LLM Text-to-Query

Convert natural language to SQL queries using local LLMs.

Interactive Shell

# Start all services
docker compose up -d

# Enter interactive REPL
docker compose exec app uv run text2query

Benchmark

# Start benchmark
docker compose --profile benchmark up --build orchestrator

The automated TPC-H benchmark runs in three phases:

Phase 1 — Setup & Validation

1. Generates/Validates TPC-H test data using tpchgen-cli
2. Loads database schema, data, and indexes
3. Generates ground truth answer CSVs from reference queries

Phase 2 — LLM Query Generation & Execution

4. Prompts natural language questions to the LLM with the database schema
5. Executes LLM-generated SQL against the database

Phase 3 — Analysis & Archiving

6. Evaluates each query using similarity criteria
7. Generates the reports to benchmark/results/YYYY-MM-DD_HH-MM-SS/