A language model can turn a customer’s review into book recommendations, summarize a new arrival, or produce conversational text. Azure OpenAI Service makes these models available to applications through a deployed endpoint. The practical question is not simply which model is most powerful, but which one fits the job. section 7.1 section 7.6
Choose and test a model against the work it will actually do:
- Balance quality, cost, and speed. A smaller model may suit quick recommendations from short reviews; a larger one may produce a more nuanced, polished summary at greater expense. section 7.1 section 7.4.3
- Respect the limit on text the model can view at once. Summarize long inputs, split them into chunks, or retrieve only relevant passages. section 7.1
- Compare candidates in the AI Foundry Model catalog using capabilities, versions, pricing, and benchmarks. Prototype with review data, check quality and fairness, then consider a different model or fine-tuning if needed. section 7.4.3
Recommendations from a review
A customer praises the dystopian setting and social commentary in *1984* and asks for similar books. A prompt asks the deployed model for three recommendations, each with a brief description. This is a small, immediate task; a long summary of a rare technical book calls for a different quality–cost decision. section 7.1
C:/> az rest --method post --url "${ENDPOINT}openai/deployments/${DEPLOYMENT_NAME}/completions?
api-version=${API_VERSION}" --headers "api-key=${API_KEY}"
"Content-Type=application/json" --body "{\"prompt\": \"${PROMPT}\",
\"max_tokens\": 150, \"temperature\": 0.7, \"n\": 1}"
An Azure CLI REST call sends the prompt to a named deployment. The request also sets output parameters, including a 150-token maximum. section 7.1
