Jev Index Implementation

Sep 19

This project tests and implements Jev, an AI tool that provides consistent, near-deterministic outputs for natural language queries, leading to performance and cost improvements in workloads like recipe importing.

I'm testing and implementing Jev (@typesafeai), and it's working great so far. No failed requests, and always extremely fast (sub-second). Its responses are consistent, so it's nearly deterministic. This means a natural-language query gets a consistent, static output. One workload I'm working on is recipe importing (by image, text, or URL). I can only offload some of the work, but there's still an average 14% time improvement, and I assume about that much cost benefit too. It appears Jev has been trained on a large recipe and ingredient subset, because it's confidently and correctly handling ingredient categorization. All my existing tests are green. I have it falling back to the LLM(s) when needed, so I'm not introducing another point of failure. Which means I get everything that I had before, plus a 14% performance boost. This is just my first workload, but I'm excited for the possibilities.

Original post