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:
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.
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'sgpt-4-0215-previewby default.--embeddings-model: OpenAI'stext-embedding-3-largeby default.--diff: A git diff to include in the knowledge graph. By default, the active diff (if any) is included with each code feature.
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())
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent