Azure AI Search: Adding a vector field returns empty results

Question

Hi,

I’m encountering significant issues with adding a vector search in the following scenarios:

1. Using an existing index
2. Creating a new index with 'Import Data' and adding the vector field during the setup

While the "Import and Vectorize Data" feature works, it only supports Blob Storage/ADLS. However, my goal is to implement it with the hotels-sample dataset.

I've already created an embedding model in Azure OpenAI Studio, which integrates successfully when used for solutions with Vector Search—whether indexing via OpenAI Studio or through the "Import and Vectorize Data" option.

Scenario 1:

I create an index with "Import Data" and select hotels-sample, then add a vector field.

Then I add a vector field

However, the vector index size remains at 0 bytes, and no results are returned.

When I disable vector search, the search returns results correctly.

Scenario 2:

I follow the same process but do not add the vector field during the initial creation. Instead, I add it afterward. Unfortunately, the outcome is the same—adding the vector field does not seem to impact the search results.

Could you advise on the correct approach for adding a vector field in these scenarios?

Answer 1

AK,

Based on my understanding of your scenario, the "import data" wizard only works for non-vector scenarios.
You can import to your Azure Blob Container the data from hotels, this is here: Azure-Samples/azure-search-sample-data: Sample data shared across several Azure Search sample code projects. (github.com) - which is the one that is in the hotels-sample data and use the import and vectorize data.

Just to highlight -adding just a field for a vector won't work since behind the scenes the indexer works with several processes (set in the skillset configuration (please refer this doc for more info Skillset concepts - Azure AI Search | Microsoft Learn)) that will :: get the captured data, chunk it in smaller parts with the split skill, vectorize using an AOAI embedding model provided by you and then mapping the vectorized chunks and original chunks via index projections. 

 

For better insight into the flow of what the import and vectorize data does under the cover can be found sample (here): https://github.com/Azure/azure-search-vector-samples/tree/main/demo-python/code/integrated-vectorization - You may also use this script and modify for your own scenario.

Once again, apologies for the delayed response. Much appreciate your patience!

---

 If the answer helped (pointed you in the right direction) > please click Accept Answer Or please share the requested/more info to help you better.