Hybrid RAG & NL-to-SQL Engine

Problem

Enterprise users needed to query both unstructured documents and structured databases using plain English. LLM-generated SQL is unreliable and dangerous without validation. Single-vector retrieval misses keyword matches. And any LLM dependency is a reliability risk without fallback.

Solution

• Indexed 100K+ chunks in Aurora PostgreSQL with pgvector (1024-dim Titan embeddings) for semantic search.
• Hybrid retrieval fuses pgvector semantic search with BM25 keyword search via Reciprocal Rank Fusion (RRF) — eliminates blind spots of single-mode retrieval.
• LLM-based reranking pushes relevance to 85%+; RAGAS-aligned evaluation (MRR, correctness, groundedness, completeness) with golden test datasets validates quality continuously.
• NL-to-SQL engine with circuit breaker and Sonnet-to-Haiku failover for 99.9% uptime.
• Every generated SQL query parsed by sqlglot into an AST — structurally validated to prevent injection before execution.
• PII detection and column-level access controls applied at the query layer; S3 audit logging on every query for compliance.

System Flow

Architecture

• 01pgvector (1024-dim Titan embeddings) + BM25 hybrid retrieval via RRF
• 02LLM-based reranking for 85%+ relevance on golden test datasets
• 03RAGAS evaluation: MRR, correctness, groundedness, completeness
• 04S3 audit logging on every query for compliance tracking
• 05Circuit breaker with Sonnet-to-Haiku failover for 99.9% uptime
• 06sqlglot AST validation - SQL injection prevention on every query
• 07PII detection + column-level access controls at query layer

Impact

• 85%+ retrieval relevance across 100K+ indexed document chunks
• 500+ NL-to-SQL queries processed daily at 99.9% uptime
• Zero SQL injection incidents in production via sqlglot AST-level validation
• S3 audit logging on every query for compliance