Our goal is to build tooling to eliminate hallucinations for data-driven queries.

We have built a demo that showcases an application of this technology - natural language basketball statistic search. https://basketball.offbynone.ai

This enables asking data-driven questions in ways that have never been possible before.

Try asking it things like "If Shaq scored half as many points, which victories turn into defeats?" or "Of the players who have missed half their shots, who has played in the most games?" These questions are unanswerable by current LLM technology, even when enhanced by RAG.

The tooling we have built works from URLs or uploaded files. It auto-detects the structure and semantic meanings of the data into a blueprint. This blueprint is used to import the data into a mongo database. Additionally, it uses the blueprint to convert human-language questions into database queries.

You can have a high degree of confidence (but as with all llm technology, not 100% confidence) in the generated answers because you can verify that the LLM isn't making up the data. You can see the queries used to fetch the data as well as the raw data fetched from the database. This approach eliminates issues with the LLM making up data, but the generated answers are only as accurate as the provided source data. For example, some of the NBA records for 3 point attempts before 1982 can be a little spotty, so questions involving percentage accuracy before that date might not be accurate. Like all LLM technology, there is also a chance that it misunderstand what you are asking for.

I'm open to feedback and questions. We are open to the possibility of open-sourcing the tooling if there is enough interest.