ai-provider-cohere-sdk

Cohere SDK Patterns

Quick Guide: Use the cohere-ai npm package with CohereClientV2 for all new Cohere integrations. V2 API requires model on every call. Use chatStream for streaming with content-delta events. Embeddings require inputType matching your use case (search_document for indexing, search_query for querying). Rerank scores documents by relevance. RAG works by passing documents to chat() -- the model returns inline citations automatically. Tool use follows a 4-step loop: user message, model returns tool_calls, you execute and return results, model generates cited response.

---

<critical_requirements>

CRITICAL: Before Using This Skill

All code must follow project conventions in CLAUDE.md (kebab-case, named exports, import ordering, import type, named constants)

(You MUST use CohereClientV2 (not CohereClient) for all new code -- V2 is the current API with required model parameter)

(You MUST specify inputType on every embed call -- search_document for indexing, search_query for querying -- mismatched types produce garbage similarity scores)

(You MUST handle the tool use loop correctly: append the full assistant message (with tool_calls) to messages, then append tool role results with matching tool_call_id)

(You MUST check finish_reason in responses -- MAX_TOKENS means the output was truncated)

(You MUST never hardcode API keys -- pass via token constructor parameter sourced from environment variables)

</critical_requirements>

---

Auto-detection: Cohere, cohere-ai, CohereClientV2, CohereClient, command-a, command-r, command-r-plus, embed-v4, rerank-v4, chatStream, content-delta, inputType, search_document, search_query, embeddingTypes, topN, CO_API_KEY, COHERE_API_KEY

When to use:

• Building applications with Cohere Command models (chat, generation, summarization)
• Creating semantic search pipelines with Cohere embeddings
• Adding relevance scoring to search results with Cohere Rerank
• Implementing RAG with inline document grounding and automatic citations
• Building agentic workflows with Cohere tool use / function calling
• Streaming chat responses for real-time user interfaces

Key patterns covered:

• Client setup with CohereClientV2 (token, timeout, platform configs)
• Chat and streaming (chat, chatStream, event types)
• Embeddings with inputType for search/classification/clustering
• Rerank for relevance scoring and search result ordering
• RAG with documents and automatic citation handling
• Tool use / function calling with multi-step loops
• Model selection (Command-A, Command-R, Embed v4, Rerank v4)

When NOT to use:

• Multi-provider applications needing OpenAI/Anthropic/Google switching -- use a unified provider SDK
• React-specific chat UI hooks -- use a framework-integrated AI SDK
• Simple text completion without Cohere-specific features (rerank, citations)

---

Examples Index

• Core: Setup, Chat & Error Handling -- CohereClientV2 init, basic chat, streaming, error handling
• Embeddings & Rerank -- Semantic search, input types, rerank scoring, RAG pipeline
• Tool Use & RAG -- Function calling, document grounding, citation handling
• Quick API Reference -- Model IDs, method signatures, event types, error classes

---

Philosophy

The Cohere TypeScript SDK (cohere-ai) provides direct access to Cohere's API surface -- chat, embeddings, rerank, and RAG with citations. The SDK is auto-generated from Cohere's API spec using Fern.

Core principles:

1. V2 API is current -- CohereClientV2 provides the modern API. model is required on every call. V1 methods on CohereClient are legacy.
2. Embeddings are typed -- The inputType parameter (search_document, search_query, classification, clustering) is mandatory for v3+ models. Mismatching input types between indexing and querying silently degrades results.
3. RAG is first-class -- Pass documents directly to chat() and the model returns grounded answers with inline citations. No external retrieval framework required for the grounding step.
4. Rerank is a standalone primitive -- Score and reorder search results without building a full RAG pipeline. Feed any list of documents and a query, get relevance scores back.
5. Citations are automatic -- When documents are provided (via RAG or tool results), the model generates fine-grained citations with start/end positions and source references.

When to use the Cohere SDK directly:

• You want Cohere-specific features: rerank, citation grounding, multilingual embeddings
• You need semantic search with embed + rerank pipeline
• You want RAG with automatic inline citations
• You are building on Cohere's platform (or Bedrock/Azure/OCI with Cohere models)

When NOT to use:

• You need to switch between multiple LLM providers -- use a unified provider SDK
• You want React-specific chat UI hooks -- use a framework-integrated AI SDK
• You only need basic chat completion without Cohere differentiators

---

Core Patterns

Pattern 1: Client Setup

Initialize CohereClientV2. The token parameter is required (pass from environment).

