AI models¶
Many Agate steps use AI models to read text, identify details, or create semantic embeddings. The AI models settings create an approved catalog so flow builders choose named configurations rather than entering provider and billing details on every node.
Custom configurations use LiteLLM-compatible model identifiers. See the LiteLLM model catalog for available providers and models.
The model catalog¶
Organization administrators can add presets or define custom model configurations. Presets are generated from the LiteLLM model catalog so current flagship provider models appear without waiting for a product release. Preset lists emphasize recent generative models (for example GPT-5 and newer OpenAI offerings) and omit retired catalog entries such as GPT-4 presets. A catalog entry records:
- a readable name shown to flow builders;
- the provider model identifier;
- whether it is generative, embedding-oriented, or a decision model used for structured enrichment lookups (such as Person Enrich and Organization Enrich);
- capabilities that determine which nodes may use it;
- the saved provider credential or endpoint;
- optional input and output prices used for estimates;
- whether the configuration is active or disabled.
Disabling a configuration removes it from new choices without erasing the history of runs that used it.
Decision models¶
Decision models are a separate catalog kind from generative extraction models. They power structured lookup steps—such as choosing a Wikidata match during person or organization enrichment—and do not appear in ordinary language-model pickers on extractors or Article Metadata.
Organization administrators add decision models in Settings → AI models, assign credentials, and test the connection the same way as other model types. Projects enable the decision models their flows need and can set a default for that role on the Models tab.
Organization and project choices¶
- Administrators define the organization catalog and save provider credentials.
- Projects inherit approved models, can enable the subset they use, and can choose defaults for supported roles.
- Flow builders see only compatible, enabled choices in each node's settings.
Some deployments allow project-specific provider credentials. These let a project use its own billing boundary while keeping the model configuration under organization control.
Testing and credentials¶
Saved credentials are not displayed after entry. Administrators can see whether a credential is configured, replace or remove it, and test a model connection before teams depend on it in a flow.
Removing a credential can affect every model linked to it, so check the linked model list before changing a shared credential.
Cost estimates¶
Backfield records the estimated cost of each run, broken down by step, so teams can compare the estimated model cost of their flows.
Estimates use the prices stored with the model configuration and the usage reported during execution. They are useful for comparison and monitoring, but the provider's invoice remains the authoritative billing record.