Measure whether the five-document café FAQ retriever works

A plausible result is not yet evidence. Evaluate with retrieval hit rate on a labeled question set, answer support rate, abstention accuracy, and citation correctness. The test fixture should contain an easy positive case, an easy negative or baseline case, and the boundary case most likely to flip. Separate assertions about software contracts from claims about model quality: both matter, but they answer different questions.

The artifact's user-facing goal is specific: answer a café question only when a relevant local note is retrieved and cite the note identifier used. Its accepted input is five short documents with stable IDs plus one normalized user question. Those statements are intentionally narrower than “build an AI system.” Narrow scope lets us inspect every input and expected result, and it prevents a toy result from being presented as a production claim. The runnable check below turns one success criterion into an assertion, so a regression exits loudly.

Save this as lesson.py and run python3 lesson.py. It uses only the language standard library, so the example is reproducible offline.

gold=[('When is the cafe open?','hours'),('How do I get wifi?','wifi')]
docs={'hours':'cafe open weekdays','wifi':'get wifi card'}
def hit(q): return max(docs,key=lambda k:len(set(q.lower().split())&set(docs[k].split())))
assert all(hit(q)==want for q,want in gold); print('hit_rate',1.0)

Expected output: hit_rate 1.0. Exact floating-point formatting may vary slightly, but the asserted behavior must not. Read the output as evidence about this stage, not merely proof that the interpreter started.

Print normalized question tokens, each document's overlap score, the winning ID, and the exact cited text. A fluent wrong answer often starts as a retrieval miss, so inspect retrieval before prompt wording. Zero overlap must abstain. For ties, choose and document a deterministic rule. Treat instructions embedded in a document as untrusted text, not commands for the application.

At the evaluation stage, save the smallest failing fixture beside the expected result. Change one cause at a time and rerun the exact command printed above; that makes the repair reviewable and keeps this chapter's progressive artifact reproducible.

Create labeled questions for each of the five documents plus unsupported questions. Measure retrieval hit rate, citation correctness, supported-answer rate, and correct abstention. Read every answer beside its cited note; lexical overlap can retrieve a document for the wrong reason. Version the notes so a score can be reproduced after content changes.

For this evaluation page, preserve the fixture and result as evidence for the next page. Label observations separately from conclusions: a passing assertion establishes the behavior it names, while broader usefulness requires the chapter's full evaluation set and stated operating limits.