Ctrl+K

Text-to-SQL Engine

Text-to-SQL Engine#

The Text-to-SQL agent is a core module which translates the Natural Language question received through the /question endpoint to SQL. The implementation can leverage business and data logic stored in the Context Store module. to generate accurate SQL given the DB schema. Currently the following NL-to-SQL implementations are included in the codebase:

  • Langchain SQL Agent - A wrapper around the Langchain SQLAgent

  • Langchain SQL Chain - A wrapper around the Langchain SQLChain

  • LlamaIndex SQL Generator - A wrapper around the LlamaIndex SQL Generator

  • Dataherald SQL Agent - Our in-house Natural Language-to-SQL agent which uses uses in-context learning

Dataherald SQL Agent#

The dataherald_sqlagent is an agent that outperforms the Langchain SQL Agent by 12%-250% in our benchmarking. It does this by leveraging up to 7 tools to generate valid SQL:

  1. QuerySQLDataBaseTool: Executes a given SQL query on the database and returns the string representation of the results.

  2. GetCurrentTimeTool: Provides access to the current date and time, allowing the agent to address temporal questions, such as “how much did the income increase this month?”

  3. TablesSQLDatabaseTool: Returns a list of table names, accompanied by a relevance score indicating their potential relevance to the given question.

  4. SchemaSQLDatabaseTool: Grants access to the schema of the tables, aiding the agent in locating relevant columns.

  5. InfoRelevantColumns: Offers additional information about specific columns, which can include descriptions from data analysts, categories, and sample rows, enriching the context for query generation.

  6. ColumnEntityChecker: Facilitates entity searches within specific columns, presenting a list of syntactically relevant entities from the selected column.

  7. GetFewShotExamples: Allows the agent to request relevant Question/SQL pairs dynamically. The agent can ask for more examples based on question complexity, fostering adaptive learning.

Abstract SQLGenerator Class#

Base class that all SQL generation classes inherit from.

SQLGenerator#

This base class defines the common structure for SQL generation classes.

create_sql_query_status(db, query, response, top_k)

Creates a SQL query status using provided parameters.

Parameters:

  • db (SQLDatabase) – The SQL database instance.

  • query (str) – The SQL query.

  • response (Response) – The Response instance.

  • top_k (int) – The number of results to return.

Returns:

The updated Response instance with the SQL query status.

Return type:

Response

generate_response(user_question, database_connection, context=None)

Generates a response to a user question based on the given user question, database connection, and optional context.

Parameters:

  • user_question (Question) – The user’s natural language question.

  • database_connection (DatabaseConnection) – The database connection information.

  • context (List[dict], optional) – (Optional) Additional context information.

Returns:

The Response containing the generated response.

Return type:

Response

Contents