GitHub

I build Sifaka as an experiment in reducing hallucinations through iterative feedback using a different LLM. As a began researching it, I found several research papers that deal with LLM "reflection" -- having an LLM reflect on its output to improve it. The idea is incredibly fascinating to me, and I used the insights of those papers to develop what is now Sifaka. I believe Sifaka has the most promise in reviewing legal documents (using the ollama/legal_model for example. I think it might also be useful for highly regulated documents such as government policy docs, or highly technical ones like product documentation. I'm looking for contributors to the project, if you're interested :) Right now, I'm working on evaluating improvement quality. I have a few ideas, but I'd love to come up with additional methods to evaluate it.