Create a RAG-based client application

When you've created an Azure AI Search index for your contextual data, you can use it with an OpenAI model. To ground prompts with data from your index, the Azure OpenAI SDK supports extending the request with connection details for the index.

The following Python code example shows how to implement this pattern.

from openai import AzureOpenAI

# Get an Azure OpenAI chat client
chat_client = AzureOpenAI(
    api_version = "2024-12-01-preview",
    azure_endpoint = open_ai_endpoint,
    api_key = open_ai_key
)

# Initialize prompt with system message
prompt = [
    {"role": "system", "content": "You are a helpful AI assistant."}
]

# Add a user input message to the prompt
input_text = input("Enter a question: ")
prompt.append({"role": "user", "content": input_text})

# Additional parameters to apply RAG pattern using the AI Search index
rag_params = {
    "data_sources": [
        {
            "type": "azure_search",
            "parameters": {
                "endpoint": search_url,
                "index_name": "index_name",
                "authentication": {
                    "type": "api_key",
                    "key": search_key,
                }
            }
        }
    ],
}

# Submit the prompt with the index information
response = chat_client.chat.completions.create(
    model="<model_deployment_name>",
    messages=prompt,
    extra_body=rag_params
)

# Print the contextualized response
completion = response.choices[0].message.content
print(completion)

In this example, the search against the index is keyword-based - in other words, the query consists of the text in the user prompt, which is matched to text in the indexed documents. When using an index that supports it, an alternative approach is to use a vector-based query in which the index and the query use numeric vectors to represent text tokens. Searching with vectors enables matching based on semantic similarity as well as literal text matches.

To use a vector-based query, you can modify the specification of the Azure AI Search data source details to include an embedding model; which is then used to vectorize the query text.

rag_params = {
    "data_sources": [
        {
            "type": "azure_search",
            "parameters": {
                "endpoint": search_url,
                "index_name": "index_name",
                "authentication": {
                    "type": "api_key",
                    "key": search_key,
                },
                # Params for vector-based query
                "query_type": "vector",
                "embedding_dependency": {
                    "type": "deployment_name",
                    "deployment_name": "<embedding_model_deployment_name>",
                },
            }
        }
    ],
}