feat: RAG pipeline for privacy-first document Q&A

[wydrox]

Summary

• Adds a complete RAG (Retrieval-Augmented Generation) pipeline for chatting with local documents -- the chore: add .gitignore entries for dev artifacts #1 use case for local LLMs
• New ppmlx/rag.py module with document loader (30+ file types), recursive text chunker, SQLite-backed vector store, and RAG chain
• Four new CLI commands: ppmlx rag ingest <path>, ppmlx rag chat, ppmlx rag list, ppmlx rag rm <collection>
• Privacy-first: all embeddings and generation happen locally, no data leaves the machine
• 37 new tests covering all components with mocked embeddings (no GPU required)

Key Design Decisions

• Uses SQLite for vector storage (consistent with existing db.py patterns), with cosine similarity computed in pure Python
• Supports collections so users can organize different document sets
• Change detection via SHA-256 hashing skips unchanged files on re-ingestion
• Embedding model is configurable per collection (defaults to embed:all-minilm)
• No new dependencies beyond what ppmlx already uses

CLI Usage

# Ingest documents
ppmlx rag ingest ./my-docs --collection myproject

# Chat with your documents
ppmlx rag chat --collection myproject --model llama3

# List collections
ppmlx rag list

# Remove a collection
ppmlx rag rm myproject

Test plan

• 37 unit tests pass: chunking, vector encoding, cosine similarity, document loading, file discovery, vector store CRUD, ingest pipeline, retrieval, prompt building, CLI integration
• Full test suite (212 tests) passes with no regressions
• Code review and cleanup completed (removed unused import, deduplicated test helper, fixed cursor-level row_factory)

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.