How googlecloudplatform/nl2sql works
NL2SQL is a composable Python framework that turns a natural-language question into SQL through selectable task modules. Its linear executor fetches a database, optionally narrows tables and columns, generates SQL, and can invoke evaluation-and-repair before returning a result. Datasets support connected databases and Excel-derived custom data; task prompts and LLM integrations are configurable, and optional analytics persist execution artefacts. The map preserves modular join selection and ReAct generation as separate capabilities. Several source excerpts are partial, so uncertain runtime links are omitted or marked unverified.
Main entry points: linear_executor/core.py Linear Executor · base.py Dataset · prompts Prompt Assets · table_selection/core.py Table Selection
Components 20
Workflow
- Linear Executor
nl2sql/executors/linear_executor/core.py - SQL Result
nl2sql/executors/linear_executor/core.py
NL2SQL Tasks
- Table Selection
nl2sql/tasks/table_selection/core.py - Column Selection
nl2sql/tasks/column_selection/core.py - Join Selection
nl2sql/tasks/join_selection/core.py - SQL Generation
nl2sql/tasks/sql_generation/core.py - ReAct Generation
nl2sql/tasks/sql_generation/react.py - SQL Eval Repair
nl2sql/tasks/eval_fix/core.py - Prompt Assets
nl2sql/assets/prompts - Yes-No Classifier
nl2sql/commons/utils/classifiers.py
Data Access
- Dataset
nl2sql/datasets/base.py - Standard Dataset
nl2sql/datasets/standard.py - Excel Dataset
nl2sql/datasets/custom.py - SQL Database
LLM Integrations
- LLM Provider
- Vertex AI Client
nl2sql/llms/vertexai.py - Google PaLM Client
nl2sql/llms/google_palm.py
Analytics
- Analytics Handler
nl2sql/commons/reporting/persist.py - System Fingerprint
nl2sql/commons/reporting/fingerprint.py
Other components
- Questioner
Connections 23
- Questioner → Linear Executor: submits question
README.md - Linear Executor → Dataset: gets database
nl2sql/executors/linear_executor/core.py - Linear Executor → Table Selection: runs selection
nl2sql/executors/linear_executor/core.py - Linear Executor → Column Selection: runs selection
nl2sql/executors/linear_executor/core.py - Linear Executor → Dataset: filters schema
nl2sql/executors/linear_executor/core.py - Linear Executor → SQL Generation: generates SQL
nl2sql/executors/linear_executor/core.py - Linear Executor → SQL Eval Repair: repairs query
nl2sql/executors/linear_executor/core.py - Linear Executor → SQL Result: returns result
nl2sql/executors/linear_executor/core.py - SQL Generation → LLM Provider: calls model
nl2sql/tasks/sql_generation/core.py - SQL Eval Repair → LLM Provider: calls model
nl2sql/tasks/eval_fix/core.py - Table Selection → Prompt Assets: uses prompts
nl2sql/tasks/table_selection/core.py - Column Selection → Prompt Assets: uses prompts
nl2sql/tasks/column_selection/core.py - SQL Generation → Prompt Assets: uses prompts
nl2sql/tasks/sql_generation/core.py - SQL Eval Repair → Prompt Assets: uses prompts
nl2sql/tasks/eval_fix/core.py - Table Selection → Yes-No Classifier: classifies output
nl2sql/tasks/table_selection/core.py - Standard Dataset → Dataset: implements dataset
nl2sql/datasets/standard.py - Excel Dataset → Dataset: builds dataset
nl2sql/datasets/custom.py - Dataset → SQL Database: connects schema
- Vertex AI Client → LLM Provider: provides client
- Google PaLM Client → LLM Provider: provides client
- Join Selection → Dataset: uses database
nl2sql/tasks/join_selection/core.py - ReAct Generation → LLM Provider: uses model
- Analytics Handler → System Fingerprint: uses fingerprint
nl2sql/commons/reporting/persist.py