[kchandan]

What

The local RAG lab was a tour of a finished library (RAGLabClient). Learners now build each part themselves, one notebook at a time, with a self-check after every exercise.

#	Notebook	Learner writes
1	chunking	fixed and paragraph-aware chunkers
2	embeddings	cosine similarity, ranking
3	vector store	numpy store (add, search)
4	generation	grounded prompt, full pipeline
5	Chroma and filters	ChromaStore, metadata filters
6	hybrid search	keyword score, hybrid ranking
7	evaluation	hit rate and MRR scorer, settings grid, own cases
8	break it	newest-version filter, refusal rule, injection detection, safer prompt

09_sec_filings_rag is the old SEC notebook, unchanged and optional.

• One source file per notebook holds both the stub and the solution (labagent/ai/rag/lab/), so learner and instructor versions cannot drift.
• labagent rag init creates the lab folder (README, nine notebooks, eight sample docs). --solutions adds solutions/ for instructors.
• Five new sample documents (eight in total).
• Docs: docs/rag-local.md rewritten, new docs/rag-lab-instructor.md, README line updated.
• Setup note: the pip bundled with macOS Python 3.9 cannot do pip install -e here, so the docs say to upgrade pip first.

Tested

• All 8 solution notebooks run through a real Jupyter kernel with every check passing.
• Tests execute every solution notebook and assert that every student notebook stops at a failed check, not a crash. Execution tests skip when sentence-transformers or chromadb is missing (so CI, which installs only the dev extra, runs the structural tests only).
• Full suite: 473 passed, 1 skipped (before the last two small commits; the RAG and hygiene tests were rerun after them).

Not tested

• Real LLM answers (Groq, Ollama): only the mock provider was exercised.
• Timings in the instructor notes are estimates, not measured with learners.
• Windows and Python 3.11 or newer.
• The abstain threshold in notebook 8 is tuned to this model and corpus, with a thin margin.

🤖 Generated with Claude Code