Session 4 · Embeddings + RAG · TCE · 50 min · pairs · The weekend's main build.
Runs entirely in your browser (Google Colab). Any laptop or lab PC works — nothing to install, no GPU, no payment. See
ZERO-SETUP.
You need: the 2–3 documents you gathered last night (PDF/txt). Forgot? Grab any lecture PDF from your class group right now — 2 minutes.
Load your document → chunk (paragraph + overlap) → embed → numpy vector store. Read the Cell 2 sanity print — if the PDF extracted garbage, swap documents now, not at minute 40.
✓ Checkpoint 1: chunk count + vector-store shape printed.
Three test queries about your material. Eyeball the top chunks: right topic? complete thoughts? If not, tune chunk target size and re-run Cells 3–4.
✓ Checkpoint 2: three sane searches shown.
Grounded template (ONLY the context · cite chunks · "I don't know" escape hatch) → ask 5 real questions about your material. Use show_chunks=True — watching retrieval is how you debug RAG.
✓ Checkpoint 3: one honest failure + your fix (or why it's genuinely hard).
RAG eval — 5 Q + expected through your S2 harness · k=1 vs 3 vs 5 · add your second document · print similarity scores in answers.
Going deeper (S3, S4 in the notebook). Rerank — retrieve 20, let a model re-order them, keep 4: the biggest RAG upgrade after chunking, in ten lines. Two scores, never one — label which chunk should win, then measure recall@k (your ceiling on accuracy) and MRR (did it arrive near the top) separately from answer quality. High recall with low MRR is the exact signature that says add the reranker. Both are unpacked in the deck's <|deeper|> panels — press D.
Save this notebook. It IS your capstone foundation: S5 (after lunch) bolts tools onto it; S6 attacks and hardens it, then you demo. Choose documents you actually care about — the demo is only as interesting as the data.