zucchetti-llm-course

Lab 04: Query Router for Multi-Source RAG

Lecture: 5, Advanced RAG and Intelligent Routing Duration: 30 minutes Difficulty: Intermediate

Objectives

Build an intent classifier that routes queries to the correct data source
Use LLM to generate SQL from natural language
Handle routing errors with fallback strategies
Test with mixed query types (policy questions vs. data questions)

Prerequisites

Requirement	How to Check
HR policy documents	Read from `lab_07_evaluation/data/policies/`
Dependencies	`pip install openai chromadb sentence-transformers rich python-dotenv`
LLM access	Ollama or DeepSeek API key

Quick Start

# Create the employee database (auto-runs if missing)
cd lab_04_query_router
python setup_db.py

# Run the lab
python lab_04_query_router.py

Files

File	Description
`lab_04_query_router.py`	Main router script with TODO section
`setup_db.py`	Creates the SQLite employee database
`solutions.py`	Embedding-based classification solution
`data/employees.db`	50 Italian employee records

Tasks

Task 1: Understand the Data Sources

Two sources are available:

Vector DB (ChromaDB): HR policy documents from Lab 07
SQL DB (SQLite): 50 employee records with salary, department, etc.

Task 2: Implement Query Classification

Replace the keyword-based classify_query() with embedding-based classification.

Task 3: Test with Mixed Queries

Watch how the router handles SQL vs. policy questions. Notice error handling when SQL generation fails.

Task 4: Interactive Testing

Try ambiguous questions like “Can a new employee work remotely?” — does it route correctly?

Expected Output

Q1: How many employees work in Engineering?
  Route: sql
  SQL: SELECT COUNT(*) FROM employees WHERE department = 'Engineering'
  Answer: There are 12 employees in the Engineering department.

Q4: What is the remote work policy?
  Route: vector
  Answer: According to the Remote Work Policy, eligible employees may
  work remotely up to 3 days per week...

🆘 Troubleshooting

Issue	Solution
Policy docs not found	Check that `lab_07_evaluation/data/policies/` is in your clone
SQL errors	The script auto-retries with error context
Wrong routing	Improve the `classify_query()` function — see solutions.py