Skip to content

ARCHITECTURE.md

Latest commit

 

History

History
93 lines (76 loc) · 3.33 KB

ARCHITECTURE.md

File metadata and controls

93 lines (76 loc) · 3.33 KB

🏗️ System Architecture

Overview

The Enterprise Document Analyzer operates as a monolithic application with a microservices-style logical separation between the Frontend (Streamlit) and Backend (FastAPI). It uses a "Hybrid RAG" approach, leveraging both local vector stores (ChromaDB) and Metadata stores (FileSystem/JSON) to manage complex document artifacts like images and tables.

🧩 High-Level Components

1. Frontend (frontend/app.py)

  • Technology: Streamlit
  • Role: User Interface and State Management.
  • Responsibilities:
    • File Upload handling.
    • Chat State (Session State).
    • PDF Rendering (via base64 embedding).
    • API Communication (via requests).

2. Backend (backend/main.py)

  • Technology: FastAPI
  • Role: REST API and Orchestrator.
  • Endpoints:
    • POST /upload: Entry point for the ingestion pipeline.
    • GET /search: Semantic search and LLM synthesis.
    • GET /database/inspect: Database transparency.
    • DELETE /...: Lifecycle management.

3. Ingestion Pipeline (backend/pipeline.py)

The heart of the system. It processes raw PDFs into structured knowledge.

Steps:

  1. Ingest: Load PDF using PyMuPDF (fitz).
  2. Classify: Detect if the document is Digital, Scanned, or Mixed.
  3. Parsers:
    • DoclingParser: Advanced layout analysis for tables and text.
    • VisionProcessor: Extracts images/figures.
    • GeminiProcessor: Fallback for complex visual documents (Handwriting).
  4. Chunking: Splits text into 500-1000 character overlapping windows.
  5. Indexing:
    • Text -> Embedding -> ChromaDB.
    • Images/Tables -> JSON Metadata Store.

4. Storage Layers

  • Vector Store (backend/custom_storage/vector.py): Uses ChromaDB (Persistent) to store embeddings for fast semantic retrieval.
  • Metadata Store (backend/custom_storage/metadata.py): JSON-based storage for linking chunks to their original page number, bounding box (bbox), and parent document.
  • Static Assets (data/static/):
    • pdfs/: Original uploaded files.
    • images/: Extracted figures and crops.

🔄 Data Flow

sequenceDiagram
    participant User
    participant UI as Streamlit Frontend
    participant API as FastAPI Backend
    participant Pipe as Pipeline
    participant DB as ChromaDB/FS

    User->>UI: Upload Document (PDF)
    UI->>API: POST /upload (File + Mode)
    API->>Pipe: Run Pipeline
    
    rect rgb(20, 20, 20)
        Note right of Pipe: Processing
        Pipe->>Pipe: Classify Document
        alt OCR Mode
            Pipe->>Pipe: Docling Extraction
        else Gemini Mode
            Pipe->>Pipe: Gemini Vision API
        end
        Pipe->>Pipe: Chunk Text
        Pipe->>DB: Store Vectors & Metadata
    end
    
    Pipe-->>API: Success (Doc ID)
    API-->>UI: Ready
    
    User->>UI: Ask Question
    UI->>API: GET /search?q=...
    API->>DB: Query Vectors
    DB-->>API: Top Chunks
    API->>API: Gemini Synthesis
    API-->>UI: Answer + Citations
Loading

📂 Directory Structure Details

  • backend/modules/vision.py: Logic for extracting images from PDFs.
  • backend/modules/ingestion.py: Helper class for initial PDF loading and classification.
  • backend/custom_storage/: Custom wrappers around database technologies to allow easy swapping (e.g., to Postgres/Supabase in future).