omniquery-ai

OmniQuery AI is an enterprise-grade Natural Language to SQL Assistant designed for complex relational database architectures, specifically Microsoft SQL Server / AdventureWorks. It is powered by Hybrid Vector Search (Qdrant + FastEmbed), Llama 3.3 70B on Groq, dynamic Primary/Foreign Key Graph Context Injection, and an autonomous Self-Healing Error Correction Loop. The project addresses large enterprise databases such as AdventureWorks with 90+ tables and hundreds of columns, where dumping the entire database schema into the LLM context window is inefficient, expensive, and leads to severe hallucinations and invalid JOINs.

OmniQuery AI uses a 4-Stage Hybrid RAG Pipeline. It indexes pre-computed 1-sentence business summaries, column definitions, and primary/foreign keys into Qdrant. It uses sentence-transformers/all-MiniLM-L6-v2 for semantic meaning and Qdrant/bm25 for exact column/table name matching. It dynamically pulls active foreign key relationships so the LLM constructs multi-table JOIN statements. If SQL Server returns a runtime or syntax error, the error message is fed back to the LLM to automatically repair and re-execute the query up to 3 retries.

OmniQuery AI includes enterprise DDL/DML guardrails that protect against SQL injection and destructive statements by validating AST tokens and regex patterns before execution. It also provides a zero-setup offline mock demo mode that enables recruiters, reviewers, and macOS/Linux users to test the assistant immediately without needing a live Microsoft SQL Server instance running locally. The repository includes an interactive CLI, non-interactive query runner, configuration, tests, and an MIT License.