ragdaemon

Ragdaemon

Ragdaemon is a Retrieval-Augmented Generation (RAG) system for code. It runs a daemon (background process) to watch your active code, put it in a knowledge graph, and query the knowledge graph to (among other things) generate context for LLM completions.

Three ways to use Ragdaemon:

1. Help me write code

Ragdaemon powers the 'auto-context' feature in Mentat, a command-line coding assistant. You can install Mentat using pip install mentat. Run with the --auto-context-tokens <amount> or -a (default=5000) flag, and ragdaemon-selected context will be added to all of your prompts.

2. Explore the knowledge graph

Install locally to visualize and query the knowledge graph directly. Install using pip install ragdaemon, and run in your codebase's directory, e.g. ragdaemon. This will start a Daemon on your codebase, and an interface at localhost:5001. Options:

• --chunk-extensions <ext>[..<ext>]: Which file extensions to chunk. If not specified, defaults to the top 20 most common code file extensions.
• --chunk-model: OpenAI's gpt-4-0215-preview by default.
• --embeddings-model: OpenAI's text-embedding-3-large by default.
• --diff: A git diff to include in the knowledge graph. By default, the active diff (if any) is included with each code feature.

3. Use ragdaemon Python API

Ragdaemon is released open-source as a standalone RAG system. It includes a library of python classes to generate and query the knowledge graph. The graph itself is a NetworkX MultiDiGraph which saves/loads to a .json file.

import asyncio
from pathlib import Path
from ragdaemon.daemon import Daemon

async def main():
    cwd = Path.cwd()
    daemon = Daemon(cwd)
    await daemon.update()

    results = daemon.search("javascript")
    for result in results:
        print(f"{result['distance']} | {result['id']}")

    query = "How do I run the tests?"
    context_builder = daemon.get_context(
        query, 
        auto_tokens=5000
    )
    context = context_builder.render()
    messages = [
        {"role": "user", "content": query},
        {"role": "user", "content": f"CODE CONTEXT\n{context}"}
    ]
    print(messages)

asyncio.run(main())