A production-ready multilingual document Question-Answering system using Retrieval-Augmented Generation (RAG) with specialized agents for intelligent document processing and querying.

Click here for more In-Depth Documentation and to chat with the codebase

• 🌐 Multilingual Support: Process documents in English, Chinese, Hindi, Bengali, and Urdu
• 🤖 Agentic Architecture: 11 specialized agents for end-to-end document intelligence
• 🔍 Hybrid Retrieval: Combines semantic search (Gemini embeddings) + BM25 keyword matching
• 📊 Advanced Reranking: Gemini Flash-based relevance scoring
• 💬 Chat Memory: Multi-turn conversations with context retention
• 🧠 Query Decomposition: Breaks complex queries into manageable sub-queries
• ⚡ Production-Ready: Docker containerization, comprehensive logging, error handling
• 🎨 Modern UI: Dark-mode NiceGUI interface with real-time streaming responses

┌─────────────────────────────────────────────────────────────────┐
│                     Document Processing Pipeline                │
└─────────────────────────────────────────────────────────────────┘
                               │
                               ▼
┌──────────┐     ┌──────────┐     ┌──────────┐     ┌──────────┐
│   PDF    │────▶│   OCR    │────▶│ Chunking │────▶│Embedding │
│  Upload  │     │(Tesseract│     │(Lang-Aware)    │ (Gemini) │
└──────────┘     └──────────┘     └──────────┘     └──────────┘
                                                            │
                                                            ▼
┌─────────────────────────────────────────────────────────────────┐
│                      Qdrant Vector Database                     │
│                  (106+ documents, 96+ indexed)                  │
└─────────────────────────────────────────────────────────────────┘
                               │
                               ▼
┌──────────┐     ┌──────────┐     ┌──────────┐     ┌──────────┐
│  Query   │────▶│ Retrieval│────▶│ Reranker │────▶│   LLM    │
│          │     │ (Hybrid) │     │ (Gemini) │     │ (Ollama) │
└──────────┘     └──────────┘     └──────────┘     └──────────┘
                                                            │
                                                            ▼
                                                   ┌──────────────┐
                                                   │   Answer +   │
                                                   │ Chat Memory  │
                                                   └──────────────┘

• Windows 10/11 or Linux
• Python 3.11+
• Docker Desktop (for Qdrant)
• Ollama (for local LLM)
• Tesseract OCR with language packs

# Install Python
winget install Python.Python.3.11

# Install Docker Desktop
winget install Docker.DockerDesktop

# Install Ollama
winget install Ollama.Ollama

# Install Tesseract
winget install UB-Mannheim.TesseractOCR

# Python 3.11+
sudo apt install python3.11 python3.11-venv python3-pip

# Docker
curl -fsSL https://get.docker.com -o get-docker.sh
sudo sh get-docker.sh

# Ollama
curl https://ollama.ai/install.sh | sh

# Tesseract
sudo apt install tesseract-ocr tesseract-ocr-eng tesseract-ocr-chi-sim \
                 tesseract-ocr-hin tesseract-ocr-ben tesseract-ocr-urd

# Clone repository
git clone https://github.com/yourusername/multilingual-rag.git
cd multilingual-rag

# Create virtual environment
python -m venv .venv

# Activate virtual environment
# Windows:
.venv\Scripts\activate
# Linux:
source .venv/bin/activate

# Install Python dependencies
pip install -r requirements.txt

# Copy environment template
cp .env.example .env

# Edit .env and add your Gemini API keys
nano .env  # or notepad .env on Windows

Edit .env and add your Google Gemini API keys:

GEMINI_API_KEYS=your-key-1,your-key-2,your-key-3

Get API keys from: https://makersuite.google.com/app/apikey

# Start Qdrant vector database
docker-compose --profile cpu up -d

# Pull Ollama model
ollama pull gemma3:4b

# Verify services
curl http://localhost:6333/healthz  # Qdrant
curl http://localhost:11434/api/tags  # Ollama

# Clean and reindex (first time)
python tests/clean_and_reindex.py

# Process sample document
python scripts/process_one_sample.py

# Process all documents
python -m src.main

# Launch NiceGUI interface
python -m src.main

# Open browser to http://127.0.0.1:8080

গবেষণা নির্দেশিকায় কী কী বিষয় আছে?
(What topics are covered in the research guidelines?)

