AskPDF is a Retrieval-Augmented Generation (RAG) system designed to process large documents (PDFs) and provide accurate, context-aware answers to user queries. By leveraging vector embeddings and Large Language Models (LLMs), it allows users to interact with their document repository naturally.
The system is composed of three main microservices:
- Client (SvelteKit): A responsive frontend for document management and chat interface.
- Server (FastAPI): The core API handling user requests, database interactions, and LLM communication.
- PDF Processor (Python/Celery): An asynchronous worker responsible for parsing PDFs, generating embeddings, and updating the vector store.
When a user uploads a file, it goes through a secure signed URL process directly to S3, triggering an event-driven processing pipeline.
User queries are processed to retrieve relevant document contexts from ChromaDB, which are then fed into the Gemini LLM to generate precise answers with citations.
- Document Ingestion: Secure upload and asynchronous processing of PDF documents.
- Semantic Search: Uses advanced embeddings (Sentence-Transformers) to find the most relevant document sections.
- Contextual Answers: Generates answers using Gemini LLM based strictly on the provided context.
- Citations: Every answer includes citations pointing back to the specific source document and text.
- Conversation History: Maintains chat sessions for continuous context (partially implemented).
- Frontend: SvelteKit, TypeScript, TailwindCSS
- Backend: FastAPI, Python, SQLAlchemy
- AI/ML: Sentence-Transformers, Gemini API, ChromaDB (Vector Store)
- Database: PostgreSQL, In-memory SQLite
- Async Processing: Celery, RabbitMQ/SQS
- Infrastructure: AWS Lambda, S3, Docker