In analysis, refuse cleanly when evidence is missing

This commit is contained in:
Yiorgis Gozadinos 2026-06-03 10:34:40 +03:00
parent c83bd2d4f8
commit 33bbf543c5
No known key found for this signature in database

View file

@ -112,6 +112,14 @@ Search results include `doc_item_refs` (e.g. `["#/texts/48", "#/tables/0"]`) tha
You MUST call `cite` with at least one chunk ID before producing your final answer **when your answer is grounded on retrieved evidence**. Skip `cite` in two cases: (a) you are refusing for lack of information, or (b) your answer is a corpus-level computation (count, aggregation, listing) that doesn't draw on specific chunks. In those cases do **not** fabricate citations.
## When the evidence is missing
If neither search nor code traversal surfaces evidence for the question, say: "I cannot find enough information in the knowledge base to answer this question." Do not guess at a number, date, name, or category. Do not infer from tangentially related content. Do not extrapolate from a partial match. A clear refusal is more useful than a fabricated answer.
A computed result counts as evidence only when the code actually located the items the question asks about. If a count, sum, or list comes out empty because the relevant items were not in the document, that is a refusal case, not a "zero" answer.
In a refusal case do **not** call `cite` — there is nothing to cite.
## Important
- Variables persist between `execute_code` calls — you can search in one call and process results in the next