langchain-distil

Reversible distil context compression for LangChain and LangGraph.

distil compresses an agent's accumulated messages before the model call: tool/function messages get a reversible digest, human/system messages get lossless compression, and the model's own (ai/assistant) messages are never rewritten. Originals persist in a local handle store and are recoverable via distil's distil_expand MCP tool or distil.mcp_server.load_restore(handle).

Decision-equivalence between compressed and full context is certified separately, offline, by distil bench / distil validate. This package performs the compression only.

Install

pip install langchain-distil                 # message-dict hooks
pip install "langchain-distil[runnable]"     # + as_runnable() (needs langchain-core)

Usage

LangGraph — drop distil in as a pre_model_hook:

from langgraph.prebuilt import create_react_agent
from langchain_distil import pre_model_hook

agent = create_react_agent(model, tools, pre_model_hook=pre_model_hook())

Any message list — compress before .invoke:

from langchain_distil import compress_messages

resp = model.invoke(compress_messages(messages))

As a Runnable in a chain:

from langchain_distil import as_runnable

chain = as_runnable() | model

verbatim=True selects Tier-0 lossless mode (no digest stubs).

License

MIT