Langchain agent support

[leobenkel]

Can I connect this UI to Langchain agent and tooling ?

Thank you

[MikeHago]

That would be a very cool feature :-)

[mynumu]

Me too! I want this…or an option to bring your own langchain code. Especially to change the RAG implementation

[alex-roc]

Hi, I have the same question as OP. Maybe we can connect it with LangServe/FastAPI?

[gradimirs]

Looking for this feature as well!

[ShaunBoughey]

Would be an amazing feature to see, would allow custom RAG implimentations to still have a nice UI!

[tcztzy]

Is there any idea to implement this feature? Is namespace package as plugin system a good idea?

[tcztzy]

I used dify recently, maybe we can implement it like dify did?

[tcztzy]

I have two implementation plans:

1. namespace package plugin system
2. dify-like implementation

Plan 1: We needs load third party with open-webui.(apps|agents|tools) at main.py start server. The third party package can be upload to pypi.org or just on github (pip can install github dependencies). And then we can treat the agent as a normal model or add tools to chat like ChatGPT's plugin.

Plan 2: We can add an openapi app as tool, and then design a page to create agent. All agent will store in database instead of code compare with plan 1.

We can implement both of them.

[aouicher]

ok for me

[jlcoo]

Me too! I want this…or an option to bring your own langchain code. Especially to change the RAG implementation

[jaaacki]

me too!!!!

[HorstA]

Me too! That would be almost vital for me ;-)

[PlebeiusGaragicus]

Came here also looking for this.

I'd like to be able to serve my LangGraph agents via LangServe and use OpenWebUI to call that endpoint. I just started looking into this.

Also, if I could get per-user usage paid via micropayments this would become a my perfect solution!

[dannongruver]

Feels like a perfect solution. Langgraph also.

Could be the middleware dev env for LLM app development. A dev builds this middle tier (RAG+Tools+Lang chain/DSPy) and tests via webUI frontend…and later have an OpenWebUI API to reuse this tested middleware for downstream apps.

This is what I’m struggling with. I want to develop an LLM app and quickly test ideas for multi-agent implementation. Only option rn is python dev/test but no UI interface. So I have to build this type of testing ground UI.

But I guess that might change the name Open “WebUI” 😅 to Open “API + WebUI”

[PlebeiusGaragicus]

I have figured out a solution for this using Pipelines. Just build a pipeline that calls your LangChain/Graph agent. Instead of using LangServe, I just build an endpoint with FastAPI that invokes my graph. Then, you can also host your own vector database and whatever else your agent will need.

[solargoe]

Have you published it for the general audience? That would be nice

[PlebeiusGaragicus]

Refer to this PR: https://github.com/open-webui/pipelines/pull/158/files

It's close enough to my solution.

Although, now that I'm using LangGraph Studio I'm going to have to rework this. It seems LangGraph is moving towards the "assistants API" and I'm struggling to make it compatible with open webui at the moment. Instead of calling an .invoke() or api/chat endpoint, you use "threads" and "runs." There's probably an "orchestration layer" that Open WebUI needs in order to make it work well.

[hasani114]

I tried this because it seemed pretty simple. Wasn't sure there was anything else to it? It works...

`"""
title: Langchain Pipeline
author: open-webui
date: 2024-05-30
version: 1.0
license: MIT
description: A pipeline for creating and managing conversational agents using the Langchain library.
requirements: langchain-core, langchain-openai
"""

from typing import List, Union, Generator, Iterator
from schemas import OpenAIChatMessage
from pydantic import BaseModel
from langchain_openai import ChatOpenAI

class Pipeline:
class Valves(BaseModel):
CHAT_MODEL: str = ""

def __init__(self):
    # Optionally, you can set the id and name of the pipeline.
    # Best practice is to not specify the id so that it can be automatically inferred from the filename, so that users can install multiple versions of the same pipeline.
    # The identifier must be unique across all pipelines.
    # The identifier must be an alphanumeric string that can include underscores or hyphens. It cannot contain spaces, special characters, slashes, or backslashes.
    # self.id = "pipeline_example"

    # The name of the pipeline.
    self.name = "Langchain Pipeline"
    self.valves = self.Valves(**{"CHAT_MODEL": "gpt-4o"})

async def on_startup(self):
    # This function is called when the server is started.
    print(f"on_startup:{__name__}")
    print(f"Model: {self.valves.CHAT_MODEL}")
    print(f"Valves: {self.valves}")
    pass

async def on_shutdown(self):
    # This function is called when the server is stopped.
    print(f"on_shutdown:{__name__}")
    pass

async def on_valves_updated(self):
    # This function is called when the valves are updated.
    pass

async def inlet(self, body: dict, user: dict) -> dict:
    # This function is called before the OpenAI API request is made. You can modify the form data before it is sent to the OpenAI API.
    print(f"inlet:{__name__}")

    print(body)
    print(user)

    return body

async def outlet(self, body: dict, user: dict) -> dict:
    # This function is called after the OpenAI API response is completed. You can modify the messages after they are received from the OpenAI API.
    print(f"outlet:{__name__}")

    print(body)
    print(user)

    return body

def pipe(
    self, user_message: str, model_id: str, messages: List[dict], body: dict
) -> Union[str, Generator, Iterator]:
    # This is where you can add your custom pipelines like RAG.
    print(f"pipe:{__name__}")

    check_weather = {
        "name": "check_weather",
        "description": "Check the current weather conditions for a given location.",
        "parameters": {
            "type": "object",
            "properties": {
                "location": {
                    "type": "string",
                    "description": "The city or location to check weather for",
                },
                "units": {
                    "type": "string",
                    "enum": ["metric", "imperial"],
                    "description": "The unit system for temperature (metric: Celsius, imperial: Fahrenheit)",
                },
            },
            "required": ["location", "units"],
        },
    }
    # If you'd like to check for title generation, you can add the following check
    if body.get("title", False):
        print("Title Generation Request")

    model = ChatOpenAI(model=self.valves.CHAT_MODEL)
    model_with_tools = model.bind_tools([check_weather])
    print(model_with_tools.to_json())
    response = model_with_tools.stream(messages)

    # Yield each chunk from the response
    for chunk in response:
        yield chunk.content`

[tuniverspinner]

@hasani114 how to use this? sorry noob question

[PlebeiusGaragicus]

The above is a "pipeline" which you can use with OpenWebUI. Just read the docs about how to get pipelines working, then you go into admin settings > pipelines and then import this module. It should show up as a new model to talk to.