effect-ai-language-model

Use LanguageModel.generateObject with an Effect Schema defining the expected shape. The response's value field contains the schema-validated object, and you can optionally set objectName to aid model understanding.

Use LanguageModel.streamText, which returns an Effect Stream of typed parts such as text-delta, reasoning-delta, and tool-call. Consume it with Stream.runForEach or collect text deltas with Stream.filter and Stream.runFold.

Catch failures with Effect.catchTag("AiError", ...) and match on error.reason._tag for specific cases like RateLimitError, AuthenticationError, or StructuredOutputError. Avoid constructing generic Error instances; use AiError.make with tagged reasons.

Static accessors like LanguageModel.generateText add LanguageModel to the effect's R requirement. To avoid this, yield the LanguageModel.LanguageModel tag during layer construction and capture it in the closure, keeping service method signatures clean.

Yes. Use ExecutionPlan from effect/ExecutionPlan to define multi-provider fallback with per-provider retry attempts, such as Anthropic with three attempts followed by OpenAI with two. You can also read the active provider via Model.ProviderName.

Set the toolChoice option: "auto" lets the model decide, "none" forces text-only, "required" forces a tool call, and { tool: "name" } or oneOf restricts which tools are available. Use disableToolCallResolution to receive tool calls without executing them.