| Hi HN! RΞASON is an OSS Typescript framework for developing LLM apps that uses Typescript's interfaces to get structured output from an LLM. While there are other TS LLM frameworks, I think RΞASON fills a unique space in the market: it's laser-focused on only three areas and, most importantly, actively stays away from pre-made prompting & retrieval. I've been in the LLM space since GPT-3 originally came out, and I've always had problems with other frameworks, such as LangChain. I dislike that they focus a ton on out-of-the-box prompting & pre-made agents — I, as the dev, should be the one in charge of it. My belief is that LLMs are a new primitive that programmers can use — not a new way to program; it's still up to the programmer to do the right thing & create the right abstractions. Therefore, it's the developer's job to learn the new concepts that come from this new primitive, such as prompting & retrieval. I see a similar analogy here with ORMs & SQL. What RΞASON helps with is in areas that don't differentiate your app: getting structured outputs, handling streaming, and observability. The goal of RΞASON is to make creating great LLM experiences easier. We try to accomplish this by simplifying the hard stuff & maximizing performance — decreasing as much as possible the TTUB. RΞASON is OpenTelemetry compatible — which allows observability in almost any tool (Zipkin, Jaeger, paid solutions, etc.). I'd really love to hear feedback about RΞASON! It has been a hobby project for the last months and I'm super curious to what y'all will think. By the way, contributions welcome! |
To be specific - to guarantee correctness you can reject logits for tokens that when used would produce an invalid JSON value returned from the LLM. For example if the string so far is
If the suggested token is something like "3Then you can reject, because the next character needs to be 0-9, and re-sample from the distribution.