rag-index

rag-index

Purpose

Prototype utilities for building a repo index used by retrieval-augmented prompts. The workflow is scan -> index -> retrieve, with a one-shot pipeline for ad-hoc use.

Scripts

• scanner.py — recursive directory scanner that emits deterministic JSON manifests
• indexer.py — chunk-aware SQLite index builder and lexical search (build, search)
• retrieve.py — prompt-context formatter that consumes chunk-level search results (retrieve, pipeline)

How to use

1. Generate a manifest (scanner):
   • python3 scanner.py <path1> <path2> --output manifest.json
   • Common options:
     • --file-types .py .md
     • --max-depth 2
     • --exclude "*.venv*" "*.git*"
     • --stats to print exclusion counts
2. Build/update an index (indexer):
   • python3 indexer.py build --manifest manifest.json --output index.db
   • Incremental behavior:
     • unchanged files are skipped
     • changed files update only changed chunks
     • removed files/chunks are deleted
3. Search the index (indexer):
   • python3 indexer.py search "query" --index index.db --top-k 5
   • Returns chunk-level records (path, rel_path, start_line, end_line, chunk_id, snippet)
4. Retrieve prompt-ready context (retrieve):
   • python3 retrieve.py "query" --index index.db --top-k 5
   • Useful options:
     • --max-context-chars 8000 hard output budget
     • --max-per-file 3 diversity cap
     • --mode lex|sem|hybrid (sem uses TF-IDF cosine; hybrid blends lexical + semantic scores)
5. One-shot pipeline (retrieve):
   • python3 retrieve.py pipeline "query" --dirs <path1> <path2>
   • Optional:
     • --index /tmp/rag.db to persist index
     • --max-depth 2
     • --file-types .py .md
   • This runs scan -> build -> retrieve in one command.

When to use RAG

Use RAG when:

• the question depends on repository-specific behavior or contracts
• the answer needs exact symbols, paths, config keys, or line ranges
• the agent is uncertain and needs grounded evidence from code/docs

Do not use RAG when:

• the question is purely generic knowledge
• the user already provided the exact snippet needed
• the target file is already fully in active context and retrieval adds no value