👋 Hi there! Thank you for being interested in contributing to LangGraph. As an open source project in a rapidly developing field, we are extremely open to contributions, whether it be in the form of a new feature, improved infra, or better documentation.
To contribute to this project, please follow a "fork and pull request" workflow. Please do not try to push directly to this repo unless you are a maintainer.
If you are not sure what to work on, we have a few suggestions:
- Look at the issues with the help wanted label. These are issues that we think are good targets for contributors. If you are interested in working on one of these, please comment on the issue so that we can assign it to you. And if you have any questions let us know, we're happy to guide you!
We aim to keep the same core APIs between the Python and JS versions of LangGraph, where possible. As such we ask that if you have an idea for a new abstraction, please open an issue first to discuss it. This will help us make sure that the API is consistent across both versions. If you're not sure what to work on, we recommend looking at the links above first.
Our issues page contains with bugs, improvements, and feature requests.
If you start working on an issue, please comment and a maintainer can assign it to you.
If you are adding an issue, please try to keep it focused on a single modular bug/improvement/feature. If the two issues are related, or blocking, please link them rather than keep them as one single one.
We will try to keep these issues as up to date as possible, though with the rapid rate of development in this field some may get out of date. If you notice this happening, please just let us know.
Although we try to have a developer setup to make it as easy as possible for others to contribute (see below) it is possible that some pain point may arise around environment setup, linting, documentation, or other. Should that occur, please contact a maintainer! Not only do we want to help get you unblocked, but we also want to make sure that the process is smooth for future contributors.
In a similar vein, we do enforce certain linting, formatting, and documentation standards in the codebase. If you are finding these difficult (or even just annoying) to work with, feel free to contact a maintainer for help - we do not want these to get in the way of getting good code into the codebase.
As of now, LangGraph has an ad hoc release process: releases are cut with high frequency via by a developer and published to npm.
LangChain follows the semver versioning standard. However, as pre-1.0 software, even patch releases may contain non-backwards-compatible changes.
If your contribution has made its way into a release, we will want to give you credit on Twitter (only if you want though)! If you have a Twitter account you would like us to mention, please let us know in the PR or in another manner.
You can invoke the release flow by calling yarn release
from the package root.
There are three parameters which can be passed to this script, one required and two optional.
- Required:
--workspace <workspace name>
. eg:--workspace @langchain/langgraph
(always appended as the first flag when runningyarn release
) - Optional:
--bump-deps
eg--bump-deps
Will find all packages in the repo which depend on this workspace and checkout a new branch, update the dep version, run yarn install, commit & push to new branch. - Optional:
--tag <tag>
eg--tag beta
Add a tag to the NPM release.
This script automatically bumps the package version, creates a new release branch with the changes, pushes the branch to GitHub, uses release-it
to automatically release to NPM, and more depending on the flags passed.
Halfway through this script, you'll be prompted to enter an NPM OTP (typically from an authenticator app). This value is not stored anywhere and is only used to authenticate the NPM release.
Full example: yarn release @langchain/langgraph --bump-deps --tag beta
.
This project uses the following tools, which are worth getting familiar with if you plan to contribute:
- yarn (v3.4.1) - dependency management
- eslint - enforcing standard lint rules
- prettier - enforcing standard code formatting
- jest - testing code
Clone this repo, then cd into it:
cd langgraph
Next, try running the following common tasks:
Our goal is to make it as easy as possible for you to contribute to this project.
All of the below commands should be run from within a workspace directory (e.g. langgraph
) unless otherwise noted.
cd langgraph
To get started, you will need to install the dependencies for the project. To do so, run:
yarn
Then you can build the project with:
yarn build
We use eslint to enforce standard lint rules. To run the linter, run:
yarn lint
or to automatically fix linting errors, run:
yarn lint:fix
We use prettier to enforce code formatting style. To run the formatter, run:
yarn format
To just check for formatting differences, without fixing them, run:
yarn format:check
In general, tests should be added within a tests/
folder alongside the modules they
are testing.
Unit tests cover modular logic that does not require calls to outside APIs.
If you add new logic, please add a unit test.
Unit tests should be called *.test.ts
.
To run only unit tests, run:
yarn test
To run a single test, run the following from within a workspace:
yarn test:single /path/to/yourtest.test.ts
This is useful for developing individual features.
Integration tests cover logic that requires making calls to outside APIs (often integration with other services).
If you add support for a new external API, please add a new integration test.
Integration tests should be called *.int.test.ts
.
Note that most integration tests require credentials or other setup. You will likely need to set up a langgraph/.env
file
like the example here.
We generally recommend only running integration tests with yarn test:single
, but if you want to run all integration tests, run:
yarn test:integration
To build the project, run:
yarn build
LangGraph exposes multiple subpaths the user can import from, e.g.
import { Pregel } from "@langchain/langgraph/pregel";
We call these subpaths "entrypoints". In general, you should create a new entrypoint if you are adding a new integration with a 3rd party library. If you're adding self-contained functionality without any external dependencies, you can add it to an existing entrypoint.
In order to declare a new entrypoint that users can import from, you
should edit the langgraph/scripts/create-entrypoints.js
script. To add an
entrypoint tools
that imports from tools/index.ts
you'd add
the following to the entrypoints
variable:
const entrypoints = {
// ...
tools: "tools/index",
};
This will make sure the entrypoint is included in the published package, and in generated documentation.
To run docs locally, switch to the docs/
directory:
cd docs
Then, install the required dependencies with:
pip install -r docs-requirements.txt
Finally, run:
make serve-docs
And navigate to http://127.0.0.1:8000/
to see your local build.
If you are contributing to this code base, then you need to be familiar with some of the underlying concepts that power the Graph
class - LangGraph's main entrypoint. These concepts are intentionally not documented in the LangGraph docs because users of LangGraph do not need to understand them. This knowledge is exclusively for contributors.
