Rag Starter

What is Rag Starter?

Rag Starter is a production-ready starter that turns a folder of documents (PDFs, Markdown, text) into a cited Q&A service. Every answer is grounded in the source, with each claim traceable to the exact passage. It exposes the same core functionality both as an MCP server (for Claude Desktop, Claude Code, or any MCP host) and as an HTTP API (FastAPI).

How to use Rag Starter?

Install dependencies with pip install -r requirements.txt, then either run the MCP server by adding it to claude_desktop_config.json or start the HTTP API with the rag-starter-api command. Use the provided tools (rag_ingest, rag_search, rag_answer) to index documents and ask questions.

Key features of Rag Starter

• Keyless by default – embeddings run locally with ONNX MiniLM; no API key needed.
• Citations, not hallucinations – every answer includes source references like [file.pdf p3].
• Optional answer synthesis – with an ANTHROPIC_API_KEY the server writes a cited answer; without one it returns passages for the host LLM.
• Idempotent ingestion – re-ingesting a file updates it in place without duplicates.
• Two interfaces from one codebase – MCP server and HTTP API.

Use cases of Rag Starter

• Chat with internal documentation in Claude Desktop by ingesting a folder of PDFs and Markdown files.
• Build a custom Q&A endpoint for a web app using the FastAPI HTTP API.
• Prototype a RAG pipeline for a client project with zero API costs for retrieval.
• Reskin the app for different clients by swapping the data/ folder and updating configuration.

FAQ from Rag Starter

How is this different from other RAG projects?

It is keyless by default (local embeddings), produces cited answers to prevent hallucinations, and exposes both an MCP server and an HTTP API from the same codebase. It also uses idempotent ingestion so there are no duplicates on re-index.

What dependencies do I need?

Python 3 with pip install -r requirements.txt. The project uses Chroma for vector storage and ONNX MiniLM for local embeddings. No external services are required for basic retrieval.

Where is my data stored?

Document chunks are stored in a local persistent Chroma database. All data stays on your machine unless you configure an external embedding backend like OpenAI.

Can I use an API key for better answers?

Yes. By setting an ANTHROPIC_API_KEY, the server will synthesize a cited answer for each query. Without the key, it returns ranked passages for the host LLM to answer.

How do I switch from local embeddings to OpenAI?

Change the embedding_backend setting in config.py (or via environment variables) from default to openai. The rest of the pipeline remains the same.