Azure AI Search
An Azure search service with built-in artificial intelligence capabilities that enrich information to help identify and explore relevant content at scale.
Whats the Azure OpenAI Inference API Version used by AzureOpenAIEmbeddingSkill
Jay
86
Reputation points
Reading this doc but the document doesn't have any info or a property to specify the Azure OpenAI Inference API Version details. So the question is which API version it will use ? One can find the available API Versions from https://github.com/Azure/azure-rest-api-specs/tree/main/specification/cognitiveservices/data-plane/AzureOpenAI/inference
Ideally I thought there should be a OpenAI API Property which i can use to pass the api version too along with the modelname, deploymentname and openai instance name
Azure AI Search
Azure OpenAI Service
Azure OpenAI Service
An Azure service that provides access to OpenAI’s GPT-3 models with enterprise capabilities.
{count} votes
-
Daniel Fang 1,060 Reputation points MVP2024-10-28T10:01:36.3466667+00:00 HI Jay
The link you have in [Reading this doc] points to Azure OpenAI Embedding skill. The purpose of the skill is to generate embeddings during indexing for your documents. You can specify the Azure OpenAI Embedding model using the parameters in the [Skill parameters] section like modelName, deploymentId, resourceUri.
The skill only produces embedding that is saved to a vector database like Azure AI Search. At this step, Azure OpenAI Inference API (like GPT model) is not yet invoked. Normally, inference api is invoked during the user interaction. For example, user asks a question, the question is converted into embedding which is then used to do vector search against the saved embedding in the vector db. The original question and retrieved documents (based on vector search) are finally passed onto inference api to produce final response to user.
-
Jay 86 Reputation points
2024-10-28T12:12:08.9+00:00 Hi @Daniel Fang , Thanks for your answer. I could not able to understand your answer; how could the embeddings be generated without calling the inference APIs?
If there is no need to invoke Inference API (like the GPT Model) then why we need to pass the modelname, deploymentID and resourceUri ?
The same doc also mentioned like
This skill is bound to Azure OpenAI and is charged at the existing Azure OpenAI pay-as-you go price.
This means my assumption is the skillset config making azure openai inference API calls behind the screen to generate the embeddings
P.S: Wherever I refer inference API I mean the embedding inference API
i.e,
/deployments/{deployment-id}/embeddings -
Daniel Fang 1,060 Reputation points MVP
2024-10-29T02:26:32.2133333+00:00 Hi Jay
The Azure OpenAI Embedding skill is the right side of the diagram below. It mainly used to ingest documents into AI Search, it uses an embedding model like ada (but not GPT model), however, it is only used to generate embedding.
To specify the endpoints of Embedding model, you can use these parameters. The embedding model like ada will be under their azure openai service.
Then on the left side, your client will need to use GPT model to generate response.
more details here:
https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview
-
Daniel Fang 1,060 Reputation points MVP
2024-10-29T02:32:15.16+00:00 If there is no need to invoke Inference API (like the GPT Model) then why we need to pass the modelname, deploymentID and resourceUri ?
Sorry, my bad. In this case, we can't use a GPT model, it needs to be an embedding model like ada.
This means my assumption is the skillset config making azure openai inference API calls behind the screen to generate the embeddings. i.e,
/deployments/{deployment-id}/embeddings
The embedding skill will use the configuration you provided to invoke Azure OpenAI service endpoint. In your model deployment, use the ada model (not GPT model) for embedding generation. GPT-4 is not for embedding generation. -
Jay 86 Reputation points
2024-10-29T07:27:19.2133333+00:00 ok now i understood where the confusion arise ! From the beginning itself i was referring to embeddings inference API only not the traditional GPT completion models.
So now coming back to my original question, If anyone wants to call the
/deployments/{deployment-id}/embeddingsthen they need to pass theapi-versionas a query parameter.My question: What is the api-version used by the skillset config ? Ideally as part of the parameters to skillset along with the
modelname,deploymentIDandresourceUriwe also need to pass the api-version e.g2024-06-01But there is no such property exist. So whats the value will be used by skillset then ?
-
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(185.6, 59%, 27%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3ESM%3C/text%3E%3C/svg%3E)
SnehaAgrawal-MSFT 22,721 Reputation points Moderator2024-10-30T10:25:20.3966667+00:00 @Jay If I understood right your query is for api-version used by the skillset ? You may refer to this document for latest version- https://learn.microsoft.com/en-us/rest/api/searchservice/get-skillset
-
Jay 86 Reputation points
2024-10-30T11:49:30.94+00:00 @SnehaAgrawal-MSFT No, my question is what is the
api-versionused by skillset internally to call the azure openai embedding APIs ? -
Jay 86 Reputation points
2024-10-30T11:51:26.2566667+00:00 No Sneha, MY question is what is the
api-versionused internally by skillset config to call the embedding apis -
Daniel Fang 1,060 Reputation points MVP
2024-11-02T21:30:36.5766667+00:00 Hi Jay
Understand your question, there is no configuration for the skillset at the moment to specify api_version. My best guess is that it will use the latest api version based on the deployment.
Of note, the api_version is only related to the version of restful API call, it should not normally change the quality of the response from embedding. The api_version is different to the model version as shown below. Model version is specified when creating the model deployment.
-
Jay 86 Reputation points
2024-11-05T05:35:26.84+00:00 Hi Daniel,
Yes, I would like to make sure the guess is correct :) Currently, there is no official documentation on this. Hence, I would like to seek Microsoft officials to fill this gap! Especially the documentation should cover this information. The purpose of this question is that the official doc "doesn't clearly state the lack of an API version property in the Azure Cognitive Search skillset for Azure OpenAI skills, nor does it provide clarity on how the API version is managed "
Sign in to comment
Sign in to answer