I'm looking to get better results from Open AI Chat and AI Search from my Indexd data

Question

I have completed the program necessary to enable my data to be indexed and accessed by Open AI to answer questions. The problem is the results are not very good. I need help with verifying if

1. I'm using the proper Index setup (semantic) parms
2. The proper AI Chat Model (GPT4o)
3. The proper prompt
4. and any other settings that may help.

I need a guru who knows how to get the best of this technology

Thanks All

Ron

Answer 1

Hi Ronald Wyman,

Welcome to the Microsoft Q&A Forum! Thank you for your question.

It sounds like you've made significant progress in setting up your data to be indexed and accessed by OpenAI. However, I understand that you're not getting the desired results. Let's work together to improve this. Here are a few areas we can focus on:

I'm using the proper Index setup (semantic) parms

Ensure that you're using the correct semantic parameters for your index. This will help in retrieving the most relevant data. The fields in your index have semantic titles and content, and that you're using the right analyzers. Also, checking if the index is using the latest version of the semantic ranker could help. You can refer to the Semantic ranking in Azure AI Search for detailed guidance on setting up your index with semantic configurations.

The proper AI Chat Model (GPT4o)

You mentioned GPT-4o. Confirm that you're using the correct deployment name and that the model is available in your region. Maybe you're using an older version by mistake. Combining different models for search and chat might be beneficial, like using a retrieval model for search and GPT-4 for answers. You can check the Deploy Models in Azure OpenAI to ensure you're using the correct model and deployment settings.

The proper prompt

Crafting the right prompt is crucial, you might not be structuring them effectively. The prompt should include context, user query, and clear instructions. Maybe you're not using system messages to guide the model's behavior. Examples of good prompts from Microsoft's docs could help here. Also, parameters like temperature and max tokens could affect the responses, so adjusting those might improve quality.

and any other settings that may help.

Other settings could include chunking strategies for your data. If the chunks are too big or small, the model might not retrieve the right info. Hybrid search with both keyword and vector search might cover more cases. Checking the scoring profiles and boosting certain fields could prioritize better results. Logging and monitoring would help identify where the issues are happening, like in retrieval or generation.

For more detailed guidance, you can refer to the official Microsoft documentation:

• Hybrid search using vectors and full text in Azure AI Search
• Field mappings and transformations using Azure AI Search indexers

Troubleshooting Checklist:

1. Validate your index with the REST API to ensure semantic fields are configured.
2. Test retrieval quality using the Search Explorer in Azure Portal.
3. Experiment with prompts in the Azure OpenAI Playground.
4. Use the debug=true parameter in the API to see how the model processes your prompt.

I hope this is helpful! Do not hesitate to let me know if you have any other questions.

Cheers!

Answer 2

I completed the process of updating our code to support the data types but when we post this information the error:

CannotEnableFieldForSearching: The searchable field 'invoicedate' must be of type Edm.String or Collection(Edm.String) or Collection(Edm.Single). Parameters: definition

This occurs for many of our fields that's why we assigned them all as "Edm.string" originally.

Please advise.

Thx

Ron

 {
            "name": "invoicedate",
            "type": "Edm.DateTimeOffset",
            "searchable": true,
            "filterable": "True",
            "retrievable": true,
            "stored": true,
            "sortable": false,
            "facetable": false,
            "key": false,
            "indexAnalyzer": null,
            "searchAnalyzer": null,
            "analyzer": "standard.lucene",
            "dimensions": null,
            "vectorSearchProfile": null,
            "vectorEncoding": null,
            "synonymMaps": []
        }

Answer 3

I pushed through and implemented your other suggestions. Please see the attached updates.

FieldDefinition.txt

Sample Data.txt

Thx
Ron