Kollektiv is a Retrieval-Augmented Generation (RAG) system designed for one purpose - allow you to chat with your favorite docs (of libraries, frameworks, tools primarily) easily.
This project aims to allow LLMs to tap into the most up-to-date knowledge in 2 clicks so that you don't have to worry about incorrect replies, hallucinations or inaccuracies when working with the best LLMs.
This project was born out of a personal itch - whenever a new feature of my favorite library comes up, I know I can't rely on the LLM to help me build with it - because it simply doesn't know about it!
The root cause - LLMs lack access to the most recent documentation or private knowledge, as they are trained on a set of data that was accumulated way back (sometimes more than a year ago).
The impact - hallucinations in answers, inaccurate, incorrect or outdated information, which directly decreases productivity and usefulness of using LLMs
But there is a better way...
What if LLMs could tap into a source of up-to-date information on libraries, tools, frameworks you are building with?
Imagine your LLM could intelligently decide when it needs to check the documentation source and always provide an accurate reply?
Meet Kollektiv -> an open-source RAG app that helps you easily:
- parse the docs of your favorite libraries
- efficiently stores and embeds them in a local vector storage
- sets up an LLM chat which you can rely on
Note this is v.0.1.* and reliability of the system can be characterized as following:
- in 50% of the times it works every time!
So do let me know if you are experiencing issues and I'll try to fix them.
- 🕷️ Intelligent Web Crawling: Utilizes FireCrawl API to efficiently crawl and extract content from specified documentation websites.
- 🧠 Advanced Document Processing: Implements custom chunking strategies to optimize document storage and retrieval.
- 🔍 Vector Search: Employs Chroma DB for high-performance similarity search of document chunks.
- 🔄 Multi-Query Expansion: Enhances search accuracy by generating multiple relevant queries for each user input.
- 📊 Smart Re-ranking: Utilizes Cohere's re-ranking API to improve relevancy of search results
- 🤖 AI-Powered Responses: Integrates with Claude 3.5 Sonnet to generate human-like, context-aware responses.
- 🧠 Dynamic system prompt: Automatically summarizes the embedded documentation to improve RAG decision-making.
- Language: Python 3.7+
- Web Crawling: FireCrawl API
- Vector Database: Chroma DB
- Embeddings: OpenAI's text-embedding-3-small
- LLM: Anthropic's Claude 3.5 Sonnet
- Re-ranking: Cohere API
- Additional Libraries: tiktoken, chromadb, anthropic, cohere
-
Clone the repository:
git clone https://github.com/Twist333d/kollektiv.git cd rag-docs -
Set up environment variables: Create a
.envfile in the project root with the following:FIRECRAWL_API_KEY="your_firecrawl_api_key" OPENAI_API_KEY="your_openai_api_key" ANTHROPIC_API_KEY="your_anthropic_api_key" COHERE_API_KEY="your_cohere_api_key" -
Install dependencies:
poetry install -
Run the application:
poetry run python app.py
or through a poetry alias:
python app.py
-
Crawl Documentation:
Update the root url you want to parse. For example:
urls_to_crawl = ["https://docs.anthropic.com/en/docs/"]
Ensure you include the url patterns of sub-pages you want to parse and exclude url patterns of sub-pages you don't want to parse:
"includePaths": ["/tutorials/*", "/how-tos/*", "/concepts/*"], "excludePaths": ["/community/*"],
Set the maximum number of pages you want to crawl:
crawler.async_crawl_url(urls_to_crawl, page_limit=250)
-
Chunk FireCrawl parsed docs:
Next step is to chunk all the parsed documents
poetry run python -m src.processing.chunking
-
Configure basic parameters:
Set up the following parameters in the
app.py:- Whether to load only specified or all processed chunks in
PROCESSED_DATA_DIR
docs = ["docs_anthropic_com_en_20240928_135426-chunked.json"] initializer = ComponentInitializer(reset_db=reset_db, load_all_docs=True, files=[])
- Whether to reset the database, which will clear all the data in local ChromaDB - use with caution. Defaults to false.
if __name__ == "__main__": main(debug=False, reset_db=False)
- Whether to load only specified or all processed chunks in
-
Chat with documentation: You can run application via the following command:
python app.py
- Only terminal UI (no Chainlit for now)
- Image data not supported - ONLY text-based embeddings.
- No automatic re-indexing of documents
- Basic chat flow supported
- Either RAG tool is used or not
- if a tool is used -> retrieves up to 5 most relevant documents (after re-ranking)
- Either RAG tool is used or not
For a brief roadmap please check out project wiki page.
Evaluation is currently done using ragas library. There are 2 key parts assessed:
- End-to-end generation
- Faithfulness
- Answer relevancy
- Answer correctness
- Retriever (TBD)
- Context recall
- Context precision
Kollektiv is licensed under a modified version of the Apache License 2.0. While it allows for free use, modification, and distribution for non-commercial purposes, any commercial use requires explicit permission from the copyright owner.
- For non-commercial use: You are free to use, modify, and distribute this software under the terms of the Apache License 2.0.
- For commercial use: Please contact [email protected] to obtain a commercial license.
See the LICENSE file for the full license text and additional conditions.
The project has been renamed from OmniClaude to Kollektiv to:
- avoid confusion / unintended copyright infringement of Anthropic
- emphasize the goal to become a tool to enhance collaboration through simplifying access to knowledge
- overall cool name (isn't it?)
If you have any questions regarding the renaming, feel free to reach out.
- FireCrawl for superb web crawling
- Chroma DB for easy vector storage and retrieval
- Anthropic for Claude 3.5 Sonnet
- OpenAI for text embeddings
- Cohere for re-ranking capabilities
For any questions or issues, please open an issue
Built with ❤️ by AZ