Week 07: Exercises & Checklist

Graded exercises

Easy: Re-run 01-rag-policy-bot.ipynb with k=3 and k=10; report how recall changes and write one sentence on what recall@k rewards.
Standard: Add two questions to the golden eval set (make one span two policy documents); re-measure recall and fix any miss you can.
Stretch: In 02-knowledge-graph-lab.ipynb, add a via constraint: cheapest route through a chosen port, returning the full path with total cost.
Portfolio: Package the RAG bot as projects/policy_bot.py (CLI: question in → cited answer out), with the golden eval set and a saved recall number.

Easy: Re-run recall_at_k(vector_search, k=3) and k=10; recall@k is monotone non-decreasing in k, so explain why it rewards larger k even when precision falls.
Standard: Add two (question, [doc_ids]) tuples to QUERIES; make one question answerable only by combining POL-001 and POL-002. Re-run both recall functions and inspect which arm (vector vs. keyword) recovered each.
Stretch: Filter candidate SERVES edges so the path passes through a chosen port node; you'll need to traverse carrier → lane → port → lane → port and sum the two leg costs, not just one edge.
Portfolio: Wrap the chunk→embed→index→hybrid→generate chain in a function that takes a question string and returns (answer, [citations]); save recall_hybrid to a JSON file and print it on --eval.

[ ] Mon: Study retrieval fundamentals (knowledge-base 05).
[ ] Tue: Chunk and embed the policy corpus; build a vector index; measure baseline retrieval accuracy.
[ ] Wed: Add hybrid search and reranking; cite sources in answers; measure groundedness.
[ ] Thu: Build the carrier-lane-port graph in Neo4j; query shortest and cheapest routes.
[ ] Fri: Use case: the RAG bot answers 10 unseen questions with citations; log retrieval misses.
[ ] Sat: Take the Week 7 quiz (quiz.md), pass with 8/10; record the score in Notes.
[ ] Milestone: Update the Excel tracker; commit RAG and graph labs.