Doxie is an experiment in LLM-based information retrieval. Or in plain words: Doxie uses ChatGPT to answer questions based on the content of one or more websites, documents, or whatever other text content you want to feed it. It exposes a simple, ChatGPT like interface.
Doxie implements the most basic RAG loop imaginable, with simple context management. Have a look at these source files:
- Srapping and preparing data from a specific source src/server/berufslexikon.ts
- Embedding data
- src/server/embedder.ts chunking, embedding
- src/server/embedder-cli.ts takes the output of the scrapper and converts it into a
xxx.embeddings.binfile to be loaded into the vector database
- Vector database, loads a file generated by
embedder-cli.tsinto a collection and lets you query it src/server/rag.ts - Chat session management, keeps track of chat sessions, expands queries for RAG, and submits user queries + RAG context to GPT 3.5-turbo to get hopefully meaningful answers with source citations. [src/server/chatsessions.ts]
To run Doxie you need the following software installed on your system:
Doxie currently does not have a proper ingestion pipeline, and is hard coded for the test use case. This will probably, maybe amended in the coming days/weeks, depending on my mood.
If you want to play around with Doxie on some real-world data, run the ./download-testdata.sh script in the root folder. It will download a data set with embeddings generated from AMS Berufslexikon to ./docker/data/berufslexikon.embeddings.bin. You can then follow the instructions in the Development section below.
npm run dev
In VS Code run the dev launch configurations, which will attach the debugger to the server, spawn a browser window, and also attach to that for frontend debugging.
- Deploy backend & frontend:
./publish.sh server - Deploy just the frontend:
./publish.sh