langchain-glean

name	langchain-glean
description	Use the langchain-glean package to call Glean from LangChain — chat models (ChatGlean / ChatGleanAgent), retrievers over Glean search and the people directory, and Glean tools for LangChain agents. Load when writing Python that imports langchain_glean or wires Glean into a LangChain app, chain, or agent.

langchain-glean

LangChain integration for Glean. It wraps the Glean Python SDK (glean-api-client) behind LangChain's Runnable interface, exposing Glean as chat models, retrievers, and agent tools.

When to use

Load this skill when writing Python that imports langchain_glean, or when wiring Glean into a LangChain app — a Glean-backed chat model, a retriever over Glean search or the people directory, or Glean tools inside a tool-calling agent. For talking to the Glean API without LangChain, use glean-api-client directly instead.

Install & import

pip install -U langchain-glean

Every public entry point is re-exported from the top-level package:

from langchain_glean import (
    ChatGlean, ChatGleanAgent,
    GleanSearchRetriever, GleanPeopleProfileRetriever,
    GleanToolkit,
    GleanSearchTool, GleanPeopleProfileSearchTool, GleanChatTool,
    GleanListAgentsTool, GleanGetAgentSchemaTool, GleanRunAgentTool,
)

The simple request models live beside their component: from langchain_glean.chat_models import ChatBasicRequest, from langchain_glean.retrievers import SearchBasicRequest, PeopleProfileBasicRequest.

Authoritative API

The package ships a py.typed marker and is fully typed (mypy strict). Read the inline type hints in the installed package for exact constructor arguments, request-model fields, and return types — do not guess them:

• the component classes under langchain_glean/chat_models/, retrievers/, tools/, and toolkit.py
• the simple request models (ChatBasicRequest, SearchBasicRequest, PeopleProfileBasicRequest) defined alongside them
• for full-control requests, the Glean SDK models in the glean-api-client dependency (from glean.api_client import models) — read that package's types too

Components implement LangChain's Runnable interface, so invoke / ainvoke / stream / astream come from langchain-core; defer to its docs for the runnable surface.

Usage patterns

• Auth is environment-first. Set GLEAN_API_TOKEN and GLEAN_SERVER_URL (the full backend URL; GLEAN_INSTANCE is a deprecated fallback). With a global token, also set GLEAN_ACT_AS to the user to act as. All components share this resolution through a common mixin, so you don't build an SDK client by hand — set the env (or pass the same values as constructor args) and construct the component.
• Three input styles; escalate only when needed. (1) a plain query string (retrievers), (2) a *BasicRequest Pydantic model for the common options, (3) a full glean.api_client model (SearchRequest, ChatRequest, ListEntitiesRequest) for complete control. Start at (1)/(2); reach for (3) only for fields the basic model doesn't expose.
• Chat. ChatGlean wraps the Glean Assistant; ChatGleanAgent(agent_id=...) targets one specific agent. Both take LangChain messages and return LangChain messages; use stream/astream to stream, and resume a conversation with chat_id.
• Retrievers return LangChain Documents from Glean search or the people directory — drop them straight into a RAG chain.
• Agent tools wrap the same capabilities for tool-calling agents, and GleanToolkit aggregates them. The agent-discovery tools are meant to be chained in order: GleanListAgentsTool → GleanGetAgentSchemaTool → GleanRunAgentTool.
• Async is everywhere — every retriever, tool, and chat model exposes ainvoke/astream.

Common mistakes

• Hand-constructing a Glean client or threading tokens through each call. Auth resolves from env/constructor via the shared mixin; configure once, then just construct the component.
• Using GLEAN_INSTANCE in new code — it's a deprecated fallback; prefer the full GLEAN_SERVER_URL.
• Omitting GLEAN_ACT_AS with a global token — global tokens must name the acting user or calls fail. (User-scoped tokens don't need it.)
• Guessing request-model field names. *BasicRequest and the glean.api_client models are typed — read the hints instead of inventing fields.
• Reaching for a full SDK request object when a string or *BasicRequest would do — extra complexity for no gain.
• Confusing the chat classes: ChatGlean is the Assistant; ChatGleanAgent is one agent selected by agent_id.

Version notes

Check the installed version with pip show langchain-glean or python -c "import langchain_glean; print(langchain_glean.__version__)". It pins glean-api-client (the generated Glean Python SDK) and langchain-core; after upgrading either, re-read the types for changed request-model fields rather than trusting older snippets, and consult CHANGELOG.md for breaking changes. Requires Python 3.9+.