PDF-GPT

What PDF-GPT is

PDF-GPT is an open-source application for asking questions about PDF contents with GPT functionality. Users provide a document by uploading it or entering a PDF URL. The app divides the document into smaller sections, creates embeddings with a Deep Averaging Network Encoder, and searches those embeddings semantically. Its documented retrieval flow selects the top five chunks with KNN before sending relevant material to OpenAI for answer generation. Answers can include the source PDF page number in square brackets. The project describes a local Gradio playground at localhost:7860, an API exposed through langchain-serve with ask_url and ask_file endpoints, Docker Compose deployment, and Jina Cloud deployment. Sample requests pass an OPENAI_API_KEY through environment variables. The Universal Sentence Encoder download is 915 MB when fetched at runtime. PDF-GPT is MIT licensed, and its maintainer seeks voluntary contributors for backlog work.