER.
+
+    - ``EMBEDDING_PROVIDER=local`` -> ``LocalEmbedder``
+    - otherwise -> ``OpenRouterEmbedder`` (requires OPENROUTER_API_KEY)
+
+    Raises:
+        ValueError: If no provider is configured (neither local nor OpenRouter).
+    """
+    if is_local_embedding_configured():
+        return LocalEmbedder()
+    if is_openrouter_available():
+        return OpenRouterEmbedder()
+    raise ValueError(
+        "No embedding provider configured. Set OPENROUTER_API_KEY for hosted "
+        "embeddings, or EMBEDDING_PROVIDER=local (with LOCAL_EMBEDDING_* "
+        "env vars) for a local OpenAI-compatible server like Ollama."
+    )