biao994/docpaws

DocPaws is a document question-and-answer assistant that lets users upload PDF files, build a searchable knowledge base from them, and then ask questions in a chat interface. The AI finds relevant passages from the uploaded documents and cites them in its responses. If no relevant content is found above a configurable similarity threshold, the system declines to answer rather than guessing, which the project describes as "retrieval distance threshold rejection." The backend is written in Python using FastAPI, a web framework for building APIs, and stores data in SQLite with FAISS handling the vector index (a structure that allows fast similarity search over document chunks). The frontend is a Vue 3 single-page application. Users can organize documents into folders and scope a question to the full knowledge base, a specific folder, or a single file. Responses stream in real time, and a "thinking process" display is available during generation. The project connects to external AI services for both the language model and the embedding model (the part that converts text into searchable vectors). Supported providers include DeepSeek, SiliconFlow, and OpenAI-compatible APIs. API keys are stored in a local environment file that the README explicitly warns should not be committed to version control. Indexing uploaded PDFs can run synchronously or through a background task queue called Celery, which requires Redis. The Celery path is optional: without it, indexing runs inline. The README provides separate startup instructions for Windows and macOS/Linux, noting that Windows has a compatibility restriction for the Celery worker pool. The repository includes an evaluation suite called "Golden 20" with 20 test questions and reproducible scoring, intended for regression testing when the retrieval or model configuration changes. The backend also has over 79 automated test cases. The project is released under the MIT license.