Overview
The brief
A RAG-powered analytics platform that turns plain-English questions into SQL, executes them, and visualizes the results as interactive dashboards — so non-technical teams can query production databases directly.
The problem
Data teams were bottlenecked writing ad-hoc SQL for stakeholders. Existing 'AI SQL' tools hallucinated columns and produced broken queries because they lacked schema context.
Our role
- Schema extraction pipeline
- Vector embeddings + Qdrant retrieval
- Natural-language → SQL agent
- Safe query execution & guardrails
- Interactive dashboard rendering
Outcomes
Results that moved the number.
96%
Query accuracy
< 4s
Question → chart
PG + MySQL
Databases supported
0
SQL written by users
How we built it
Process.
01
Schema-aware retrieval
Table and column metadata is embedded into Qdrant so every prompt is grounded in the real database — no hallucinated fields.
02
Constrained SQL generation
The agent generates read-only SQL against a whitelisted schema, with automatic parameterization and query timeouts.
03
Chart auto-selection
Result shape drives visualization — categorical → bar, temporal → line, single value → KPI — with editable overrides.
04
Explainable answers
Every answer shows the generated SQL and the reasoning, so analysts trust and can audit what the model did.
Highlights
What makes it stand out.
- Plain-English to production SQL with 96% accuracy
- Qdrant-backed schema retrieval prevents hallucinations
- Auto-selected charts with editable overrides
- Every answer is auditable — SQL and reasoning shown
Have a similar project?