// lib/cohere.ts -- basic setup
import { CohereClientV2 } from "cohere-ai";

const client = new CohereClientV2({
  token: process.env.CO_API_KEY,
});

export { client };

// lib/cohere.ts -- production configuration
const TIMEOUT_MS = 30_000;

const client = new CohereClientV2({
  token: process.env.CO_API_KEY,
  timeout: TIMEOUT_MS,
});

Why good: Explicit token from env var, named timeout constant, named export

// BAD: Hardcoded key, default CohereClient (V1)
import { CohereClient } from "cohere-ai";
const client = new CohereClient({ token: "sk-abc123" });

Why bad: Hardcoded API key is a security breach risk, CohereClient is the legacy V1 client

See: examples/core.md for error handling, platform configs (Bedrock, Azure)

---

Pattern 2: Chat Completion

V2 chat uses messages array with system, user, assistant, and tool roles.

const response = await client.chat({
  model: "command-a-03-2025",
  messages: [
    { role: "system", content: "You are a helpful coding assistant." },
    { role: "user", content: "Explain TypeScript generics." },
  ],
});

console.log(response.message.content[0].text);

Why good: System message for instruction, model explicitly specified, correct V2 content access path

// BAD: Missing model (required in V2), wrong response access
const response = await client.chat({
  messages: [{ role: "user", content: "Hello" }],
});
console.log(response.text); // WRONG: V2 uses response.message.content[0].text

Why bad: V2 requires model, response shape is response.message.content[0].text not response.text

See: examples/core.md for multi-turn, token tracking, temperature control

---

Pattern 3: Streaming

Use chatStream with for await and check event type for content-delta.

const stream = await client.chatStream({
  model: "command-a-03-2025",
  messages: [{ role: "user", content: "Explain async/await." }],
});

for await (const event of stream) {
  if (event.type === "content-delta") {
    process.stdout.write(event.delta?.message?.content?.text ?? "");
  }
}

Why good: Checks event type before accessing delta, handles nullable content safely

// BAD: Not checking event type
for await (const event of stream) {
  console.log(event.delta?.message); // Many events don't have message delta
}

Why bad: Only content-delta events have text content -- other events (message-start, citation-start, tool-plan-delta) have different shapes

See: examples/core.md for full streaming with all event types

---

Pattern 4: Embeddings

inputType is required for v3+ models. Mismatching types between indexing and querying silently degrades results.

const EMBEDDING_MODEL = "embed-v4.0";

// Index documents with search_document
const docEmbeddings = await client.embed({
  model: EMBEDDING_MODEL,
  inputType: "search_document",
  texts: ["TypeScript is a typed superset of JavaScript."],
  embeddingTypes: ["float"],
});

// Query with search_query
const queryEmbedding = await client.embed({
  model: EMBEDDING_MODEL,
  inputType: "search_query",
  texts: ["What is TypeScript?"],
  embeddingTypes: ["float"],
});

Why good: Correct inputType pairing, embeddingTypes explicitly specified, named model constant

// BAD: Same inputType for both indexing and querying
const docs = await client.embed({
  model: "embed-v4.0",
  inputType: "search_query", // WRONG for documents
  texts: documents,
  embeddingTypes: ["float"],
});

Why bad: Using search_query for document indexing silently produces worse similarity scores -- documents must use search_document

See: examples/embeddings-rerank.md for cosine similarity, dimension control, batch embedding

---

Pattern 5: Rerank

Score documents by relevance to a query. Returns ordered results with relevance scores.

const RERANK_MODEL = "rerank-v4.0-pro";
const TOP_N = 3;

const result = await client.rerank({
  model: RERANK_MODEL,
  query: "What is TypeScript?",
  documents: [
    "TypeScript is a typed superset of JavaScript.",
    "Python is a general-purpose language.",
    "TypeScript compiles to JavaScript.",
  ],
  topN: TOP_N,
});

for (const item of result.results) {
  console.log(`Doc ${item.index}: score ${item.relevanceScore}`);
}

Why good: Named constants, topN limits results, accesses index and relevanceScore

See: examples/embeddings-rerank.md for embed + rerank pipeline, rank fields

---

Pattern 6: RAG with Documents

Pass documents to chat() and the model returns grounded answers with inline citations.

const response = await client.chat({
  model: "command-a-03-2025",
  messages: [{ role: "user", content: "What is TypeScript?" }],
  documents: [
    {
      data: {
        text: "TypeScript is a typed superset of JavaScript.",
        title: "TS Docs",
      },
    },
    {
      data: {
        text: "TypeScript was developed by Microsoft.",
        title: "History",
      },
    },
  ],
});

console.log(response.message.content[0].text);

