[Feature Proposal] csv_query: Natural Language SQL Query Interface for CSV Files Using DuckDB

[NIrajan-15]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

Proposing a new utility chain called csv_query that allows querying CSV files using natural language via DuckDB SQL.

This module would:

• Take an LLM and a path to a CSV file.
• Automatically infer schema with read_csv_auto.
• Use LangChain’s existing SQL database + prompt chain infrastructure.
• Return natural language answers to user questions via a self-contained Runnable.

It supports reproducibility, and memory-efficient querying, and removes the need for manual database setup or Pandas integration.

Motivation

Current options in LangChain, such as CSVLoader and PandasAgent, either convert CSVs to documents or require them to be preloaded into memory. That limits use in production or real-time CSV analysis.

This feature targets:

• Users with large or semi-structured CSVs
• Simple API (create_csv_runner(LLM, path))
• Better performance using DuckDB
• Cleaner integration with LangChain’s SQL Database system

Proposal (If applicable)

Create a new module csv_query:

• Contains create_csv_runner() factory
• Based on LangChain’s existing SQLDatabase support
• Uses DuckDB and creates a VIEW over the CSV
• Chains prompt → SQL generation → result → final natural answer

Optional enhancements can include multiple backends (pandas, polars) and a tool wrapper for agent usage in the future.

Would love to open a PR if maintainers approve the idea.