🏗️ BIS Standard Recommendation Engine

BIS x Sigma Squad AI Hackathon — AI-powered recommendation engine that maps building material descriptions to applicable Bureau of Indian Standards (BIS).

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                      inference.py (Judge Entry)                 │
│                   SQLite WAL Cache (sub-ms lookups)             │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  ┌──────────────┐   ┌──────────────┐   ┌──────────────────┐    │
│  │  ChromaDB    │   │    BM25      │   │  Cross-Encoder   │    │
│  │  (Dense)     │──▶│  (Sparse)    │──▶│  Reranker        │    │
│  │  Vector      │   │  Keyword     │   │  BAAI/bge-base   │    │
│  └──────────────┘   └──────────────┘   └──────────────────┘    │
│         │                                       │               │
│         └───────── RRF Fusion ──────────────────┘               │
│                         │                                       │
│                   ┌─────▼──────┐                                │
│                   │  Groq LLM  │                                │
│                   │  (llama3)  │                                │
│                   └────────────┘                                │
├─────────────────────────────────────────────────────────────────┤
│  LlamaParse (PDF → MD) → Regex Chunker ("SUMMARY OF IS")       │
└─────────────────────────────────────────────────────────────────┘

Quick Start

1. Environment Setup

# Clone and enter project
cd bissigma_engine

# Create virtual environment
python -m venv venv
source venv/bin/activate  # or venv\Scripts\activate on Windows

# Install dependencies
pip install -r requirements.txt

# Configure API keys
cp .env.example .env
# Edit .env with your GROQ_API_KEY and LLAMA_CLOUD_API_KEY

2. Data Ingestion

Place your dataset.pdf (SP 21:2005) in the data/ folder, then:

python -m src.ingest

3. Run Inference (Judge's Script)

python inference.py --input data/queries.json --output data/results.json

4. Launch Demo UI

streamlit run app.py

5. Docker (Full Stack)

docker-compose up --build

• API: http://localhost:8000/docs
• UI: http://localhost:8501

Project Structure

├── src/
│   ├── __init__.py          # Package init
│   ├── config.py            # Environment configuration
│   ├── ingest.py            # LlamaParse + Regex chunking pipeline
│   ├── retriever.py         # Hybrid search (Vector + BM25) + Reranking
│   └── generator.py         # Groq LLM generation
├── data/                    # PDFs, ChromaDB, SQLite cache
├── inference.py             # Judge's bulletproof entry point
├── app.py                   # Streamlit demo UI
├── app_api.py               # FastAPI microservice backend
├── Dockerfile               # Container image
├── docker-compose.yml       # Local dev stack
└── requirements.txt         # Python dependencies

Key Design Decisions

Concern	Solution
Zero-crash	Every query wrapped in try/except → empty [] fallback
Latency < 5s	SQLite WAL cache, Groq ultra-fast inference, local ChromaDB
Accuracy >80%	Hybrid search (dense + sparse) + Cross-Encoder reranker
Chunking	Regex split at "SUMMARY OF IS" — one chunk = one standard
Compliance	Strict system prompt prevents LLM hallucination

API Endpoints

Method	Path	Description
GET	/health	Health check
POST	/recommend	Get BIS recommendations for a query
POST	/ingest	Trigger data ingestion pipeline
GET	/docs	Interactive Swagger UI

---

Built with ❤️ by Sigma Squad