AskDocs — Intelligent Document Query System

RAG-based IDP · FastAPI · React · ChromaDB · Google Gemini
6th Semester — Advanced Web Technology (AWT)

---

What is AskDocs?

AskDocs is a privacy-first Intelligent Document Processing (IDP) system built on Retrieval-Augmented Generation (RAG). Upload any document — PDF, Word, Excel, or image — and ask natural-language questions about its content. The system retrieves the most relevant passages and generates a concise, cited answer using Google Gemini.

All document data stays on your machine. Only the retrieved text snippets are sent to the Gemini API for answer generation.

---

Features

Feature	Details
Multi-format ingestion	PDF, DOCX, XLSX/XLS, PNG, JPG, JPEG
OCR fallback	Scanned PDFs and images processed via Gemini Vision
Semantic search	ChromaDB vector store with cosine-similarity HNSW index
Grounded answers	Gemini 2.5 Flash — answers cite source document + page
Sliding-window chunking	500-word chunks, 100-word overlap
Batch embedding	Gemini text-embedding-004 (768-dimensional)
Document management	List, re-upload, and delete documents with full vector cleanup
Modern UI	React + Vite + Tailwind CSS with drag-and-drop upload

---

Architecture

┌─────────────────────────────────────────────────────────────┐
│                     React Frontend                          │
│   FileUpload │ DocumentList │ ChatInterface │ CitationPanel │
└──────────────────────┬──────────────────────────────────────┘
                       │  HTTP / REST  (localhost:8000)
┌──────────────────────▼──────────────────────────────────────┐
│                    FastAPI Backend                          │
│   POST /api/upload   │   POST /api/query   │   GET /api/documents │
└──┬─────────────┬─────┴──────────────┬──────────────────┬────┘
   │             │                    │                  │
┌──▼──┐     ┌───▼───┐           ┌────▼────┐        ┌────▼────┐
│Extr-│     │Chunker│           │Embedder │        │   LLM   │
│actor│     │500w/  │           │Gemini   │        │Gemini   │
│PyMu-│     │100ovlp│           │text-004 │        │2.5Flash │
│PDF  │     └───────┘           └────┬────┘        └─────────┘
│docx │                              │
│Exce-│                        ┌─────▼──────┐
│l/OCR│                        │  ChromaDB  │
└─────┘                        │ (local fs) │
                               └────────────┘

---

Prerequisites

Requirement	Version
Python	3.10+
Node.js	18.x+
Google Gemini API Key	Valid key from Google AI Studio

---

Quick Start

1. Clone / navigate to the project

cd "AWT Project"

2. Backend setup

cd backend
python -m venv venv

# Windows
venv\Scripts\activate

# macOS / Linux
source venv/bin/activate

pip install -r requirements.txt

Create backend/.env:

GEMINI_API_KEY=your_api_key_here

Start the backend:

uvicorn main:app --reload --port 8000

3. Frontend setup

cd frontend
npm install
npm run dev

4. Open in browser

Navigate to http://localhost:5173

One-command start (Windows)

start.bat

---

Usage

1. Upload a document — drag and drop a file into the upload panel or click to browse
2. Wait for processing — the progress bar shows extraction → chunking → embedding status
3. Ask a question — type any natural-language question in the chat box
4. Read the answer — view the Markdown-formatted answer with inline citations
5. Check sources — the Citation Panel shows which document and page each answer came from
6. Manage documents — use the Document List to view or delete uploaded files

---

API Reference

Method	Endpoint	Description
POST	/api/upload	Upload and process a document
GET	/api/documents	List all stored documents
DELETE	/api/documents/{name}	Delete document and all its vectors
POST	/api/query	Ask a question; returns answer + citations
GET	/health	Health check

POST /api/upload

Request: multipart/form-data with file field

Response:

{
  "message": "Document processed successfully.",
  "document_name": "report.pdf",
  "chunks_created": 42
}

POST /api/query

Request:

{
  "question": "What are the key findings?",
  "top_k": 5
}

Response:

{
  "answer": "The key findings are... *(Source: report.pdf, Page 3)*",
  "citations": [
    {
      "document_name": "report.pdf",
      "page_number": 3,
      "chunk_text": "The study found that..."
    }
  ]
}

---

Project Structure

AWT Project/
├── backend/
│   ├── main.py                  # FastAPI app, CORS, router registration
│   ├── requirements.txt
│   ├── .env                     # GEMINI_API_KEY (not committed)
│   ├── routers/
│   │   ├── upload.py            # POST /api/upload, GET/DELETE /api/documents
│   │   └── query.py             # POST /api/query
│   ├── services/
│   │   ├── extractor.py         # Multi-format text extraction + OCR
│   │   ├── chunker.py           # Sliding-window word chunking
│   │   ├── embedder.py          # Gemini text-embedding-004 calls
│   │   ├── vector_store.py      # ChromaDB CRUD operations
│   │   └── llm.py               # Gemini 2.5 Flash answer generation
│   ├── models/
│   │   └── schemas.py           # Pydantic request/response schemas
│   ├── uploads/                 # Uploaded files (local, not committed)
│   └── chroma_db/               # ChromaDB persistent storage (local)
├── frontend/
│   └── src/
│       ├── App.jsx              # Root layout: sidebar + chat
│       ├── components/
│       │   ├── FileUpload.jsx   # Drag-and-drop upload with progress
│       │   ├── DocumentList.jsx # Uploaded documents with delete
│       │   ├── ChatInterface.jsx# Conversation UI
│       │   └── CitationPanel.jsx# Source document citations
│       └── api/
│           └── api.js           # Axios API client
├── docs/
│   └── SRS.docx                 # Software Requirements Specification
├── generate_srs.py              # Script to regenerate SRS.docx
├── start.bat                    # One-click startup (Windows)
└── README.md

---

Technology Stack

Layer	Technology	Purpose
Frontend	React 18 + Vite	SPA with fast HMR
Styling	Tailwind CSS	Utility-first CSS
File Upload	React Dropzone	Drag-and-drop with validation
HTTP Client	Axios	REST API calls
Backend	FastAPI + Uvicorn	Async Python REST API
Validation	Pydantic v2	Schema validation
LLM	Gemini 2.5 Flash	Answer generation
Embeddings	Gemini text-embedding-004	768-dim semantic vectors
OCR	Gemini Vision (2.5 Flash)	Scanned document/image text
Vector DB	ChromaDB	Local HNSW similarity search
PDF	PyMuPDF	Text extraction
Word	python-docx	DOCX parsing
Excel	Pandas + openpyxl	Spreadsheet parsing
Images	Pillow	Image preprocessing

---

How RAG Works in AskDocs

Document Ingestion:
  File → Extract Text → Chunk (500w/100w overlap) → Embed → Store in ChromaDB

Query:
  Question → Embed → Cosine Search (top-5) → Inject Context → Gemini LLM → Answer + Citations

The key insight: Gemini only sees the retrieved passages, not the entire document. This keeps API costs low while ensuring answers are grounded in your actual documents.

---

Built for the 6th Semester Advanced Web Technology course.