Skip to main content
Glama
Sign In
rag-mcp

rag-mcp

by Kamalesh-Kavin
OverviewSchemaSourceRelated ServersDiscussionsAdmin
Python
Local

rag-mcp

A personal knowledge base MCP server for Claude Desktop.

Drop in files (PDF, Markdown, plain text, code), and ask Claude questions that span your entire document collection. Powered by local embeddings via Ollama and ChromaDB for persistent vector storage.

Features

  • Index any document — PDF, .md, .txt, .py, .js, .ts, .json, .yaml

  • Semantic search — finds relevant content by meaning, not just keywords

  • Local & private — all embeddings generated locally via Ollama (no data leaves your machine)

  • Persistent — ChromaDB persists to disk; re-index only when documents change

  • Re-index safe — indexing the same file twice replaces old chunks cleanly

Related MCP server: Mimir

Tools exposed to Claude

Tool

Description

index_document

Index a file into the knowledge base

search_docs

Semantic search across all indexed documents

list_indexed_docs

List every document currently in the index

delete_document

Remove a document and all its chunks

Resource: doc://{filename} — read all raw chunks for a specific document

Requirements

  • Python 3.13+

  • uv

  • Ollama running locally with nomic-embed-text pulled:

    ollama pull nomic-embed-text

Setup

git clone https://github.com/Kamalesh-Kavin/rag-mcp
cd rag-mcp
cp .env.example .env
uv sync

Claude Desktop configuration

Add this to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "rag-assistant": {
      "command": "/path/to/uv",
      "args": [
        "--directory",
        "/path/to/rag-mcp",
        "run",
        "rag-mcp"
      ]
    }
  }
}

Usage in Claude

Index a document:
  "Index the file /Users/me/notes/architecture.md"

Ask a question:
  "What does my architecture doc say about the database layer?"

List what's indexed:
  "What documents are in my knowledge base?"

Delete a document:
  "Remove architecture.md from the knowledge base"

Architecture

File on disk
    │
    ▼
read_file()          ← pypdf (PDF) or open() (text/code)
    │
    ▼
chunk_pages()        ← sliding window: 1000 chars, 200 overlap
    │
    ▼
embed_chunks()       ← POST http://localhost:11434/api/embeddings
    │                   nomic-embed-text → 768-dim vector
    ▼
VectorStore.add()    ← ChromaDB PersistentClient, cosine similarity
    │
    ▼
search_docs()        ← embed query → cosine nearest-neighbour lookup

Project structure

src/rag_mcp/
├── __init__.py
├── ollama_client.py   # async httpx wrapper for Ollama embeddings API
├── ingestion.py       # file readers, chunker, ingest pipeline
├── vector_store.py    # ChromaDB wrapper (add, search, list, delete)
└── server.py          # MCP server — 4 tools + 1 resource
data/documents/        # drop files here to index them

License

MIT

Available Tools

4 tools
delete_documentA

Remove a document and all its chunks from the knowledge base.

This is useful when a document has been updated (you'll delete then re-index it) or when you no longer need it to be searchable.

Args: source: The filename of the document to remove, e.g. "design_doc.pdf". Use list_indexed_docs to see valid filenames.

Returns: Confirmation message with the number of chunks deleted.

Inputs: source (required). Returns: result

index_documentA

Index a document into the knowledge base so it can be searched.

Reads the file at file_path, splits it into overlapping chunks, embeds each chunk using nomic-embed-text via Ollama, and stores the vectors in ChromaDB. Re-indexing the same file is safe — old chunks are replaced automatically.

Args: file_path: Absolute path to the file to index. Supported types: .pdf .md .txt .py .js .ts .json .yaml .yml

Returns: A summary string describing what was indexed.

Inputs: file_path (required). Returns: result

list_indexed_docsA

List all documents currently indexed in the knowledge base.

Returns one entry per document with its filename, file type, number of chunks, and when it was last indexed.

Returns: Formatted string listing all documents, or a message if empty.

Inputs: no named parameters. Returns: result

search_docsA

Semantically search the knowledge base for content related to query.

The query is embedded using the same model used during indexing. ChromaDB finds the N most similar chunks using cosine similarity. Lower distance = more relevant.

Args: query: Natural-language question or keyword phrase. n_results: How many results to return (default 5, max 20). source_filter: If non-empty, restrict search to chunks from this specific document (exact filename, e.g. "notes.md").

Returns: Formatted string with matching chunks, their sources, and distances. Returns a helpful message if the knowledge base is empty.

Inputs: query (required), n_results, source_filter. Returns: result

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Deploy Server
    One endpoint for every MCP toolRayrun runs every MCP server behind one URL and holds the credentials. Each call is decided per client, tool, and argument, then recorded.ray.runAd

    TDQS

    A4.5/5.0

    Scored across 4 tools

    Disambiguation5/5

    Each tool has a clearly distinct purpose: index, delete, list, and search. No overlap or ambiguity in their functionality.

    Naming Consistency5/5

    All tools follow a consistent verb_noun pattern (delete_document, index_document, search_docs, list_indexed_docs). The slight variation in list_indexed_docs is still predictable.

    Tool Count5/5

    4 tools is well-scoped for a RAG knowledge base server. It covers the essential operations without unnecessary bloat or missing core functionality.

    Completeness5/5

    The tool surface covers the complete lifecycle: indexing (create), listing (read), searching (query), and deleting. Update is handled via delete+re-index, which is acceptable for this domain.

    Maintenance

    ActivityInactive
    ResponsivenessOutdated data

    Resources

    Related MCP Connectors

    Related MCP Servers

    • Local RAG

      shinpr
      A
      license
      A
      quality
      A
      maintenance
      Privacy-first local document search using semantic search. Runs entirely on your machine with no cloud services, supporting PDF, DOCX, TXT, and Markdown files.
      9
      4,501 npm
      429
      MIT

    • Mimir

      jcode-works
      A
      license
      Not graded
      quality
      B
      maintenance
      Local-first memory and retrieval for private project knowledge. Enables indexing files, searching, and asking questions about project documents using local embeddings and LLM.
      6
      AGPL 3.0

    • mcp-rag

      elhassane1230
      A
      license
      A
      quality
      D
      maintenance
      Enables semantic search and question answering over a knowledge base using hybrid retrieval and grounded answers, all running offline with no API keys.
      4
      MIT