Hi community, looking over controlflow, looks super awesome, some q’s I have , I would likely want to use this with a vector db etc and implement RAG are there plans to add plugins or maybe even piggyback off of llama index’s plugins similar to what langchain does? I noticed it uses the llm providers from langchain so I assume this should be possible, if we can already do RAG stuff, are there any examples?
n
Nate
09/22/2024, 11:15 PM
hey @Cody Webb - just fyi there is a #C079VLLH5D3 channel here!
on RAG in controlflow, I'd defer to @Jeremiah but I would say that ingestion / embedding / querying from vectorstores falls outside of the intended scope of what controlflow would offer natively
that's not to say you shouldn't build a RAG system using controlflow, for example
its more to say that integrations with a given vectorstore, you might want to use another tool for, or make yourself
c
Cody Webb
09/22/2024, 11:16 PM
Yea that’s what I was thinking, just do the RAG pipeline ingestion/ querying etc as a tool
👍 1
j
Jeremiah
09/22/2024, 11:30 PM
Hey @Cody Webb, welcome! As Nate said, controlflow is more focused on coordinating the agentic behavior and doesn't express an opinion on any RAG framework. As long as your RAG entrypoint can be expressed as a Python function, any CF agent should happily accept it as a tool (just pass the function directly to the agent or task's tool attribute). Nate's (very real) example is exactly what I'd have in mind
Jeremiah
09/22/2024, 11:31 PM
I am actually adding a first class vector-based feature this evening to give agents flexible memories and even that uses this pattern
🙌 1
c
Cody Webb
09/22/2024, 11:32 PM
Very cool thanks , that makes total sense I wouldn’t want it to get as bloated as LC etc :) was mostly wondering how much would work hand in hand since I noticed you are using the llm providers
j
Jeremiah
09/22/2024, 11:36 PM
Ya, it's a good exercise in focusing on something I think CF is uniquely good at (and honestly I would like to offer an opinionated RAG framework if time and bandwidth allowed, but more because it works better with CF than because we have a unique opinion on what is the "best" way). So instead we're just trying to make sure our agents are as compatible as possible with whatever other AI ecosystem tools you have. If you hit any rough edges please let us know, we'd love to help!
🙌 1
c
Cody Webb
09/22/2024, 11:41 PM
Also just noticed you’re in DC (me too!) , I do DS/AI
advising/POCs for a few diff agencies and have been trying to introduce them to prefect as they are struggling with airflow/hadoop and trying to modernize 🙂 will ping you if successful
j
Jeremiah
09/22/2024, 11:42 PM
🙌 awesome! We actually have some parallel stuff going on there -- will DM