chainml-copilot
An AI agent that is trained on documentation of projects and is then able to answer user questions related to that project.
Project Description
The project consists of (1) training data which is multiple PDF representations of different web3 projects' docs;
(2) a model that represents the AI agent;
(3) a server that wraps the model and exposes an API to interact with it via JSON requests;
(4) an XMTP service that receives messages from users and then relays them to the model server; and
(5) the frontend which in this case is the live XMTP chatbot interface but could be replaced by Telegram, Discord, Slack, WhatsApp, or any custom frontend.
How it's Made
By component: (1) i scraped the docs subdomains of the projects to download the html (and assets) of each page on the docs website. I then flattened the html files from a deeply nested directory to one flat one. I then converted these html files to PDFs and finally mered all the PDFs into a single file. Then I repeated this for each web3 project. I had planned to just use the .md files on their github but there was no infallible way to convert multiple .md files with references to images and other files into a single PDF.
(2) this is written in Python3 and uses APIs from OpenAI, Google News and Google Search. The original model is open sourced by ChainML.
(3) the server is Flask (Python)
(4) XMTP service is written in Javascript.
