Claude
1d4625a1c2
Co-Authored-By: Claude Sonnet 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01B38FGKKam5SCGgwgduVAh3
29 lines
2.1 KiB
Markdown
29 lines
2.1 KiB
Markdown
# 5. Generation and the faithfulness check
|
|
|
|
prev: [4. Retrieval and reranking](04-retrieval-and-reranking.md) · [Index](../README.md) · next: [6. Observability and a production checklist](06-observability-and-production-checklist.md)
|
|
|
|
Source: [`RagQueryService`](../src/main/java/com/ankurm/rag/query/RagQueryService.java). Test:
|
|
[`FaithfulnessTest`](../src/test/java/com/ankurm/rag/FaithfulnessTest.java) ([output 09](output/09-faithfulness-check.txt)).
|
|
|
|
After the model answers, `RagQueryService` asks a second model call, `FactCheckingEvaluator`, whether the answer is supported by the
|
|
chunks that were in the prompt. The response reports one of three statuses:
|
|
|
|
| status | when | fact check runs? |
|
|
|---|---|---|
|
|
| `answered` | chunks retrieved and the judge said "yes" | yes |
|
|
| `ungrounded` | chunks retrieved and the judge said anything other than "yes" (case aside), or the judge call failed | yes |
|
|
| `no_context` | nothing retrieved | no |
|
|
|
|
Behaviours the test pins down:
|
|
|
|
- **The verdict is matched literally.** A judge reply of `Yes.` counted as *not* grounded; `YES` counted as grounded. The evaluator
|
|
compares against "yes" ignoring case, not ignoring punctuation. A real judge model that likes to add a full stop would fail
|
|
every answer; instruct it to answer with one word, and watch `rag_queries_total{status="ungrounded"}`.
|
|
- **A failing judge is not a passing judge.** If the judge call throws (a rate limit, a timeout) the answer is reported `ungrounded` and
|
|
`rag.faithfulness.judge_failures` increments. The answer text is still returned, so the caller decides what to do with an unchecked answer.
|
|
- **No chunks, no check.** With nothing to check against, the service skips the judge (zero fact-check calls in the output) and reports `no_context`.
|
|
- **The check is only as good as the judge.** The tests script the judge, so they prove the wiring and the four outcomes, not that a real model
|
|
spots an invented claim. Measure that on your own questions before you rely on it, and note that it doubles the model calls per question.
|
|
|
|
Next: [6. Observability and a production checklist](06-observability-and-production-checklist.md)
|