A cheaper way to search SEC filings with AI agents
Two founders with s in have spent seven months building a finance search system. The main idea is to avoid making an AI agent rediscover the same links between facts every time someone asks a question. More of that work is done once when documents are added, then reused later.
The system uses an that automatically reorganizes documents and connects related information in a . It is described as different from a classic and different from GraphRAG. The test data was the latest S&P 500 10-K filings.
Changing the data structure and moving more compute to the stage appeared to make faster and cheaper.
Key points
- The system shifts repeated work from question time to document time.
- An restructures filings and links related information in a .
- The test used recent S&P 500 10-K filings.
- The claimed benefit is faster and cheaper for dense finance documents.
- No clear benchmark numbers for cost, speed, or accuracy are included.