Equity research search with PineGap

PineGap’s analysts were spending hours reading filings to answer questions that should take minutes. The answers were in the documents. Finding them meant reading and summarizing by hand, across a corpus of more than a million filings.
Speed alone wasn’t enough. An analyst can’t put an answer in front of a client unless they can see where it came from, so every answer had to point back to the filing behind it.
Getting the Filings Into Shape
Filings are long, inconsistent documents full of tables. LlamaParse handles the ingestion, and a chunking pipeline breaks each filing into passages small enough to search while keeping track of the filing each passage came from. That source link is what lets the final answer cite its evidence.
Retrieval That Finds the Right Passage
Search runs in two stages. Hybrid retrieval over OpenAI embeddings in Qdrant combines matching by meaning with matching by exact terms. That pulls in a wide set of candidates, including filings that use different wording and ones that hinge on a specific term or figure. A cross-encoder reranker then reads each candidate against the question and puts the most relevant passages first, which is where precision comes from.
Answers Grounded in the Sources
GPT-4o writes the final answer from the reranked passages, using structured prompts that keep it tied to what was retrieved. Analysts use it through a dashboard where they ask questions and read the answers alongside their sources.
Impact
Research that took hours now takes minutes. Reports go out faster, and analysts respond to clients sooner.
In short, the engagement delivered:
- An ingestion and chunking pipeline for 1M+ filings.
- Hybrid vector search with cross-encoder reranking for precision.
- LLM synthesis of final answers from the retrieved passages.
- A Next.js and FastAPI dashboard for analysts.
The system is built with LlamaParse, OpenAI embeddings, Qdrant, a cross-encoder reranker, GPT-4o, Next.js and FastAPI.