Financial Services
Faster search turnaround versus the legacy keyword system
System Usability Scale score achieved (out of 100)
About
The firm is a boutique investment research firm whose analysts needed plain-English access to thousands of SEC filings instead of building arcane Boolean search strings by hand. We built a fine-tuned semantic search layer with a full compliance audit trail.
Industry
Financial Services
Company size
100 – 500 employees
Founded
2014
The Company
An analyst team spending more time querying than analysing
The firm provides fundamental investment research grounded in public SEC filings, covering a wide universe of listed companies for institutional clients. Analysts needed reliable, fast access to thousands of 10-K, 10-Q, 8-K, and proxy filings for comparative analysis.
The firm's existing search tooling relied on 256-character Boolean query strings against the legacy EDGAR interface — powerful in principle, but slow and error-prone to construct correctly for a nuanced comparative question.
The challenge
Arcane query syntax was costing three staff-hours a day
Constructing a Boolean search precise enough to surface the right filings — without either missing relevant documents or drowning in irrelevant ones — consumed roughly three staff-hours across the analyst team every day, time that came directly out of the interpretive work that was the firm's actual value-add.
Compliance also required an audit trail of what was searched and what was found, which the legacy Boolean interface did not support in any structured, exportable way — creating a second layer of manual record-keeping on top of the search itself.
The Solution
A QLoRA-tuned semantic search layer with a built-in audit trail
We fine-tuned a MiniLM retrieval model using QLoRA on a corpus of SEC filings and deployed it behind a Streamlit interface, letting analysts search in plain English rather than Boolean syntax. Every query and its results are logged to DuckDB, giving compliance the exportable, long-term audit trail the legacy system couldn't provide.
The interface supports comparative queries across companies and reporting periods directly, surfacing the specific filing passages that support a result rather than returning a flat list of document links for the analyst to open one by one.
The Results
3.4x faster search and a usability score in the top tier
Search speed improved 3.4x against the legacy Boolean interface, and the system scored 91 on the System Usability Scale — a strong result that reflected how directly analysts could go from question to answer without translation into query syntax. The three staff-hours a day previously spent on query construction across the team were eliminated entirely.
Compliance gained a structured, exportable audit trail for the first time, satisfying an internal requirement that the legacy tooling had never been able to meet without a parallel manual logging process.
KEEP READING
Explore more studies on AI, analytics, and enterprise intelligence.
