In one line
Sage retrieves evidence with both keyword and semantic search, reranks it, and returns page-level citations. The goal is a useful answer that a reader can verify, not a fluent answer that asks to be trusted.
Question
Can a research assistant answer questions across long financial filings while showing exactly which pages support the answer?
What I built
- Combined BM25 keyword search and vector retrieval with reciprocal-rank fusion so exact financial terms and semantically related passages both matter.
- Balanced retrieval across companies before cross-encoder reranking to prevent one large filing from dominating the context.
- Resolved citations back to source pages and built an evaluation set from three real SEC filings.
Main result
- documented evaluation checks passed
- 19/19
- backend tests passing
- 290
What the work showed
- Hybrid retrieval was more dependable than relying on keyword or vector search alone for financial language.
- Company-balanced retrieval improved multi-company questions by keeping evidence from each requested filing in play.
- Browser-based user testing exposed citation and retrieval failures that ordinary unit tests did not reveal.
What it does not prove
- A citation proves where a statement came from; it does not guarantee that the model interpreted the passage correctly.
- The recorded evaluation covers three SEC filings and should not be treated as evidence for every filing type or research question.
Conclusion
Sage is useful when auditability matters more than a polished paragraph. Its design keeps retrieval, reranking, and citations visible enough to debug—and gives the reader a direct path back to the source.