This project provides a visualization and debugging tool for LangChain workflows, targeting developers who need to understand and troubleshoot their LLM applications. It offers a richer, more interactive debugging experience compared to LangChain's built-in tracer by leveraging the Ought ICE visualizer.

How It Works

The tool intercepts and visualizes LangChain interactions, presenting them in a user-friendly UI. Key features include highlighting hardcoded versus templated prompt sections, displaying execution flow, and showing LLM call costs (for specific OpenAI models). It adapts the Ought ICE visualizer, offering a distinct UI preference for users who favor its approach to inspecting LLM calls and tool execution.

Quick Start & Requirements

• Install: pip install langchain-visualizer
• Prerequisites: libyaml-dev (on Linux), google-search-results, openai. For replaying recorded interactions: vcr-langchain.
• Usage: Import langchain_visualizer and call langchain_visualizer.visualize() on an async function. Jupyter notebooks require langchain_visualizer.jupyter.
• Example: https://github.com/amosjyng/langchain-visualizer/blob/main/tests/agents/test_langchain_getting_started.py

Highlighted Details

• Visualizes prompt text, distinguishing hardcoded and templated parts.
• Displays execution flow and tool calls (e.g., PythonREPL).
• Shows costs for individual LLM calls and total run cost (OpenAI text-davinci-003).
• Supports visualization of document chunking for embeddings via visualize_embeddings.

Maintenance & Community

• Developed by amosjyng.
• Related project: VCR LangChain for testing and demos.
• Users are encouraged to open issues or PRs for missing visualization features.

Licensing & Compatibility

• License: Not explicitly stated in the README.
• Compatibility: Designed for use with LangChain.

Limitations & Caveats

The project is a personal adaptation and may not cover all LangChain functionalities. Users are encouraged to contribute missing features. The README notes that LangChain's built-in tracer is "definitely better supported."