Generate embeddings for search queries and documents
Azure AI Search doesn't host vectorization models, so one of your challenges is creating embeddings for query inputs and outputs. You can use any supported embedding model, but this article assumes Azure OpenAI embedding models for the steps.
We recommend integrated vectorization, which provides built-in data chunking and vectorization. Integrated vectorization takes a dependency on indexers, skillsets, and built-in or custom skills that point to a model that executes externally from Azure AI Search.
If you want to handle data chunking and vectorization yourself, we provide demos in the sample repository that show you how to integrate with other community solutions.
How embedding models are used in vector queries
Query inputs are either vectors, or text or images that are converted to vectors during query processing. The built-in solution in Azure AI Search is to use a vectorizer.
Alternatively, you can also handle the conversion yourself by passing the query input to an embedding model of your choice. To avoid rate limiting, you can implement retry logic in your workload. For the Python demo, we used tenacity.
Query outputs are any matching documents found in a search index. Your search index must have been previously loaded with documents having one or more vector fields with embeddings. Whatever embedding model you used for indexing, use that same model for queries.
Create resources in the same region
Integrated vectorization requires resources to be in the same region:
Generate an embedding for an improvised query
The following Python code generates an embedding that you can paste into the "values" property of a vector query.
!pip install openai
import openai
openai.api_type = "azure"
openai.api_key = "YOUR-API-KEY"
openai.api_base = "https://YOUR-OPENAI-RESOURCE.openai.azure.com"
openai.api_version = "2024-02-01"
response = openai.Embedding.create(
input="How do I use Python in VSCode?",
engine="text-embedding-ada-002"
)
embeddings = response['data'][0]['embedding']
print(embeddings)
Output is a vector array of 1,536 dimensions.
Tips and recommendations for embedding model integration
Identify use cases: Evaluate the specific use cases where embedding model integration for vector search features can add value to your search solution. This can include multimodal or matching image content with text content, multilingual search, or similarity search.
Design a chunking strategy: Embedding models have limits on the number of tokens they can accept, which introduces a data chunking requirement for large files. For more information, see Chunk large documents for vector search solutions.
Optimize cost and performance: Vector search can be resource-intensive and is subject to maximum limits, so consider only vectorizing the fields that contain semantic meaning. Reduce vector size so that you can store more vectors for the same price.
Choose the right embedding model: Select an appropriate model for your specific use case, such as word embeddings for text-based searches or image embeddings for visual searches. Consider using pretrained models like text-embedding-ada-002 from OpenAI or Image Retrieval REST API from Azure AI Computer Vision.
Normalize Vector lengths: Ensure that the vector lengths are normalized before storing them in the search index to improve the accuracy and performance of similarity search. Most pretrained models already are normalized but not all.
Fine-tune the model: If needed, fine-tune the selected model on your domain-specific data to improve its performance and relevance to your search application.
Test and iterate: Continuously test and refine your embedding model integration to achieve the desired search performance and user satisfaction.