langchain-ai/langchain

Configuration

Knobs that affect runtime behavior.

Environment variables

Variable	Effect
`LANGCHAIN_TRACING_V2` / `LANGSMITH_TRACING`	Enable LangSmith tracing for every Runnable invocation
`LANGSMITH_API_KEY` (or `LANGCHAIN_API_KEY`)	API key for LangSmith uploads
`LANGSMITH_ENDPOINT` (or `LANGCHAIN_ENDPOINT`)	Override the LangSmith API base URL
`LANGSMITH_PROJECT` (or `LANGCHAIN_PROJECT`)	Project name to attach traces to
`LANGCHAIN_DEBUG`	Print verbose debug events to stdout
`LANGCHAIN_VERBOSE`	Less verbose stdout printing
`LANGCHAIN_HANDLER`	Reserved for legacy handlers
`OPENAI_API_KEY`	Used by `langchain-openai`
`ANTHROPIC_API_KEY`	Used by `langchain-anthropic`
`GROQ_API_KEY`, `MISTRAL_API_KEY`, `XAI_API_KEY`, `DEEPSEEK_API_KEY`, `FIREWORKS_API_KEY`, `OPENROUTER_API_KEY`, `PERPLEXITY_API_KEY`, `EXA_API_KEY`, `NOMIC_API_KEY`	Per-partner API keys
`OLLAMA_HOST`	Base URL for the Ollama runtime (defaults to `http://localhost:11434`)
`HUGGINGFACEHUB_API_TOKEN`	Hugging Face Hub token
`LANGSMITH_TRACING_V2_DISABLED` / `LANGCHAIN_TRACING_V2_DISABLED`	Force-disable tracing for a specific subprocess

The full set of LangSmith vars is documented in the langsmith SDK; LangChain reads them through the SDK rather than directly.

Global toggles

libs/core/langchain_core/globals.py exposes module-level functions:

from langchain_core.globals import set_debug, set_verbose, set_llm_cache, get_debug, get_verbose, get_llm_cache

set_debug(True)        # register a debug tracer that prints events to stdout
set_verbose(True)      # register a less verbose printer
set_llm_cache(InMemoryCache())   # process-wide LLM response cache

These are convenience hooks for development. Production code attaches handlers via RunnableConfig.callbacks instead.

`RunnableConfig`

libs/core/langchain_core/runnables/config.py defines the per-invocation config TypedDict:

Key	Purpose
`callbacks`	List of `BaseCallbackHandler` instances (or a `CallbackManager`)
`tags`	List of strings for filtering in LangSmith
`metadata`	Arbitrary dict attached to runs
`run_name`	Display name for the run
`run_id`	UUID identifier for the root run
`max_concurrency`	Cap for `batch`, `abatch`, `RunnableParallel`
`recursion_limit`	Cap for nested `RunnableLambda` recursion
`configurable`	Dict overriding `ConfigurableField` attributes on Runnables in the chain

Pass it to any Runnable:

chain.invoke({"input": "..."}, config={"tags": ["pr-42"], "metadata": {"user": "alice"}})

Cache backends

Built-in caches in libs/core/langchain_core/caches.py:

InMemoryCache — process-local, by-key cache.
BaseCache — extend this to plug in Redis, SQLite, etc. (most concrete cache implementations live in langchain-community).

Set globally via set_llm_cache(...), or per-model via model.cache = ....

Rate limiters

libs/core/langchain_core/rate_limiters.py provides InMemoryRateLimiter:

from langchain_core.rate_limiters import InMemoryRateLimiter

rate_limiter = InMemoryRateLimiter(
    requests_per_second=10,
    check_every_n_seconds=0.1,
    max_bucket_size=10,
)
model = ChatOpenAI(rate_limiter=rate_limiter)

The model awaits the rate limiter before each call.

Model-level kwargs

Every partner's chat model accepts:

model — model name
temperature, top_p, max_tokens — sampling
timeout, max_retries — HTTP behavior
api_key, base_url, organization — auth and routing
default_headers, default_query — request-level overrides
streaming, stream_usage — streaming behavior
cache — per-model cache override
callbacks — per-model callback handlers (legacy; prefer config["callbacks"])
tags, metadata — defaults for every call

Provider-specific kwargs vary; see the partner page for each.

CI configuration

CI behavior is configured per-workflow under .github/workflows/:

check_diffs.yml — which packages run on which file changes
_test.yml — Python version matrix (3.10, 3.11, 3.12, 3.13, 3.14)
_test_pydantic.yml — Pydantic version matrix
pr_lint.yml — allowed PR title types and scopes
pr_labeler.yml — automatic labeling rules