local_faiss_mcp

The Local FAISS MCP Server provides a self-contained vector store powered by FAISS for retrieval-augmented generation workflows. It ingests documents, builds a local vector index, and serves MCP-compatible tool calls for ingestion and querying. You can use the ingest_document tool to add content to the vector store and the query_rag_store tool to retrieve relevant document chunks based on a natural language query. Optional built-in prompts like extract-answer and summarize-documents help you generate structured, citation-backed responses from the retrieved chunks. The server exposes three main usage modes: running the installed CLI, running as a Python module, or running the server directly for development/testing. When running, you specify an index directory where FAISS and metadata are stored, and you can customize the embedding model and re-ranking options via command-line arguments.

To interact with the MCP server, use:

• ingest_document to add new content to the index (e.g., local and remote documents).
• query_rag_store to search the vector store for relevant chunks.
• Then apply prompts like extract-answer to generate a formatted answer with citations, or summarize-documents to create focused summaries from multiple chunks.

Prerequisites:

• Python 3.10+ installed on your system
• Pip available for installing Python packages

Install the MCP server package from PyPI:

pip install local-faiss-mcp

Usage variants after installation:

1. Run the server using the installed CLI (easiest):

local-faiss-mcp --index-dir /path/to/index/directory

2. Run the server as a Python module (recommended for MCP workflows):

python -m local_faiss_mcp --index-dir /path/to/index/directory

3. For development or testing, run the server module directly:

python local_faiss_mcp/server.py --index-dir /path/to/index/directory

Optional to enable broader file format support, install pandoc if you plan to ingest DOCX, HTML, EPUB, and more formats:

# macOS
brew install pandoc

# Linux
sudo apt install pandoc

Note: PDF, TXT, and MD ingestion works without pandoc. To upgrade later:

pip install --upgrade local-faiss-mcp

Tips and common considerations:

• The index directory is created automatically if it does not exist. Ensure the path has read/write permissions.
• By default, embeddings use the configured Hugging Face model; you can pass --embed to customize the model (e.g., all-MiniLM-L6-v2, all-mpnet-base-v2).
• Re-ranking can improve results. Use --rerank to enable the default cross-encoder-based reranker, or specify a model with --rerank <model-name>.
• The server listens for MCP tool calls via stdin/stdout. When integrating with an MCP client, ensure proper I/O handling in your environment.
• If you encounter memory issues with large corpora, consider reducing top_k during query_rag_store or using smaller embedding models for resource-constrained setups.