StellarMIND — Chat-to-SQL with RAG

Problem

Business users need to query databases without knowing SQL. Existing tools lack context-aware query generation and safety guarantees.

Why It Matters

Data democratization requires non-technical users to access insights without engineering bottlenecks. Raw LLM-to-SQL is unreliable. RAG with schema context fixes this.

System Approach

• Spring Boot MCP server with Tool interface for executeDataQuery
• pgvector for storing schema knowledge chunks and embeddings
• Spring AI for LLM integration (provider-agnostic — works with OpenAI, Anthropic, etc.)
• Chain-of-Thought (CoT) web interface for query debugging and transparency
• Read-only SQL enforcement via query parsing (only SELECT, WITH allowed)

Key Decisions & Trade-offs

• Read-only restriction limits use cases but ensures database safety
• pgvector requires PostgreSQL — not database-agnostic, but worth the trade-off
• MCP transport (stdio) over HTTP for better AI assistant integration
• Separate stellarmind-server and stellarmind-client for modularity

Current Status

Core query flow working. CoT UI functional. Newman test suite passing.

Roadmap

• Add support for streaming responses
• Implement query history and favorites
• Add schema auto-discovery

What I'd Improve Next

• Could add query result visualization
• Consider supporting multiple database connections