Software engineer,
AI systems
end to end

Ask your database in English

Natural language to SQL over enterprise databases the model has never seen.

What if someone who doesn't write SQL could just ask? TRIDENT SQL turns plain-English business questions into working SQL against unfamiliar multi-table schemas, in four phases. The highlighted components were designed and built for this thesis.

Target database6,601 training questions

1. Phase 1

   Pre-processing

   • Value index
   • Example index
   • Query-Tree Rules

   Indexes & rules

2. Phase 2

   Schema linking

   + user's question

   • Value retrieval
   • Column finding

   Focused schema

3. Phase 3

   SQL generation

   • Divide-and-conquer ×4
   • Example-based ×4
   • Plan-and-Fragment ×4

   12 candidate SQLs

4. Phase 4

   Correction & selection

   • Repair 8 checks
   • Query Tree
   • Cross-Writer Agreement

   Final SQL

1. Phase 1 · Pre-processing. Before a single question arrives, the system studies its terrain. Every text value in the database is embedded for lookup, 6,601 solved training questions are indexed by the shape of their queries, and Query-Tree Rules are distilled from the mistakes the model makes on the training set, each rule kept only if it measurably raises accuracy.
2. Phase 2 · Schema linking. The question is anchored to the database values it mentions and to the columns it truly needs, narrowing a sprawling schema to a focused one.
3. Phase 3 · SQL generation. Three independent writers each draft four candidate queries, twelve in all. One decomposes the problem, one learns from similar solved examples, and the third, Plan-and-Fragment, reasons before it writes: it first commits to exactly what the question asks to return, then translates every domain hint into a precise SQL fragment and assembles the query from those pieces verbatim, so no formula, literal or unit is lost in translation.
4. Phase 4 · Correction and selection. Each candidate passes eight repair checks. The Query Tree then performs a structural audit: the English question is decomposed into a tree of what it demands, the SQL is parsed into a tree of what it actually does, and the two are compared branch by branch. Only the flagged branches are repaired; every correct part of the query is left untouched. Finally, Cross-Writer Agreement chooses the answer: each writer casts a single vote, so no writer can win on volume alone, and an LLM judge settles the genuinely close calls.

It achieved 70% execution accuracy on the BIRD dev set, running on open-weight models only and training-free. No proprietary API, which means it can be deployed inside a company's own environment with no data leaving it — the difference between a demo and something a regulated customer can actually run.

Results on the BIRD dev set

Execution accuracy	70%
Benchmark questions	1,534
Distinct databases	11
Proprietary APIs	0

github.com/VineetV2/Trident-SQL ↗