// Citations reference which documents support each claim
if (response.message.citations) {
  for (const citation of response.message.citations) {
    console.log(`"${citation.text}" from doc ${citation.sources}`);
  }
}

Why good: Documents passed inline with metadata, citations accessed from response, no external retrieval framework needed

See: examples/tools-rag.md for full RAG pipeline with embed + rerank + chat

---

Pattern 7: Tool Use / Function Calling

4-step loop: user message -> model returns tool_calls -> execute tools -> return results with tool_call_id.

const tools = [
  {
    type: "function" as const,
    function: {
      name: "get_weather",
      description: "Get weather for a city",
      parameters: {
        type: "object",
        properties: {
          location: { type: "string", description: "City name" },
        },
        required: ["location"],
      },
    },
  },
];

const response = await client.chat({
  model: "command-a-03-2025",
  messages: [{ role: "user", content: "Weather in Paris?" }],
  tools,
});

// Check if model wants to call tools
if (response.message.toolCalls) {
  // See examples/tools-rag.md for the complete tool execution loop
}

Why good: Standard JSON Schema tool definition, checks for toolCalls before executing

See: examples/tools-rag.md for complete multi-step tool loop with tool result submission

---

Pattern 8: Error Handling

Catch CohereError for API errors, CohereTimeoutError for timeouts.

import { CohereError, CohereTimeoutError } from "cohere-ai";

try {
  const response = await client.chat({
    model: "command-a-03-2025",
    messages: [{ role: "user", content: "Hello" }],
  });
} catch (error) {
  if (error instanceof CohereTimeoutError) {
    console.error("Request timed out");
  } else if (error instanceof CohereError) {
    console.error(`API Error [${error.statusCode}]: ${error.message}`);
    console.error("Body:", error.body);
  } else {
    throw error; // Re-throw unknown errors
  }
}

Why good: Specific error types with status codes, re-throws unexpected errors, timeout handled separately

See: examples/core.md for production error handling patterns

---

Performance Optimization

Model Selection for Cost/Speed

General purpose (best)      -> command-a-03-2025 (256K context, strongest)
Reasoning tasks             -> command-a-reasoning-08-2025 (multi-step reasoning)
Vision/document analysis    -> command-a-vision-07-2025 (images, charts, OCR)
Translation                 -> command-a-translate-08-2025 (23 languages)
Lightweight / edge          -> command-r7b-12-2024 (7B, fast, 128K context)
Legacy (still supported)    -> command-r-08-2024, command-r-plus-08-2024
Embeddings (best)           -> embed-v4.0 (multimodal, 128K context, flexible dims)
Embeddings (English)        -> embed-english-v3.0 (1024 dims)
Embeddings (multilingual)   -> embed-multilingual-v3.0 (23 languages)
Rerank (quality)            -> rerank-v4.0-pro (32K context, multilingual)
Rerank (speed)              -> rerank-v4.0-fast (32K context, latency-optimized)

Key Optimization Patterns

• Batch embeddings -- pass up to 96 texts per embed() call instead of calling per-document
• Use topN in rerank -- limit results to reduce response size and cost
• Use outputDimension with embed-v4 -- reduce dimensions (256/512/1024) for faster similarity search at minimal quality loss
• Check finish_reason === "MAX_TOKENS" -- detect truncated output
• Use temperature: 0 for deterministic output (enables caching)
• Use embed-v4 int8/binary types for compressed storage with minimal quality loss
• Use strictTools: true to force tool calls to follow the schema exactly (structured outputs)
• Use thinking: { type: "enabled" } with reasoning models for complex multi-step tasks
• Use toolChoice: "REQUIRED" when you always want the model to call a tool (command-r7b+ only)

---

<decision_framework>

Decision Framework

Which Client Class to Use

New project?
+-- YES -> CohereClientV2 (always)
+-- Existing V1 code?
    +-- Working fine? -> Keep CohereClient but plan migration
    +-- Need V2 features? -> Migrate to CohereClientV2

Which Model to Choose

What is your task?
+-- General chat/generation -> command-a-03-2025 (most capable)
+-- Reasoning / multi-step -> command-a-reasoning-08-2025
+-- Image/document analysis -> command-a-vision-07-2025
+-- Translation -> command-a-translate-08-2025
+-- Lightweight / low latency -> command-r7b-12-2024
+-- Embeddings -> embed-v4.0 (or embed-english-v3.0 for English-only)
+-- Rerank quality -> rerank-v4.0-pro
+-- Rerank speed -> rerank-v4.0-fast

