DocPaws
A self-hosted RAG document assistant with per-user knowledge bases, PDF indexing, source citations, and a configurable distance threshold that makes it decline out-of-scope questions.
DocPaws is a retrieval-augmented-generation assistant for personal document knowledge bases, with a FastAPI and SQLModel backend and a Vue 3 and Vite frontend. Users upload PDFs into folders, which are indexed incrementally, and can then ask questions scoped to the whole library, a folder, or a single file, receiving streamed answers with source citations and a view of the model's reasoning.
When a question falls outside the indexed material, or the closest match does not clear a configurable retrieval-distance threshold, DocPaws declines to answer rather than guessing, to reduce hallucination. It runs behind Docker Compose services for PostgreSQL, Redis, and MinIO object storage, uses Celery for background PDF indexing, and can connect to DeepSeek, an OpenAI-compatible endpoint, or SiliconFlow for its language and embedding models. A reproducible evaluation set is included.