DocETL Adds Python DSL And Features To Reduce LLM Output Slop
Reactions from ranked influencers
2 posts
Really great to hear. We have put in lots of engineering cycles to improve DocETL, including having a Python DSL and skill engineering. Also some exciting new features coming soon around deslopifying long form LLM outputs :-)
Wikis are pure slop Pure slop is LLM content from summaries / groups of other LLM content in a big sloppy mess using other peoples prompts built for generic projects (my definition). It's my current belief that if you want to do analysis of large scale unstructured documents with AI today, you currently need constrained pipelines with steps that were verified by a domain expert in the loop. DocETL from @sh_reya has gotten a lot easier to use since I used it last. It's skill guides you through building your own AI enabled ETL pipelines and helps you verify after each step - it also handles all of the LLM caching and helps you test with multiple LLMs. If you want to build a Wiki for your repo, there are lots of good learnings from slop cannon wiki builders, but I'm of the belief that if you're introducing a wiki, you probably should understand, tune and verify how it was produced.
Wikis are pure slop Pure slop is LLM content from summaries / groups of other LLM content in a big sloppy mess using other peoples prompts built for generic projects (my definition). It's my current belief that if you want to do analysis of large scale unstructured documents with AI today, you currently need constrained pipelines with steps that were verified by a domain expert in the loop. DocETL from @sh_reya has gotten a lot easier to use since I used it last. It's skill guides you through building your own AI enabled ETL pipelines and helps you verify after each step - it also handles all of the LLM caching and helps you test with multiple LLMs. If you want to build a Wiki for your repo, there are lots of good learnings from slop cannon wiki builders, but I'm of the belief that if you're introducing a wiki, you probably should understand, tune and verify how it was produced.