to start:

docker compose up

access web ui:

http://127.0.0.1:8000/

how it works:

PDF Upload & Processing

1. PDF Upload (FastAPI /upload endpoint)

   • File validation and UUID generation
   • Save file and store metadata in MongoDB
   • Queue background processing task

2. PDF Processing Pipeline

   • Extract text, images, and tables using PyMuPDF, Camelot, PDFPlumber
   • Generate markdown with image placeholders
   • Run OCR on images using PaddleOCR
   • Clean and normalize markdown content
   • Process diagrams using LLaVA vision model
   • Split content into chunks and create vector documents
   • Generate document summary using Ollama LLM
   • Store documents in ChromaDB vector store

Vector Storage & Indexing

3. Vector Store (ChromaDB)

   • Persistent storage with document embeddings
   • Metadata tracking (file_id, chunk_index, source_file)
   • Embeddings generated by nomic-embed-text model

4. RAG Pipeline (LlamaIndex)

   • Create index from ChromaDB documents
   • Configure with Ollama LLM and embeddings
   • Set up query engine for similarity search

Question Answering

5. Question Processing (FastAPI /ask endpoint)
   • Receive question and query vector store
   • Retrieve top-k similar chunks
   • Generate answer using Ollama LLM with retrieved context
   • Return structured response with sources and confidence score

Supporting Services

6. Core Services
   • Ollama Service: LLM (llama3.2:3b), embeddings (nomic-embed-text), vision (llava:7b)
   • MongoDB: File metadata and status tracking
   • OCR Service: PaddleOCR for image text extraction
   • Markdown Cleaner: Content normalization

Data Flow: PDF -> Text/Image Extraction -> Markdown -> OCR -> Cleaning -> Chunking -> Vector Store -> Question -> Similarity Search -> LLM Response