@thomasfosterau/effect-ai-cohere

A first-party Effect AI provider for Cohere. Effect ships no @effect/ai-cohere package, so this implements each capability directly against Cohere's v2 API.

• CohereClient — the low-level HTTP client for /v2/chat (regular + streaming), /v2/embed, and /v2/rerank, every failure mapped to AiError. Its apiUrl override is the BYOK routing contract used by the Cloudflare AI Gateway package.
• CohereLanguageModel — the LanguageModel implementation over Command chat models: prompt / tool / structured-output conversion and response decoding, exposed through layer, model, and make.
• CohereEmbeddingModel — the EmbeddingModel implementation over embed-v4.0 (pinned model + 1024-dim output; retrieval-aware inputType).
• CohereRerankerModel — the effect-ai-ext RerankerModel implementation over rerank-v3.5.
• CohereClassifierModel — the effect-ai-ext ClassifierModel implementation, backed by Cohere chat + structured output.

Install

pnpm add @thomasfosterau/effect-ai-cohere effect

effect (>=4.0.0-rc.118) is a peer dependency (it must be a singleton).

Usage

import { Effect, Layer, Redacted } from "effect";
import { FetchHttpClient } from "effect/http";
import { LanguageModel } from "effect/ai";
import { CohereClient, CohereLanguageModel } from "@thomasfosterau/effect-ai-cohere";

const ModelLayer = CohereLanguageModel.layer({ model: "command-a-03-2025" }).pipe(
  Layer.provide(CohereClient.layer({ apiKey: Redacted.make(process.env.COHERE_API_KEY!) })),
  Layer.provide(FetchHttpClient.layer),
);

const program = LanguageModel.generateText({ prompt: "Explain Effect in one line" }).pipe(
  Effect.map((response) => response.text),
  Effect.provide(ModelLayer),
);

CohereClient.layer accepts an apiUrl override — the same client can be repointed at the Cloudflare AI Gateway route for BYOK. Structured output uses Cohere's native response_format JSON-schema mode; tool-call parameters are returned as parsed JSON and validated against the toolkit. Streaming is available via LanguageModel.streamText.

Embeddings, reranking, classification

import { CohereClient, CohereEmbeddingModel, CohereRerankerModel } from "@thomasfosterau/effect-ai-cohere";
import { EmbeddingModel } from "effect/ai";
import { RerankerModel } from "@thomasfosterau/effect-ai-ext";

const EmbedLayer = CohereEmbeddingModel.layer({ inputType: "search_document" }).pipe(
  Layer.provide(CohereClient.layer({ apiKey: Redacted.make(process.env.COHERE_API_KEY!) })),
  Layer.provide(FetchHttpClient.layer),
);

A note on classification

Cohere's dedicated Classify surface is v1-legacy: there is no v2/classify, and /v1/classify now requires a fine-tuned Embed model (default-model classification was deprecated 2025-01-31). CohereClassifierModel therefore backs the ClassifierModel abstraction with Cohere chat + structured output over a Command model — covering the full ClassifyOptions contract (caller label sets, open-set, multi-label, example-shot) on a supported v2 surface. See CohereClassifierModel's module docs for the full rationale.

License

MIT