[kchandan]

What this adds

Local RAG lab (labagent.ai.rag, extra rag-local)

• RAGLabClient: ingest files or pre-chunked rows, hybrid retrieval (vector + keyword rerank), metadata filters, citations, evaluate() with a pass mark.
• Stores: Chroma (default, embedded), Qdrant, in-memory. Local MiniLM embeddings.
• LLM gateway: Groq, OpenAI, Anthropic, Ollama; falls back to a clearly marked mock answer when no key works. HTTP status is reported in warnings.
• labagent rag init | ide | playground.
• Workspace notebooks: a starter, and an SEC 10-K tutorial built on the pinned dataset.

Web IDE image: Dockerfile.lab (code-server pinned by SHA-256, model pre-cached, non-root) and a Codespaces devcontainer.

rag-ide pack (extra modal): labagent launch rag-ide starts a private browser VS Code per learner as a Modal Sandbox with its own volume, random password in a 0600 CSV, idle shutdown and a hard time cap; destroy removes sandboxes and volumes. Modal calls sit behind a gateway interface so another platform (EC2) can be added.

Verified

• 222 tests pass (Python 3.9). Clean wheel install without extras: rag init, packs, launch rag-ide --dry-run work; a missing modal package gives an install hint, not a traceback.
• Real Modal run: image build, login, non-root code-server (no sudo), volume persistence across a restart, notebook run, destroy cleanup.
• Real Chroma, Qdrant (in-process) and MiniLM against the pinned SEC snapshot.

Not verified

• Written answers from a real LLM in the SEC notebook (only mock locally).
• 30 simultaneous learners on Modal; Modal billing (sandboxes bill at the non-preemptible rate).
• The Codespaces devcontainer has never been opened in Codespaces; remote Qdrant Cloud untested.
• image=pypi needs a release containing labagent.ai.rag (not yet cut).

Not included

Docs for these features (follow-up) and the throwaway Modal spike.

🤖 Generated with Claude Code