Hi Yogeshwar,
To respond to your questions:
1. Is the Google Custom Search JSON API the recommended way to do this?
Not for a new implementation. Google has already announced that the Custom Search JSON API is closed to new customers and will be retired in 2027. I would avoid designing a new solution around a service that is already being phased out.
2. Has anyone implemented snippet-level evaluation in a microflow?
From an architectural perspective, I would use snippet-level evaluation as a filtering step rather than as the final decision criteria.
The main concern is false positives. A result may contain both the customer name and keyword, but that alone does not prove there is a meaningful relationship between them.
As long as the provider returns a consistent JSON structure, I would expect the import mapping to be relatively straightforward. The bigger challenge is defining reliable business rules for determining what qualifies as a true match.
3. Has anyone integrated an LLM step for context-based classification?
This is the approach I would recommend for the cases that remain after the initial filtering. Instead of checking only whether both terms appear in the result, an LLM can help evaluate whether the content actually indicates a relevant relationship.
Mendix already provides OpenAI/Azure OpenAI connectors, making this type of integration relatively straightforward. I would also request a confidence score and brief justification so reviewers can understand why a result was classified a certain way.
My recommendation would be:
I hope this helps answer your question. If you need any additional details, feel free to reply below.
Best regards,
Oswaldo Valles