retrieval-only-RAG

A PDF retrieval tool wrapped as an MCP server. It handles the R in RAG — your IDE agent (Cursor, Kiro, Claude Code) handles generation.

PDFs ──> [ load → chunk → embed → store → retrieve ]
                          │
                 returns matching chunks
                          │
              [ MCP server wraps the retriever ]
                          │
     IDE agent calls it ──┘  →  IDE agent writes the answer

No LLM inside this tool. Embeddings run locally (no cloud key needed).

Setup

python -m venv .venv
.venv\Scripts\activate        # Windows
pip install -r requirements.txt

Usage

Index your PDFs — drop PDF files into pdfs/ then run:

python -m pdf_rag.cli index

Only new or changed PDFs are processed on subsequent runs — unchanged files are skipped. Deleted PDFs have their chunks removed automatically.

Search — retrieve the top-k chunks for a question:

python -m pdf_rag.cli search "What is the difference between ArrayList and LinkedList?"

Output includes source filename, page number, and similarity score for each chunk.

MCP Server

Exposes one tool — search_pdfs(query) — that any MCP-compatible IDE agent can call.

python mcp_server.py

Claude Code (.mcp.json in project root)

A .mcp.json is already included in this repo:

{
  "mcpServers": {
    "pdf-rag": {
      "command": "C:\\Projects\\Retrieval\\.venv\\Scripts\\python.exe",
      "args": ["C:\\Projects\\Retrieval\\mcp_server.py"],
      "cwd": "C:\\Projects\\Retrieval"
    }
  }
}

Update the paths to match your machine, then Claude Code picks it up automatically.

Cursor (.cursor/mcp.json)

{
  "mcpServers": {
    "pdf-rag": {
      "command": "path/to/.venv/Scripts/python.exe",
      "args": ["path/to/mcp_server.py"],
      "cwd": "path/to/project"
    }
  }
}

Once connected, ask your IDE agent a question about your PDFs — it calls search_pdfs, gets the chunks, and writes the answer. You own retrieval; the agent owns generation.

Configuration (config.yaml)

pdf_folder: pdfs              # folder to scan for PDFs
vector_store: vector_store    # where ChromaDB persists the index
embedding_model: BAAI/bge-small-en-v1.5   # local HuggingFace model
top_k: 5                      # chunks returned per query

Project structure

pdf_rag/
  config.py      # load + validate config.yaml
  indexer.py     # PDF loading, chunking, embedding, ChromaDB persistence
  retriever.py   # similarity search + result formatting
  cli.py         # index / search commands
mcp_server.py    # MCP wrapper exposing search_pdfs()
config.yaml
requirements.txt
.mcp.json        # Claude Code MCP config (update paths for your machine)
pdfs/            # drop your PDFs here (not committed)
vector_store/    # ChromaDB index + manifest.json (not committed)

How the RAG split works

The MCP server is editor-agnostic — swap Cursor for Kiro (or any MCP client) by changing only the connection config, no code changes needed.