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 non-sensitive documents you gathered last night (PDF/txt). Use a short, text-based document or one chapter. Forgot? Use the instructor’s sample document sample-os-notes.txt (OS lecture notes + an illustrative page of academic regulations — Cell 2 falls back to it automatically, and Cells 5–7 are pre-filled with queries that work on it; see sample-queries.md). Do not upload private marks, IDs, passwords, or confidential college records.
Load your document → chunk (paragraph + overlap) → embed → numpy vector store. Read the Cell 2 sanity print — if extraction is empty, suspiciously short, or visibly garbled, stop and switch to the instructor sample now (FILENAME = SAMPLE), not at minute 40. Cell 1 mounts your Google Drive — accept the prompt: Cell 4 saves your vector store there so Labs 5 and 6 can reload it after the break. The notebook indexes only the configured chunk limit; a score applies only to the content actually indexed.
✓ Checkpoint 1: chunk count + vector-store shape printed.
Three test queries about your material (the pre-filled ones are for the sample document — replace them). 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 (this becomes 25% of your capstone score — aim for 10 examples, not 5) · 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 — and note that File → Save keeps your code, not your uploaded PDF or embeddings: Cell 4 writes chunk_vecs.npy + chunks.json to your Drive (MyDrive/genai), and Labs 5 and 6 have a reload cell that reads them back instead of re-embedding. 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.