Show HN: Pongo – 80% Fewer LLM Hallucinations with one LoC (joinpongo.com)

We’re building Pongo, a new retrieval layer that can cut LLM hallucinations by as much as 80% in RAG pipelines. You can see the full benchmarks on our website, where we saw RAG errors/hallucinations fall from ~19% to ~3% after adding Pongo.

This has huge implications for AI applications, getting the right answer can be the difference between a user churning and adopting a product. For a 3 step agent, each step having a success rate of 80% compared to 97%, is the difference between a 51% and 92% success rate for the workflow as a whole.

Pongo sits at the end of existing retrieval pipeline, whether it’s a vector database or a lexical search engine. You send in your top 100-300 results along with the query, then it uses a mix of models and retrieval methods to score and order the results in as little as 0.5s. The API is just 1 line of code.

The accuracy difference comes from a two main factors

1. Surfacing relevant results that the initial search ranked outside of the top 10 results. 2. Moving the right results from ranks 5-15 to ranks 1-3, which significantly decreases strain on the LLM’s attention window.

This jump in performance stems from out multi-model approach since each method has non-overlapping failure cases. We utilize multi-vector models, cross-encoder models, and sparse vector results in our ranking algorithm. This is a pretty compute intensive process and took quite a bit of work to get this to run with as little latency as it does.

Would love to get some feedback, and see how it holds up in your projects.