عارضی ملازمین کی ملازمت کی مدت کب تک بڑھائی گئی ہے؟
(Until when has the employment period of adhoc employees been extended?)

这个文件是关于什么的？
(What is this document about?)

What types of administrative documents are available?
Compare research guidelines and employment procedures.

# Run quick tests (5 core features)
python tests/test_rag_quick.py

# Expected output:
# ✅ PASS - Query Decomposition
# ✅ PASS - Document Retrieval
# ✅ PASS - Reranking
# ✅ PASS - Single-Turn Chat
# ✅ PASS - Chat Memory
# Passed: 5/5 (100.0%)

# Run comprehensive tests with proof
python scripts/run_comprehensive_tests.py
# Generates docs/RAG_TEST_PROOF.json

• Process multilingual policy documents
• Extract specific regulations, dates, amounts
• Compare procedures across departments

• Analyze research guidelines in local languages
• Find funding limits and application procedures
• Cross-reference multiple guideline documents

• Employment extensions and notifications
• Budget allocation information
• Organizational hierarchies

• Unified search across language barriers
• Semantic understanding of queries
• Context-aware answers with citations

• Technical Documentation - Architecture, components, API docs
• User Guide - Operating and maintaining the system
• Performance Report - Benchmarks and test results
• Presentation - Project overview for interviews
• Test Questions - Comprehensive test scenarios
• Qdrant Fix Guide - Filesystem issue resolution

Key configuration options in .env:

# Vector Database
QDRANT_URL=http://localhost:6333
QDRANT_COLLECTION_NAME=multilingual_docs

# Embeddings
EMBEDDING_MODEL=models/text-embedding-004
GEMINI_API_KEYS=key1,key2,key3

# LLMs
LLM_PRIMARY=ollama:gemma3:4b
LLM_FALLBACK=gemini-flash-latest
OLLAMA_HOST=http://localhost:11434

# Retrieval
ENABLE_BM25=true
DENSE_WEIGHT=0.6
KEYWORD_WEIGHT=0.4

# Reranking
ENABLE_RERANK=true
RERANK_TOP_K=30

# OCR
OCR_ENGINE=tesseract
OCR_LANGS=eng,chi_sim,hin,ben,urd

1. Qdrant OutputTooSmall Error

• Solution: Ensure using Docker named volumes (not bind mounts)
• See: Qdrant Fix Guide

2. Embedding API Errors

• Check: API keys in .env
• Verify: Model name is models/text-embedding-004

3. Ollama Connection Failed

• Start: ollama serve
• Pull model: ollama pull gemma3:4b

4. Tesseract Language Packs Missing

• Windows: Install from UB-Mannheim installer
• Linux: sudo apt install tesseract-ocr-[lang]

5. Upload Not Working

• Fixed: NiceGUI event handler updated
• Check: PDF files only, max 50MB

Contributions welcome! Please:

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Commit changes (git commit -m 'Add amazing feature')
4. Push to branch (git push origin feature/amazing-feature)
5. Open a Pull Request

This project is licensed under the MIT License - see LICENSE file for details.

• Google Gemini: For powerful multilingual embeddings
• Ollama: For easy local LLM deployment
• Qdrant: For efficient vector search
• Tesseract: For OCR capabilities
• NiceGUI: For beautiful Python UI

• Email: contactayushdhiman@gmail.com
• LinkedIn: Ayush Dhiman

If you find this project useful, please consider giving it a star ⭐

Built with ❤️ for multilingual document intelligence by Ayush Dhiman