A LangChain playground using TypeScript

A playground for LangChain.js, LangGraph, Slack, Model Context Protocol (MCP) and other LLM-related tools.

This project provides both REST API endpoints or Slack bot integration for interacting with different language models and LangChain and LangGraph workflows.

Architecture

Core components

• langchain.js: Framework for building applications with LLMs.
• langgraph: Framework for building applications with advanced workflow orchestration for multi-step processes.
• slack/bolt: Integration with Slack for building Slack apps.
• Model Context Protocol (MCP): MCP is a protocol for building LLM-powered tools.

LLM providers

• openai
• ollama
• groq

Document Loaders

• DirectoryLoader: Loads documents from a directory via Unstructured API.
• ConfluenceLoader: Loads documents from Confluence.
• RecursiveCharacterTextSplitter: Splits documents into chunks of text.
• GitHubLoader: TODO.

Services

• ollama: Ollama enables the execution of LLM models locally.
• openweb-ui: OpenWeb UI is a self-hosted WebUI that interacts with Ollama.
• unstructured-api: The Unstructured API is designed to ingest/digest files of various types and sizes.
• qdrant: Qdrant serves as a vector database.
• chroma: Chroma serves as an embedding database. Not used anymore.
• redis: Redis is an open-source in-memory data structure store.

Server mode

• fastify: serves as a web server in src/api
• slack: serves as a Slack app in src/slack

Sentry log analysis

In this project, I used LangGraph to build a workflow to analyze Sentry logs.

The workflow in big picture is as follows:

1. Get Sentry issue and first event
2. Normalize the issue and event and extend the stacktrace to source code fetching from GitHub
3. Generate a summary of the investigation using the normalized issue and event

New Relic log analysis

In this project, I used LangGraph to build a workflow to analyze New Relic logs.

The workflow in big picture is as follows:

1. Get New Relic logs
2. Analyze New Relic logs to get the request timeline, service error logs and relevant URLs
3. Generate a summary of the investigation by analyzing the request timeline, service error logs and relevant URLs

Answer from Retriever-Augmented Generation (RAG)

In this project, there are following routes to answer user's question from the document RAG retrieval.

Routes:

• DELETE /document/reset: Reset the document RAG retrieval.
• PUT /document/load/directory: Load documents from a directory using Unstructured API + Parent document retriever.
• PUT /document/load/confluence: Load documents from Confluence + Parent document retriever.
• POST /document/query: Answer user's question from the document RAG retrieval.

Document loader process

Document query process

AWS Bedrock Knowledge Base

Parent document retriever

Slack integration

In this project, I used slack/bolt and LangGraph to build a Slack app.

• When a user mentions the bot in a channel, the bot will respond with a message.
• It will execute the following steps:
  • Intent classifier: Classify the intent of the user's message.
  • Intent router: Route the user's message to the appropriate node.
  • Get message history: Get the message history of the channel.
  • MCP tools: Use MCP tools to get information from Model Context Protocol.
  • Summarise thread: Summarise the thread.
  • Translate message: Translate the message to the user's language.
  • Find information: Find information from the RAG database.
  • General response: Generate a general response.
  • Final response: Respond to the user's message.

How to start

docker-compose up -d --build

Endpoints

TBD

Todo

• Add more examples
• Add tests
• Make better documentations