[SQL] How to Customize Input/Output Variable Names in LangChain Multi-Step SQL Query Chain?

I’m setting up a LangChain workflow to generate, validate, execute, and format SQL queries in multiple steps:

1. Generate SQL Query - Create an SQL query from user input.
2. Validate the Query - Prompt the LLM with the user question, database context, and generated SQL query.
3. Execute the Query - Run the validated query and store the result.
4. Convert to Natural Language - Take the result and reformat it into a human-readable response.

The problem I’m facing is in configuring custom variable names between steps. For example: • The built-in create_sql_query_chain function seems to accept only “query” as an input, but I want to label it as unvalidated_query to pass it into the validation step. • During validation, the prompt step only recognizes variables passed to the main invoke() function, which limits my ability to manage intermediate input/output names.

Here’s an example of the chain structure I’m trying to achieve:

validated_query_chain = validation_prompt | llm | StrOutputParser()

final_chain = RunnablePassthrough.assign(
query=create_sql_query_chain,
validated_query=validated_query_chain,
result=execute_query_chain
) | {
"answer": answer_back_to_natural_language_prompt | llm | StrOutputParser(),
"query": query,
"validated_query": validated_query
}


I tried various ways to rename or reassign input/output variables at each step, hoping it would enable a flexible chain where each step could reference custom-named outputs from the previous step. Unfortunately, this led to errors, as it seems LangChain’s intermediate steps are tightly bound to default naming conventions.

I expected to be able to customize these names so that each stage (like validation) could access custom-named inputs such as unvalidated_query. Instead, the constraints around input handling make it hard to link steps flexibly.

Continue reading...