FinanceBench evidence-page finder
Use this skill when you must point at the exact page of a 10-K or 10-Q that supports an answer, and you want a reproducible citation instead of a guess.
When to use it
- A question names a company, a fiscal year and a metric, and you must return the page where that metric is reported.
- You already hold the full PDF and can extract page text, but you do not have a vector index or an embedding model available.
Method
- Split the PDF into physical pages and keep the page number with every page.
- Remove function words, then cut the question into two-word probes. Long whole sentence queries are rejected by coverage-based admission filters, so never use them as the only query.
- Search each probe separately against the page index and fuse the results with
reciprocal rank fusion (
1 / (60 + rank)), summing per page. - Return at most the requested number of distinct pages, ordered by fused score.
Why two-word probes
Short probes keep the admission floor at one matching term. A long question requires a large share of its terms to match before a page is even considered, and evidence pages in financial filings often repeat only a couple of the question's words. On the public FinanceBench PepsiCo slice this raised evidence-page Recall@20 from 0.625 to 0.750.
Limits
Lexical retrieval cannot bridge a semantic gap: a page that reports "one customer accounted for 16% of net revenue" never contains the phrase "customer concentration". Treat a lexical result as a candidate list to verify by reading, and say so when the evidence is only indirect.
<!-- OB-S:NoLSVk2qft2aG-WTRnPJjg -->
微信扫一扫