Different embedding model families need different prompt prefixes —
Gemini wants "task: ... | query: ..." and "title: ... | text: ...",
e5 wants "query: ..." / "passage: ...", and using the wrong one
silently degrades retrieval quality. Add EMBEDDING_PROMPT_STYLE
(gemini/e5/raw) so the prefix matches the chosen model.
Defaults: gemini for OpenRouter (matches the existing default
google/gemini-embedding-001), e5 for the local provider (matches
the recommended multilingual-e5-large setup). Both override via
env var or constructor.
Update README and .env.example to recommend intfloat/multilingual-
e5-large served by HuggingFace Text Embeddings Inference (one
docker run) as the Turkish-optimized local setup, with a clear env
var reference table. Ollama and OpenRouter remain documented as
alternatives.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
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>
google/gemini-embedding-001 became paid on OpenRouter, leaving
users without credit unable to run the semantic_search tool. The
old code hardcoded the model and 3072 dimensions in three places.
Make OpenRouterEmbedder accept model/dimension via constructor
args or OPENROUTER_EMBEDDING_MODEL / OPENROUTER_EMBEDDING_DIMENSION
env vars, with the previous values as backward-compatible defaults.
Switch the VectorStore and the response payload in mcp_server_main
to read embedder.dimension instead of the hardcoded 3072 so a
configured non-Gemini model does not produce shape mismatches.
Bad dimension input (non-int or non-positive) now raises a clear
ValueError instead of a downstream shape error.
Documented the new env vars in .env.example.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Replace EmbeddingGemma local model with OpenRouter API integration
- Use google/gemini-embedding-001 model via OpenRouter (3072 dimensions)
- Add conditional tool registration: auto-disable if OPENROUTER_API_KEY not set
- Add openai and numpy dependencies to pyproject.toml
- Update .env.example with OPENROUTER_API_KEY configuration
- Fix ruff lint issues in semantic_search module
