LangChain Adapter

Callback handler that hooks into on_llm_start / on_llm_end for per-LLM-call scheduling.

Setup

from loco import AsyncLOCOScheduler, SharedResource
from loco.adapters.langchain import LOCOCallbackHandler

scheduler = AsyncLOCOScheduler([], SharedResource("llm_api", capacity=3))
callback = LOCOCallbackHandler(scheduler, agent_id="rag-pipeline")

Usage

Attach the callback to your LLM:

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(model="gpt-4o", callbacks=[callback])
result = await llm.ainvoke("Summarize this...")

Each LLM call is automatically scheduled through LOCO. The callback extracts the model name from LangChain's serialized config and computes weight.

Multiple Agents

Create one callback per agent:

rag_callback = LOCOCallbackHandler(scheduler, agent_id="rag-pipeline")
qa_callback = LOCOCallbackHandler(scheduler, agent_id="qa-chain")

rag_llm = ChatOpenAI(callbacks=[rag_callback])
qa_llm = ChatOpenAI(callbacks=[qa_callback])

Both agents compete through the same scheduler.