Local FAISS MCP Server

A Model Context Protocol (MCP) server that provides local vector database functionality using FAISS for Retrieval-Augmented Generation (RAG) applications.

Features

Core Capabilities

• Local Vector Storage: Uses FAISS for efficient similarity search without external dependencies

• Document Ingestion: Automatically chunks and embeds documents for storage

• Semantic Search: Query documents using natural language with sentence embeddings

• Persistent Storage: Indexes and metadata are saved to disk

• MCP Compatible: Works with any MCP-compatible AI agent or client

v0.2.0 Highlights

• CLI Tool: local-faiss command for standalone indexing and search

• Document Formats: Native PDF/TXT/MD support, DOCX/HTML/EPUB with pandoc

• Re-ranking: Two-stage retrieve and rerank for better results

• Custom Embeddings: Choose any Hugging Face embedding model

• MCP Prompts: Built-in prompts for answer extraction and summarization

Quickstart

Or use with Claude Code - configure MCP client (see Configuration) and try:

Claude will retrieve relevant document chunks from your vector store and use them to answer your question.

Installation

⚡️ Upgrading? Run pip install --upgrade local-faiss-mcp

From PyPI (Recommended)

Optional: Extended Format Support

For DOCX, HTML, EPUB, and 40+ additional formats, install pandoc:

Note: PDF, TXT, and MD work without pandoc.

From Source

Usage

Running the Server

After installation, you can run the server in three ways:

1. Using the installed command (easiest):

2. As a Python module:

3. For development/testing:

Command-line Arguments:

• --index-dir: Directory to store FAISS index and metadata files (default: current directory)

• --embed: Hugging Face embedding model name (default: all-MiniLM-L6-v2)

• --rerank: Enable re-ranking with specified cross-encoder model (default: BAAI/bge-reranker-base)

Using a Custom Embedding Model:

Using Re-ranking for Better Results:

Re-ranking uses a cross-encoder model to reorder FAISS results for improved relevance. This two-stage "retrieve and rerank" approach is common in production search systems.

How Re-ranking Works:

1. FAISS retrieves top candidates (10x more than requested)

2. Cross-encoder scores each candidate against the query

3. Results are re-sorted by relevance score

4. Top-k most relevant results are returned

Popular re-ranking models:

• BAAI/bge-reranker-base - Good balance (default)

• cross-encoder/ms-marco-MiniLM-L-6-v2 - Fast and efficient

• cross-encoder/ms-marco-TinyBERT-L-2-v2 - Very fast, smaller model

The server will:

• Create the index directory if it doesn't exist

• Load existing FAISS index from {index-dir}/faiss.index (or create a new one)

• Load document metadata from {index-dir}/metadata.json (or create new)

• Listen for MCP tool calls via stdin/stdout

Available Tools

The server provides two tools for document management:

1. ingest_document

Ingest a document into the vector store.

Parameters:

• document (required): Text content OR file path to ingest

• source (optional): Identifier for the document source (default: "unknown")

Auto-detection: If document looks like a file path, it will be automatically parsed.

Supported formats:

• Native: TXT, MD, PDF

• With pandoc: DOCX, ODT, HTML, RTF, EPUB, and 40+ formats

Examples:

2. query_rag_store

Query the vector store for relevant document chunks.

Parameters:

• query (required): The search query text

• top_k (optional): Number of results to return (default: 3)

Example:

Available Prompts

The server provides MCP prompts to help extract answers and summarize information from retrieved documents:

1. extract-answer

Extract the most relevant answer from retrieved document chunks with proper citations.

Arguments:

• query (required): The original user query or question

• chunks (required): Retrieved document chunks as JSON array with fields: text, source, distance

Use Case: After querying the RAG store, use this prompt to get a well-formatted answer that cites sources and explains relevance.

Example workflow in Claude:

1. Use query_rag_store tool to retrieve relevant chunks

2. Use extract-answer prompt with the query and results

3. Get a comprehensive answer with citations

2. summarize-documents

Create a focused summary from multiple document chunks.

Arguments:

• topic (required): The topic or theme to summarize

• chunks (required): Document chunks to summarize as JSON array

• max_length (optional): Maximum summary length in words (default: 200)

Use Case: Synthesize information from multiple retrieved documents into a concise summary.

Example Usage:

In Claude Code, after retrieving documents with query_rag_store, you can use the prompts like:

The prompts will guide the LLM to provide structured, citation-backed answers based on your vector store data.

Command-Line Interface

The local-faiss CLI provides standalone document indexing and search capabilities.

Index Command

Index documents from the command line:

Configuration: The CLI automatically uses MCP configuration from:

1. ./.mcp.json (local/project-specific)

2. ~/.claude/.mcp.json (Claude Code config)

3. ~/.mcp.json (fallback)

If no config exists, creates ./.mcp.json with default settings (./.vector_store).

Supported formats:

• Native: TXT, MD, PDF (always available)

• With pandoc: DOCX, ODT, HTML, RTF, EPUB, etc.

  • Install: brew install pandoc (macOS) or apt install pandoc (Linux)

Search Command

Search the indexed documents:

Results show:

• Source file path

• FAISS distance score

• Re-rank score (if enabled in MCP config)

• Text preview (first 300 characters)

CLI Features

• ✅ Incremental indexing: Adds to existing index, doesn't overwrite

• ✅ Progress output: Shows indexing progress for each file

• ✅ Shared config: Uses same settings as MCP server

• ✅ Auto-detection: Supports glob patterns and recursive folders

• ✅ Format support: Handles PDF, TXT, MD natively; DOCX+ with pandoc

Configuration with MCP Clients

Claude Code

Add this server to your Claude Code MCP configuration (.mcp.json):

User-wide configuration (~/.claude/.mcp.json):

With custom index directory:

With custom embedding model:

With re-ranking enabled:

Full configuration with embedding and re-ranking:

Project-specific configuration (./.mcp.json in your project):

Alternative: Using Python module (if the command isn't in PATH):

Claude Desktop

Add this server to your Claude Desktop configuration:

Architecture

• Embedding Model: Configurable via --embed flag (default: all-MiniLM-L6-v2 with 384 dimensions)

  • Supports any Hugging Face sentence-transformers model

  • Automatically detects embedding dimensions

  • Model choice persisted with the index

• Index Type: FAISS IndexFlatL2 for exact L2 distance search

• Chunking: Documents are split into ~500 word chunks with 50 word overlap

• Storage: Index saved as faiss.index, metadata saved as metadata.json

Choosing an Embedding Model

Different models offer different trade-offs:

Important: Once you create an index with a specific model, you must use the same model for subsequent runs. The server will detect dimension mismatches and warn you.

Development

Standalone Test

Test the FAISS vector store functionality without MCP infrastructure:

This test:

• Initializes the vector store

• Ingests sample documents

• Performs semantic search queries

• Tests persistence and reload

• Cleans up test files

Unit Tests

Run the complete test suite:

Run specific test files:

The test suite includes:

• test_embedding_models.py: Comprehensive tests for custom embedding models, dimension detection, and compatibility

• test_standalone.py: End-to-end integration test without MCP infrastructure

License

MIT