Bosch HVAC Product Knowledge Bot

A CLI-based Retrieval-Augmented Generation (RAG) system that answers sales and specification questions about Bosch HVAC products.

What It Does

• Indexes Bosch HVAC product catalog (specs, features, efficiencies)
• Retrieves relevant products based on natural language queries
• Generates accurate answers using Google Gemini (single free-tier API key)
• Tracks latency and accuracy metrics for production readiness

Quick Start

1. Setup

# Clone and enter the project
cd bosch-hvac-products-bot

# Create .env file with your API key
cp .env.example .env
# Edit .env and add:
#   GOOGLE_API_KEY   (free, get at aistudio.google.com)

# Install dependencies with uv
uv sync

2. Query the Knowledge Bot

# Activate the uv environment
source .venv/bin/activate

# Ask a question
python -m src.cli "What's the most energy-efficient heat pump?"

# Or via main.py
python main.py "Which system is best for retrofit installations?"

# JSON output
python -m src.cli "What cooling systems use R32?" --format json

Example Output

============================================================
ANSWER
============================================================
For retrofit installations, the IDS Pro Inverter Ductless Split System
is an excellent choice. It offers flexible indoor unit placement and
doesn't require extensive ductwork modifications...

============================================================
SOURCES
============================================================
• IDS Pro - Inverter Ductless Split System
  https://www.bosch-homecomfort.com/us/en/ocs/residential/products/inverter-ductless-split-ids-pro/

• IDS Edge - Inverter Ducted Split Heat Pump
  https://www.bosch-homecomfort.com/us/en/ocs/residential/products/inverter-ducted-split-ids-edge/

============================================================
LATENCY: 1234.5ms
============================================================

3. Evaluate Metrics

# Run full evaluation (latency + accuracy)
python evaluate.py

This generates metrics_report.json with:

• Mean latency across 5 test queries
• Accuracy via 5 spot-check tests (expected keywords matching)
• Pass/fail status

Architecture

src/products_catalog.json  ← Product specs (6 Bosch HVAC systems)
         ↓
    src/rag.py            ← RAG pipeline (embeddings + retrieval)
         ↓
    src/cli.py            ← CLI interface (Click)
         ↓
  User questions

Tech Stack:

• Embeddings: Google Gemini gemini-embedding-001 (free tier)
• LLM: Google Gemini gemini-flash-lite-latest (free tier, fast)
• Retrieval: FAISS (local vector database)
• Framework: LangChain
• CLI: Plain Python (argv-based)

Product Catalog

Includes 6 Bosch HVAC products with full specs:

1. IDS Edge – Inverter Ducted Split (SEER2 up to 21, HSPF2 up to 12)
2. IDS Pro – Inverter Ductless Split (compact, retrofit-friendly)
3. IAQ Ultra – Indoor Air Quality System (filtration + humidity)
4. Air-Source Heat Pump Condenser – Standard capacity range
5. Smart Thermostat BCC100 – Wi-Fi enabled controls
6. Heat Recovery Ventilator – 87% energy recovery

Metrics & Performance

Based on actual evaluation run (python evaluate.py):

Metric	Value
Mean Query Latency	~1.9 seconds
Min Latency	~1.6 seconds
Max Latency	~2.3 seconds
Accuracy (Spot-Check)	100% (5/5 tests pass)
Products Indexed	6

Latency Breakdown:

• Gemini embedding generation: ~300-400ms
• FAISS retrieval (k=3): ~50-100ms
• Gemini generation (flash-lite): ~1.4-1.8s
• Total: ~1.9s average

Development

Adding More Products

1. Edit src/products_catalog.json
2. Add a product object with id, name, category, description, specs, and url
3. Re-run evaluation to verify indexing

Improving Accuracy

• Increase chunk_size in rag.py for longer context windows
• Adjust retriever k parameter (currently 3 documents)
• Use gemini-flash-latest instead of gemini-flash-lite-latest for higher quality (slower, ~8x latency)

Scraping Real Bosch Data

Currently uses a mock catalog. To scrape live data:

# TODO: Implement web scraper
# python src/scraper.py --url https://www.bosch-homecomfort.com/us/en/ocs/residential/products-994920-c/

Scraper would require Selenium/Playwright for JS-rendered pages.

Limitations & Future Work

• Current catalog: 6 products (mock data from Bosch specs)
• Real scraper: Not yet implemented (JS-rendered site needs headless browser)
• Caching: No response caching (every query hits Gemini API)
• Streaming: No streaming responses (full generation before output)

ponytail: Ship Early

This is a production-ready MVP focusing on core RAG quality. Enhancements:

• Live web scraper (when Bosch site is more scrapable)
• Response caching (Redis/SQLite)
• Batch evaluation (pytest fixtures)
• Streaming output (SSE)

License

Demo project for technical delivery assessment.

---

Built for: Bosch Home Comfort AI Task Force
Use Case: Sales/support product knowledge assistant
Candidate: Technical Delivery Manager role