Is there a way to get Wren to search for similar column values before querying.

[EoinOGara]

For the use cases of my data, Wren out of the box often falls short when the user queries on a specific column value.

Example: Show me my sales data for Amazon in the last month. Wren tries to write query with CustomerName = "Amazon" and if CustomerName is stored as "Amazon.com, Inc" then it is unable to find the customer and filter results for that customer.

There are several columns that this applies to, and there are too many values of customer to place into instructions or in the mdl file, or to add sql-question pairs for each customer.
Are there any pre-existing solutions to this issue?
I believe a tool that searches the distinct column values with string similarity to find the precise values, before executing the query could work.
Is there any flexibility in configuring such a tool or prompt without completely forking and editing Wren code?

[Y01N]

• I am same person who started this discussion (but on personal github acct)
• In a different prototype of a sql chatbot, I used a python tool to find the nearest distinct customer matches (fuzzy match). Then I would pass those results to the llm so it could use a mixture of string similarity and llm semantics to decide the proper customer.
• Is the best way to achieve similar results in Wren to create a SQL Function that finds the closest string match in a column for a given column value?
• Then use global instructions to tell SQL generation to use that specific SQL function when filtering by CustomerName.
• This solution is unable to use a the language model for selecting the proper customer, and would be fully reliant on the fuzzy match working properly to determine what exact value the user is asking for.
• Are there other ways to customize the SQL Generation pipeline to include external tools or configure the prompts?
• I can also see a solution external to Wren AI where the model used is a custom agent that has tooling to find the top 5 similiar column values that the user may be querying on, then use instructions to tell the llm to use one of those values.

My purpose of this post is to clarify: Are there any official and configurable ways to customize the SQL generation pipeline / prompts / tooling?
As well as open up any discussion for the best way to solve the issue of the LLM needing to be aware of the exact value of a column without getting the precise value from the user.