ganymedenil/document.ai

This project is a local knowledge base system that lets you ask questions against your own documents and get answers generated by GPT-3.5. It is written in Python and the README is primarily in Chinese, though the core idea is straightforward: load a set of question-and-answer pairs, convert them into a numerical format called vectors, store those vectors in a database, and then match new incoming questions to the closest stored answers before using GPT to compose a polished reply. The flow works like this. You start with a collection of documents or Q&A data. The system converts each piece of text into a vector, which is a list of numbers that captures the meaning of the text. Those vectors go into a vector database. When a user asks a question, the question is also converted into a vector and the database returns the top few most similar stored answers. GPT then takes those retrieved answers and shapes them into a coherent response. The author includes a medical Q&A demo as one example domain, but notes the same approach can be applied to any field, such as customer support or internal documentation. The code directory contains example scripts and the docs directory holds the author's notes and diagrams explaining the approach. The README also discusses limitations: queries can be imprecise if the phrasing is vague, the default OpenAI embedding model may not perform well on highly specialized topics, and fine-tuning GPT on domain-specific data improves accuracy but is expensive to do frequently. The author trained several custom Chinese-language text embedding models and shared them publicly on Hugging Face as an alternative to the default OpenAI embeddings.