Episode 185 of Thinking Elixir. Dive into the world of structured LLM prompting with our latest guest who shares insights on their innovative project, InstructorEx. We’ll explore how this tool is pushing the boundaries of large language models to return structured data, particularly JSON, and the benefits it offers when integrated into existing systems. The conversation takes a deep dive into the challenges and solutions of forcing LLMs to output data in desirable formats using Elixir technology, like Ecto schemas, and what it means to treat an LLM like a digital human, even applying it to tasks like spam detection in emails. Discover the intriguing concept of using a local database as a cache for LLM results. If you’re curious about the intersection of structured data, Python experience transition to Elixir, and the future of LLM integration, this is an episode not to miss. Tune in for these engaging topics and more!
I was trying to coerce the output completions from OpenAI for my specific use case into a more structured format & was hell-bent not to leave Elixir land.
Was looking for the Elixir equivalent of Output Parsers in Python LangChain & this was super timely and helpful, thank you!
Would something similar to Output Parsers be on the roadmap for Elixir LangChain?
(I imagine Instructor being one of many outputs that follow some defined behaviour)
Hi @shawn_leong! I’m hoping that Bumblebee can get support for the same type of output constraint that LlamaCPP supports.
If that can happen, then I’d love to add that to Elixir LangChain.
What type of output parsing were you looking for? When working with OpenAI and ChatGPT, the easiest I’ve found is creating a function that takes the data in the structure you want.