SQL-Agent

[jeevaarul]

I have a database with over 40 tables, but the table and column names are not meaningful names and a large data size, here’s how to handle the SQL agent flow. My requirement is to perform NPL (Natural Language Processing) to SQL conversions. The end user asks simple questions in English, and the system should create an SQL query based on the table names and column names in the database, which is MSSQL.

My current approach involves using the SQL agent with Langraph. This tool extracts the table names and database metadata given the context, but the generated query isn’t always accurate.

[dirkbrnd]

Hi @jeevaarul
We have faced this exact issue with some clients before. SQL is strangely really hard for models to get right always, which is required here.
I can suggest you take a look at a "Route" team. You could make a team with 40 different agents, each assigned to one table. And you give it the table metadata in the instructions of that agent (so it doesn't need to look it up).
This of course only works if you don't intend to do any joins between tables. Then I'd suggest a more flexible workflow approach.

I'm looking to improve our SQL tools and then build an example with this kind of agent setup.

[claudiomontoya]

you must have gettables functions getschema runquery retrieves tables with a comment, retrieves schema with a comment of fields, run query receives query and executes in case of error add a function that receives query and error and use claude 3.5 or 3.7 so that it corrects with that it would work well. Is traducción español is my speaking