A lightweight question runner that answers metadata-focused questions
(signatures, sign status, completion, dates) using a single
structured-output LLM call against compact per-type summary tables.
No agent loop, no vector retrieval, no full document fetching. Designed
as a cheap shortcut: when the LLM answers confidently (pass/fail) we use
the result; otherwise it throws NotEnoughInformationError and the
composite runner falls through to the full agent pipeline.
A lightweight question runner that answers metadata-focused questions (signatures, sign status, completion, dates) using a single structured-output LLM call against compact per-type summary tables.
No agent loop, no vector retrieval, no full document fetching. Designed as a cheap shortcut: when the LLM answers confidently (pass/fail) we use the result; otherwise it throws
NotEnoughInformationErrorand the composite runner falls through to the full agent pipeline.