Let's start with Pregel. Pregel (PDF) is an API for building graphs.
Some key concepts:
- Pregel graphs take an
input
andoutput
as parameters which represent where the graph starts and ends - Pregel graphs take a mapping of nodes represented as
{nodeName: node}
- Each node subscribes to (one or more) channels. This defines when a node executes. Specifically, for a given
node N
that subscribes tochannel M
, whenever the value ofchannel M
changes,node N
must be executed. Intuitively, it represents what the current node is dependent on. - Each node writes to (one or more) channels. This defines where the final value after a node is executed is stored, because nodes don't store their own value.
- More on channels below.
To form an intuition around Pregel graphs, lets look at the example tests.
In the example below, the pregel graph is defined to start at inputChannelName
and end at outputChannelName
. The graph has a single node called nodeOne
that transforms the input value by adding one to it. When the graph is invoked with an input value of 2
:
inputChannelName
gets set to a value of2
, because it is defined as the input channel.- Since
nodeOne
subscribes toinputChannelName
,nodeOne
executed. nodeOne
transforms2
to3
and3
gets written tooutputChannelName
.- Since
outputChannelName
is defined as the graph's output, the execution ends and returns3
.
const addOne = jest.fn((x: number): number => x + 1);
const chain = Channel.subscribeTo("inputChannelName")
.pipe(addOne)
.pipe(Channel.writeTo("outputChannelName"));
const app = new Pregel({
nodes: { nodeOne: chain },
input: ["inputChannelName"],
output: ["outputChannelName"],
});
expect(await app.invoke({ input: 2 })).toEqual({ output: 3 });
This was a simple example, let's look at a more complicated example.
In the example below, the graph has one node. The checkpointer
parameter in Pregel means that it persists the state at every step. If a checkpointer is specified, then thread_id
must be specified every time the graph is invoked and it represents the unique id of that invocation.
Invocation 1:
- When the graph is invoked with
2
,input
channels value becomes2
- Node
one
runs because it is subscribed toinput
. The node tranforms2
to2
by runninginputPlusTotal
. - The value of channels
output
andtotal
get set to2
because nodeone
writes to both channels - Because
memory
is passed into the graph,total
at thread_id of1
is saved as a value of2
- The graph ends with
output
's value which is2
Invocation 2:
input
channel value set to3
- Node
one
triggered. Node one transforms3
tototalValue + 3
=2 + 3
=5
total
is aBinaryOperatorAggregate
channel. Hence, it transforms the inbox value5
to5 + prevTotalValue
=5 + 2
=7
output
channel's value is written as5
and the graph returns with5
Invocation 3:
input
channel value set to5
with athread_id
of2
indicating a new id for storage- Node
one
triggered. Node one transforms5
tototalValue_in_thread_id_2 + 3
=0 + 5
=5
- Checking the value of
total
inthread_id_1
is still the same as the value in invocation 2 which is7
. output
channel's value is written as5
and the graph returns with5
it("should handle checkpoints correctly", async () => {
const inputPlusTotal = jest.fn(
(x: { total: number; input: number }): number => x.total + x.input,
);
const one = Channel.subscribeTo(["input"])
.join(["total"])
.pipe(inputPlusTotal)
.pipe(Channel.writeTo("output", "total"));
const memory = new MemorySaver();
const app = new Pregel({
nodes: { one },
channels: { total: new BinaryOperatorAggregate<number>((a, b) => a + b) },
checkpointer: memory,
});
// Invocation 1
await expect(
app.invoke(2, { configurable: { thread_id: "1" } }),
).resolves.toBe(2);
let checkpoint = memory.get({ configurable: { thread_id: "1" } });
expect(checkpoint).not.toBeNull();
expect(checkpoint?.channelValues.total).toBe(2);
// Invocation 2
await expect(
app.invoke(3, { configurable: { thread_id: "1" } }),
).resolves.toBe(5);
checkpoint = memory.get({ configurable: { thread_id: "1" } });
expect(checkpoint?.channelValues.total).toBe(7);
// Invocation 3
await expect(
app.invoke(5, { configurable: { thread_id: "2" } }),
).resolves.toBe(5);
checkpoint = memory.get({ configurable: { thread_id: "2" } });
expect(checkpoint?.channelValues.total).toBe(5);
checkpoint = memory.get({ configurable: { thread_id: "1" } });
expect(checkpoint?.channelValues.total).toBe(7);
});
Those are some of the fundamentals of how a Pregel graph works. To get a deeper understanding of how Pregel works, you can check out it's expected behavior in pregel.test.ts
.
Some concepts about channels:
- Channels are the way nodes communicate with one another in Pregel. If it were not for channels, nodes would have no way of storing values or denoting dependencies on other nodes.
- At it's core, every channel does a couple things:
- It stores a current value.
- It implements a way to
update
it's current value based on the expected parameter for the update function. - It implements a way to
checkpoint
or "snapshot" the current state of the channel. This enables persistence across a graph. - It implements a way to
empty
or "restore" a channel from a checkpoint/snapshot. This enables us to create a new channel from a checkpoint variable stored in a database.
channels/base.ts
is the base class for a channel and it can be extended to create any kind of channel. For example,last_value.ts
,binop.ts
are all types of channels.- In Pregel, there is no limitation on the number of channels a node can subscribe to or write to. In LangGraph, however, currently every node maps to two channels. (1) A channel's value that it is subscribes to, i.e - is dependent on. (2) The channel that it writes to.
src/pregel/index.ts
holds all the business logic that uses channels and nodes in a pregel graph.async *_transform
holds some of the most important logic because it is responsible for updating the channel's value and updating the checkpoint accordingly.