Search over 2,200 filings with source links
An internal research tool for Emerald Wealth Partners. It helps analysts find company evidence in SEC filings and earnings call transcripts, then check each answer against its sources.

- Client
- Emerald Wealth Partners
- Timeline
- 7 weeks
- Tools
- OpenAI, Supabase, SEC EDGAR, custom pipeline
Overview
The task was to make research evidence accessible
Three investment staff tracked a 450-company watchlist. Preparing a qualitative brief for a committee meeting took two to three days of reading annual filings and earnings call transcripts. Financial data was already available through FactSet; finding and comparing what management had said still took manual work.
The team had a prototype using uploaded PDFs and an API. It helped with single documents but could not maintain the full research library or reliably compare companies across years. Some answers included quotes that did not match their citations.
The project focused on retrieving and explaining evidence. Analysts kept responsibility for interpreting it and making investment decisions.
Corpus
Coverage came before model choice
The watchlist had 450 companies. We indexed 400 with the required filing history: 2,200 annual reports and transcripts. The pipeline pulls filings from SEC EDGAR and updates the library weekly without adding duplicate records.
The system uses small passages of about 220 words to find relevant text, then gives the answer stage larger windows of about 650 words so it can read the surrounding context. Transcripts keep each analyst question with its answer.
Retrieval
Company matching and citations needed checks
The search combines keyword and semantic matching. Names, dates, and tickers need exact matches. Questions about tone or capital discipline also need matching by meaning.
Early results sometimes came from the wrong company. We changed ticker matching to use whole words and removed common terms from legal names that could create false matches. We also expanded abbreviations, so a query with “AI” could find passages that use “artificial intelligence.”
The answer checks that citations point to retrieved passages. A second check compares the answer with those passages and flags unsupported claims. If the evidence is weak, the tool says it has no data. If coverage is limited, it reports that separately. Analysts can open source passages to review the evidence.
Evaluation
Tests changed as the research library grew
The evaluation set grew from 13 to 83 cases. It covers company matching, comparisons, historical questions, filtered searches, and questions the library cannot answer. Some early tests expected a fixed company list; as coverage grew, the correct result included more companies.
We tested 29 retrieval settings against the full test suite. Returning fewer candidates for broad questions improved ranking without reducing accuracy in those tests. Combining settings that worked well alone sometimes made results worse.
The full research question was the useful unit of evaluation: company coverage and answer support both mattered. A higher score on one retrieval measure was not enough to keep a change.
Result
Committee research took 10 minutes
At 60 days, Emerald’s internal reporting showed committee research preparation at 10 minutes, down from two to three days. The team evaluated 12–15 new positions per quarter, up from 4–5, and ran more than 50 queries per week.
The research corpus stayed in the client’s private database. Queries and selected passages went to the model provider through the client’s API account. Ongoing costs were about $150–250 per month at the team’s usage level.