Predict with Models API
The predictive endpoints of the Models API run a predictive model against your inputs and return a structured prediction. Use these endpoints when you need a numeric estimate, a category label, or a forecast rather than generated language.
Predictive endpoints return structured predictions based on your business data, such as a numeric value, probability, label, or forecast. Use a predictive model that you built using AI Models (formerly Einstein Studio) in Data 360, and then call it by its API name. Unlike generative endpoints, predictive endpoints don’t use large language models (LLMs) to generate chat responses, text, or embeddings.
Use a predictive endpoint when:
- You need structured output, such as a score, class probability, label, or forecasted value, that your application can process.
- You have historical or structured data and a model trained to score churn risk, rank leads, forecast demand, or classify records at scale.
- You want predictions based on your business data rather than language generated by an LLM.
Use a generative endpoint when you need to generate or understand natural language, such as drafting a reply, summarizing a case, answering a question, or creating embeddings for semantic search.
| Generative endpoints | Predictive endpoints | |
|---|---|---|
| Output | Natural-language text, chat, or embeddings | A numeric value, class probability, label, or forecast |
| Backed by | Salesforce-managed or bring-your-own LLMs | A predictive model you build or bring with AI Models in Data 360 |
| You provide | A prompt or messages | A deployed model’s API name plus input records |
| Best for | Drafting, summarizing, Q&A, semantic search | Scoring, ranking, forecasting, and classifying your own data |
| Result | Varies by prompt | Repeatable, structured predictions |
- Classification — Assign category labels to an input. Includes classification (probability per class), multi-outcome, topic classification, and sentiment detection.
- Regression — Predict a continuous numeric value for an input, including forecasting future values of a time series.