Repo-Aware Code RAG Assistant

Note: This repository is the public demonstration version of a private, live system under continuous development. It implements the same architecture end-to-end — AST-aware chunking, hybrid retrieval, the FastAPI service, the MCP server, and the evaluation harness — in a fully local, zero-cloud-account mode. The production deployment (Azure AI Search, Databricks ingestion jobs, MLflow tracking) is private; its integration seams are visible here as thin, documented adapters.

A retrieval-augmented assistant for codebases that answers "where is X and how does it work?" with exact file:line citations — built on the observation that code retrieval fails differently than prose retrieval, and needs code-specific treatment at every stage.

                       ┌──────────────────────────────────────────────┐
   repo on disk ──────▶│  AST-aware chunker (chunking.py)             │
                       │  functions/classes with exact line spans;    │
                       │  windows only as fallback                    │
                       └──────────────┬───────────────────────────────┘
                                      ▼
                       ┌──────────────────────────────────────────────┐
                       │  Index (index.py)                            │
                       │  BM25 w/ identifier field-boost  +  dense    │
                       │  local: numpy+rank_bm25 · prod: Azure AI Search │
                       └──────────────┬───────────────────────────────┘
                                      ▼
                       ┌──────────────────────────────────────────────┐
                       │  Retrieval (retrieval.py)                    │
                       │  query expansion → RRF fusion → rank priors  │
                       │  → optional cross-encoder rerank             │
                       └──────┬───────────────────────┬───────────────┘
                              ▼                       ▼
                   FastAPI service (api.py)   MCP server (mcp_server.py)
                   /ingest /search /ask       search_code · read_span
                                              — retrieval as a live tool

Quickstart (fully local, no keys)

pip install -e .
python -m repo_rag.ingest .        # index this repo on itself
repo-rag-api                       # http://localhost:8000/docs

Real output of the retriever, dogfooding on this repository:

$ search: 'where is reciprocal rank fusion implemented'
  src/repo_rag/retrieval.py:64-73  (function reciprocal_rank_fusion, via bm25+dense)
  src/repo_rag/eval.py:45-68  (function evaluate, via bm25+dense)
  src/repo_rag/retrieval.py:109-125  (function Retriever._prior, via bm25+dense)

$ search: 'which MCP tool reads exact source lines for a citation'
  src/repo_rag/mcp_server.py:72-82  (function read_span, via bm25+dense)
  src/repo_rag/mcp_server.py:1-16  (module mcp_server docstring, via bm25+dense)
  src/repo_rag/chunking.py:83-98  (function chunk_repo, via bm25+dense)

Ask over HTTP:

curl -s localhost:8000/search -X POST -H 'content-type: application/json' \
  -d '{"query": "how are python files split into chunks", "k": 3}' | jq .

/ask returns a synthesized answer: extractive with citations by default, or LLM-written (still citation-constrained) when ANTHROPIC_API_KEY is set.

Related MCP server: CodeGraph

MCP: retrieval as a live tool, not a hardcoded pipeline

Instead of one fixed ask→retrieve→generate chain, the index is exposed over the Model Context Protocol, so any MCP client — Claude Desktop, Claude Code, IDE agents — decides when and how to query it, iteratively, mid-conversation.

pip install ".[mcp]"

Claude Desktop config (claude_desktop_config.json):

{
  "mcpServers": {
    "repo-rag": {
      "command": "repo-rag-mcp",
      "env": { "REPO_RAG_INDEX_DIR": "/absolute/path/to/.ragindex" }
    }
  }
}

Tools exposed: search_code(query, k), read_span(path, start, end), index_stats().

Evaluation

The harness measures recall@k and MRR against gold file:line spans (hit = line-overlap in the same file), and logs to MLflow when MLFLOW_TRACKING_URI is set. A RAGAS faithfulness pass over generated answers is available behind --ragas (needs the [eval] extra and an LLM key).

repo-rag-eval eval/sample_eval.jsonl --k 5
# recall_at_5: 0.8333
# mrr: 0.6167
# n_questions: 6

(The one miss is instructive: this README quotes the demo queries, so once indexed it outcompetes the implementation for one question - the same docs-vs-code tension the rank priors exist to manage.)

Those numbers are the illustrative sample set in this repo (6 self-referential questions, offline hash embedder, no rerank). The headline metrics for this project — recall@5 lifted 61% → 84% (0.79 MRR) with RAGAS faithfulness 0.82 — were measured on the full evaluation set of the private live deployment over a ~2K-file codebase, using the same harness with production embeddings and reranking enabled. The methodology here is the methodology there; run it on your own repo with your own eval set.

Ablations are one flag away: --no-expand disables query expansion; REPO_RAG_RERANK=true enables the cross-encoder (with the [ml] extra).

Design notes (the interesting 20%)

• AST chunking over fixed windows (chunking.py) — windows slice through function bodies, poisoning embeddings and producing citations that start mid-def. AST boundaries keep units intact, give every chunk a name, and make path:start-end land exactly where a developer would open the file. Oversized definitions window internally, carrying the parent's name.

• RRF over weighted score sums (retrieval.py) — BM25 scores and cosine similarities live on incomparable scales; rank fusion is scale-free and needs no corpus-specific tuning.

• Identifier field-boosting in BM25 (index.py) — in code search, a chunk whose name matches the query nearly always beats prose that merely mentions the concept. Implemented as token repetition inside BM25 - together with the rank priors, the single biggest sample-eval lift.

• Weak rank priors — implementations over tests, definitions over module prose. Deliberately weak multipliers, so strong matches in tests still surface.

• Hermetic by default — the deterministic hash embedder keeps tests, CI, and the demo fully offline; sbert/azure backends are a config switch.

Production architecture (private deployment)

• Azure AI Search serves hybrid BM25 + vector ranking over the same chunk schema (index.AzureAISearchIndex is the adapter seam).

• Databricks runs repo_rag.ingest as a scheduled job against repository checkouts, feeding the indexer.

• MLflow tracks every eval run across chunking/retrieval configurations.

• Docker → Azure: docker build -t repo-rag . && az containerapp up --name repo-rag --source .

Roadmap

• AST-aware chunking with exact line spans

• Hybrid BM25 + dense retrieval with RRF

• Query expansion; optional cross-encoder rerank

• FastAPI service with citation-formatted responses

• MCP server (stdio): search_code, read_span, index_stats

• Eval harness: recall@k, MRR, MLflow logging, RAGAS hook

• Incremental re-indexing on file change (watch mode)

• Tree-sitter chunking for TypeScript/Go/Java (beyond line windows)

• MCP resources: expose indexed files as browsable resources

• Eval dashboard comparing configs across runs

License

MIT © 2026 Manav Kanaganapalli