🎓 AWS re:Invent 2025 Workshop | For educational purposes - demonstrates production patterns
Duration: 60 minutes | Lab 1: 25 min | Lab 2: 20 min
Learn to build enterprise-grade hybrid search combining semantic similarity, lexical matching, and fuzzy search with Aurora PostgreSQL. Integrate Model Context Protocol (MCP) for natural language database queries with Row-Level Security (RLS).
What You'll Build:
- Multi-modal search system over 21,704 products
- AI agent with natural language database access
- Secure multi-tenant system with PostgreSQL RLS
├── lab1-hybrid-search/
│ ├── notebook/
│ │ └── dat409-hybrid-search-notebook.ipynb # Lab 1: Hybrid search implementation
│ ├── data/
│ │ └── amazon-products.csv # 21,704 product dataset
│ └── requirements.txt
├── lab2-mcp-agent/
│ ├── streamlit_app.py # Lab 2: Interactive demo app
│ ├── test_personas.sh # RLS testing script
│ └── requirements.txt
├── scripts/
│ ├── bootstrap-code-editor.sh # Environment setup
│ ├── setup-database.sh # Database initialization
│ └── setup/ # Helper utilities
└── solutions/ # Reference implementations
Build a multi-modal search system combining three complementary techniques:
| Method | Technology | Use Case |
|---|---|---|
| Semantic | pgvector + HNSW + Cohere | Conceptual queries ("eco-friendly products") |
| Keyword | PostgreSQL tsvector + GIN | Exact terms ("iPhone 15 Pro") |
| Fuzzy | pg_trgm + GIN | Typo tolerance ("wireles hedphones") |
What You'll Learn:
- When to use semantic vs keyword search
- Index strategies for production workloads (HNSW vs IVFFlat)
- Result fusion with Reciprocal Rank Fusion (RRF)
- Cohere Rerank for ML-based result optimization
Hands-On:
cd /workshop/lab1-hybrid-search/notebook
# Open dat409-hybrid-search-notebook.ipynbYou'll implement fuzzy search, semantic search, and hybrid RRF queries with TODO blocks guiding you through each step.
Build an AI agent that queries databases using natural language:
User: "Show warranty info for headphones"
↓
Strands Agent (Claude Sonnet 4)
↓
MCP Tools → SQL Query
↓
Aurora PostgreSQL (RLS filtered)
↓
Results based on user persona
What You'll Learn:
- Model Context Protocol (MCP) for standardized database access
- Application-level security with PostgreSQL RLS
- AI agent patterns for database queries
- Multi-tenant data isolation strategies
Hands-On:
cd /workshop/lab2-mcp-agent
./test_personas.sh # Test RLS policies
streamlit run streamlit_app.py # Interactive demoExplore how different personas (customer, support agent, product manager) see different data through RLS policies.
For AWS re:Invent Participants:
- Access your Code Editor environment via the provided CloudFront URL
- Navigate to
/workshop/lab1-hybrid-search/notebook/ - Open
dat409-hybrid-search-notebook.ipynb - Follow the guided TODO blocks
Environment Includes:
- ✅ Aurora PostgreSQL 17.5 with pgvector
- ✅ 21,704 products pre-loaded with embeddings
- ✅ Python 3.13 + Jupyter + all dependencies
- ✅ Amazon Bedrock access (Cohere models)
- ✅ MCP server pre-configured
| Component | Technology | Purpose |
|---|---|---|
| Database | Aurora PostgreSQL 17.5 | Vector storage with pgvector extension |
| Vector Index | HNSW (pgvector) | Fast ANN search with 95%+ recall |
| Embeddings | Cohere Embed English v3 | 1024-dimensional dense vectors |
| Full-Text | PostgreSQL tsvector + GIN |
Lexical search and BM25-style ranking |
| Fuzzy Match | pg_trgm + GIN | Trigram similarity for typo tolerance |
| MCP Server | awslabs.postgres-mcp-server |
Standardized database access tools |
| AI Agent | Strands Agent Framework | Tool-calling orchestration layer |
| LLM | Claude Sonnet 4 | Natural language → SQL translation |
| RLS | PostgreSQL Row-Level Security | Declarative multi-tenancy |
| Data API | Aurora Data API | Serverless, IAM-authenticated access |
| Python | 3.13 (pandas, psycopg3, boto3) | Data loading and orchestration |
How Natural Language Queries Work:
"Show warranty info" → Strands Agent → MCP Tools → Aurora PostgreSQL → Filtered Results
↓ ↓ ↓
Claude Sonnet 4 SQL Query RLS Policies
Key Components:
| Component | Role | Technology |
|---|---|---|
| Strands Agent | Orchestration & tool calling | Python framework |
| Claude Sonnet 4 | Natural language → SQL | Amazon Bedrock |
| MCP Client | Standardized database tools | awslabs.postgres-mcp-server |
| Aurora Data API | Serverless database access | IAM authentication |
| RLS Policies | Row-level security | PostgreSQL |
Why This Pattern?
- ✅ Standard Practice: Agent uses admin access; security via application-level filtering
- ✅ Serverless: No VPC required with Data API
- ✅ Portable: MCP tools work across different AI frameworks
- ✅ Intelligent: Multi-step reasoning with context awareness
Production Considerations:
- Agent requires admin credentials (standard for AI agents)
- Data API adds ~10ms latency (acceptable for agentic workflows)
- RLS provides database-level isolation per persona
Documentation:
- Aurora PostgreSQL - Managed PostgreSQL service
- pgvector - Vector similarity search extension
- Model Context Protocol - Standardized AI tool protocol
- PostgreSQL RLS - Row-level security
Related AWS Services:
- Amazon Bedrock - Cohere embeddings & rerank
- RDS Data API - Serverless database access
- Secrets Manager - Credential management
Found this helpful?
- ⭐ Star this repository
- 🍴 Fork for your own use cases
- 🐛 Report issues
- 💡 Submit pull requests
- 📢 Share with colleagues
Contributions welcome! See CONTRIBUTING.md for guidelines.
MIT-0 License - See LICENSE
Security issues: CONTRIBUTING.md
AWS re:Invent 2025 | DAT409 Builder's Session
Hybrid Search with Aurora PostgreSQL for MCP Retrieval
© 2025 Shayon Sanyal