Funtionary vs LLMCompiler

MeetKai / functionary

Chat language model that can use tools and interpret the results

MIT License

1.37k stars 107 forks source link

Funtionary vs LLMCompiler #227

Closed xdevfaheem closed 1 month ago

xdevfaheem commented 2 months ago

Apparently LLMCompiler is extended & optimised version of ReWOO agent architechture with parallel function calling capabilities. It seem's functionary is capable of parallel function calling too. So, what is the core difference between fuctionary and LLMComplier apart from panning steps in LLMC...

Could anyone help me clear this up?

xdevfaheem commented 1 month ago

@khai-meetkai ? (Sorry for pinging)

khai-meetkai commented 1 month ago

Hi @xdevfaheem, I think the basic difference between function calling models (such as Functionary, gpt-40, ...) and LLM-Compiler is:

LLMCompiler first generates a plan (A dependency graph) ahead, then executes functions like traversing all nodes in the graph, each node can be considered as a function call.
Function calling models: do not generate a plan ahead, just analyze the query and the results of previous function calls to decide if we need to execute more functions or not. It is more like ReAct except it can execute parallel function calls at each step.

xdevfaheem commented 1 month ago

Yeah, I thought so. Now, you confirmed it. TQSM.

But in LLMCompiler and ReWOO, After parsing the plan and if there's is any error in execution stage then the agent would fail if there is no replanning stage, but as of now, replanning in LLMCompiler is not efficient and not fault tolerant. I think ReAct still stand tall but functionary is an enhanced version with parallel function calling without fancy thought process prompting. i guess if we add planning and replanning ability to functionary, it would be strainght up lit. (optionally, we can make it stateful agent with fsm lib like xstate).

i guess, i would work on those.