Embed inputType Selection

What are you embedding?
+-- Documents for a search index -> "search_document"
+-- Search queries against an index -> "search_query"
+-- Text for a classifier -> "classification"
+-- Text for clustering -> "clustering"
+-- Images -> "image" (embed-v4+ only)

When to Use Rerank

Do you have search results to re-order?
+-- YES -> Use rerank as a second-stage ranker
|   +-- Quality matters most? -> rerank-v4.0-pro
|   +-- Latency matters most? -> rerank-v4.0-fast
+-- NO -> Not applicable (rerank needs existing results to score)

RAG Approach

Do you need grounded answers with citations?
+-- YES -> Pass documents to chat()
|   +-- Have pre-retrieved documents? -> Pass directly via documents param
|   +-- Need retrieval first? -> Use embed + vector search + rerank pipeline, then pass top results to chat()
+-- NO -> Use plain chat without documents

</decision_framework>

---

<red_flags>

RED FLAGS

High Priority Issues:

• Using CohereClient instead of CohereClientV2 for new code (V1 is legacy)
• Missing model parameter in V2 API calls (required on every call, unlike V1)
• Using wrong inputType for embeddings (search_query for documents or vice versa -- silently degrades results)
• Hardcoding API keys instead of using environment variables
• Not appending the full assistant message (with tool_calls) before appending tool results in the tool use loop

Medium Priority Issues:

• Not specifying embeddingTypes (defaults may not match your storage format)
• Ignoring finish_reason: "MAX_TOKENS" (output was silently truncated)
• Not handling CohereTimeoutError separately from CohereError
• Processing all stream events without checking type (only content-delta has text)
• Using V1 parameter names (preamble, connectors, conversation_id) with V2 client

Common Mistakes:

• Accessing response.text instead of response.message.content[0].text (V2 response shape changed)
• Forgetting that embeddingTypes is required in V2 Embed API
• Not matching tool_call_id when submitting tool results (model cannot correlate results)
• Using documents with string values instead of { data: { text: "..." } } objects in V2
• Expecting response.message.citations to exist when no documents were provided (citations only appear with grounded responses)

Gotchas & Edge Cases:

• The SDK is in beta -- pin your cohere-ai version in package.json to avoid breaking changes
• V2 API is NOT yet supported for cloud deployments (Bedrock, SageMaker, Azure, OCI) -- use V1 client for cloud platforms
• inputType is camelCase in TypeScript SDK (inputType) but snake_case in the REST API (input_type)
• Embed API accepts max 96 texts per call -- batch larger sets yourself
• embed-v4.0 supports outputDimension for flexible sizing (256, 512, 1024, 1536) but v3 models have fixed dimensions
• Rerank relevanceScore is normalized 0-1 but not calibrated across queries -- compare scores within a single query only
• Stream events include tool-plan-delta before tool-call-start -- the model's reasoning about which tool to call
• V2 uses system role for instructions (V1 used preamble parameter)
• Citation sources in tool use responses reference tool_call_id values, not document indices
• The clientName constructor parameter is for logging/analytics, not authentication
• responseFormat: { type: "json_object" } is NOT supported in RAG mode (with documents, tools, or toolResults)
• toolChoice is only supported on command-r7b-12-2024 and newer models
• First requests with strictTools: true and a new tool set take longer (schema compilation)
• thinking (reasoning mode) is only available on reasoning-capable models like command-a-reasoning-08-2025

</red_flags>

---

<critical_reminders>

CRITICAL REMINDERS

All code must follow project conventions in CLAUDE.md (kebab-case, named exports, import ordering, import type, named constants)

(You MUST use CohereClientV2 (not CohereClient) for all new code -- V2 is the current API with required model parameter)

(You MUST specify inputType on every embed call -- search_document for indexing, search_query for querying -- mismatched types produce garbage similarity scores)

(You MUST handle the tool use loop correctly: append the full assistant message (with tool_calls) to messages, then append tool role results with matching tool_call_id)

(You MUST check finish_reason in responses -- MAX_TOKENS means the output was truncated)

(You MUST never hardcode API keys -- pass via token constructor parameter sourced from environment variables)

Failure to follow these rules will produce broken embeddings, missing citations, or insecure AI integrations.

</critical_reminders>