doc-indexer

doc-indexer is a local MCP server written in Rust that enables semantic search over a variety of document types using a Qdrant vector store and Voyage AI embeddings. It exposes MCP tools to index documents or directories, perform semantic searches across the indexed content, delete documents, and retrieve statistics about the index. The server integrates with Claude Code CLI and other MCP-compatible clients, allowing you to manage documents and queries through MCP tools like index_document, index_directory, search_documents, delete_document, and get_stats. To use it, configure your MCP client to point at this server, supply the necessary environment variables (authentication with Voyage, embedding model, and Qdrant access), and then issue the available MCP tool commands to index and search your documents.

Prerequisites:

• Rust 2024 Edition (rustc 1.85+) installed on your system
• Qdrant vector database up and running (server or docker) at the configured URL
• poppler-utils (pdftotext) available for PDF extraction
• Voyage AI API key (or an OpenAI-compatible embeddings endpoint)

Install steps:

1. Install Rust and set up the environment

   • On macOS with Homebrew: brew install rustup; rustup default stable
   • On Linux: follow rustup.rs installer from https://rust-lang.org

2. Install dependencies for PDF parsing (pdftotext via poppler)

   • macOS: brew install poppler
   • Linux: sudo apt-get install poppler-utils

3. Ensure Qdrant is running

   • Run locally: qdrant (or use the provided binary/docker image)
   • Verify accessible at http://localhost:6334 by default

4. Build the MCP server

   • git clone <repo-url> # if you haven’t already
   • cd doc-indexer-mcp
   • cargo build --release

5. Prepare environment variables

   • Create a .env file or export variables as needed (see example below)

6. Run the server

   • Run the server with the release binary (configured for your MCP client): cargo run --release

Note: The server reads configuration from environment variables as described in the Configuration section. You can also adapt the setup to run with your preferred environment (Docker, npx, etc.) if you port the config accordingly.

Tips and common issues:

• Ensure Qdrant is reachable at the configured QDRANT_URL and that the collection (QDRANT_COLLECTION) exists or can be created by the server.
• The embedding API key (VOYAGE_API_KEY) must be valid and have access to the Voyage AI embeddings endpoint; if using an OpenAI-compatible endpoint, confirm compatibility.
• PDF parsing relies on pdftotext; ensure it is installed and accessible in your PATH.
• If you change DOCS_PATH or INDEX_SUBDIRS, make sure filesystem permissions allow read access to the documents.
• For large document sets, adjust PDF_CHUNK_SIZE and OVERLAP to optimize indexing performance and memory usage.
• The MCP server uses stdio for MCP communication; ensure your MCP client is configured to communicate over the correct channel.
• If you deploy locally and want to test Claude Code integration, ensure the Claude Code CLI configuration includes the doc-indexer MCP server with the appropriate command path and environment variables.