Cohere RAG Pipeline (Core Workflow A)

Overview

End-to-end Retrieval-Augmented Generation using Cohere's three core endpoints: Embed (vectorize), Rerank (sort by relevance), Chat (generate grounded answer with citations).

Prerequisites

• Completed cohere-install-auth setup
• cohere-ai package installed
• Understanding of vector similarity search

Instructions

Step 1: Embed Your Documents

import { CohereClientV2 } from 'cohere-ai';

const cohere = new CohereClientV2();

// Your knowledge base
const documents = [
  { id: 'doc1', text: 'Cohere Command A has 256K context and supports tool use.' },
  { id: 'doc2', text: 'Embed v4 generates 1024-dim vectors with 128K token context.' },
  { id: 'doc3', text: 'Rerank v3.5 scores relevance from 0 to 1 across 100+ languages.' },
  { id: 'doc4', text: 'The Chat API v2 requires model as a mandatory parameter.' },
  { id: 'doc5', text: 'Cohere supports structured JSON output via response_format.' },
];

// Embed documents for storage
const docEmbeddings = await cohere.embed({
  model: 'embed-v4.0',
  texts: documents.map(d => d.text),
  inputType: 'search_document',
  embeddingTypes: ['float'],
});

// Store vectors alongside document text in your vector DB
const vectors = docEmbeddings.embeddings.float;
console.log(`Embedded ${vectors.length} docs, ${vectors[0].length} dimensions each`);

Step 2: Search — Embed the Query

async function searchDocuments(query: string, topK = 10) {
  // Embed the query (note: inputType is 'search_query', not 'search_document')
  const queryEmbedding = await cohere.embed({
    model: 'embed-v4.0',
    texts: [query],
    inputType: 'search_query',
    embeddingTypes: ['float'],
  });

  const queryVector = queryEmbedding.embeddings.float[0];

  // Cosine similarity search (replace with your vector DB query)
  const scores = vectors.map((vec, i) => ({
    index: i,
    score: cosineSimilarity(queryVector, vec),
  }));

  return scores
    .sort((a, b) => b.score - a.score)
    .slice(0, topK)
    .map(s => documents[s.index]);
}

function cosineSimilarity(a: number[], b: number[]): number {
  let dot = 0, magA = 0, magB = 0;
  for (let i = 0; i < a.length; i++) {
    dot += a[i] * b[i];
    magA += a[i] * a[i];
    magB += b[i] * b[i];
  }
  return dot / (Math.sqrt(magA) * Math.sqrt(magB));
}

Step 3: Rerank Retrieved Documents

async function rerankResults(query: string, candidates: typeof documents) {
  const response = await cohere.rerank({
    model: 'rerank-v3.5',
    query,
    documents: candidates.map(d => d.text),
    topN: 3,
  });

  return response.results.map(r => ({
    ...candidates[r.index],
    relevanceScore: r.relevanceScore,
  }));
}

Step 4: Generate Grounded Answer with Citations

async function ragAnswer(query: string) {
  // 1. Retrieve
  const candidates = await searchDocuments(query);

  // 2. Rerank
  const topDocs = await rerankResults(query, candidates);

  // 3. Generate with inline citations
  const response = await cohere.chat({
    model: 'command-a-03-2025',
    messages: [{ role: 'user', content: query }],
    documents: topDocs.map(d => ({
      id: d.id,
      data: { text: d.text },
    })),
  });

  const answer = response.message?.content?.[0]?.text ?? '';
  const citations = response.message?.citations ?? [];

  return { answer, citations, sources: topDocs };
}

// Usage
const result = await ragAnswer('What context length does Command A support?');
console.log('Answer:', result.answer);
console.log('Citations:', result.citations.length);

Complete Pipeline (Copy-Paste Ready)

import { CohereClientV2 } from 'cohere-ai';

const cohere = new CohereClientV2();

async function rag(query: string, knowledgeBase: string[]) {
  // 1. Rerank the knowledge base directly (skip embed for small corpora)
  const ranked = await cohere.rerank({
    model: 'rerank-v3.5',
    query,
    documents: knowledgeBase,
    topN: 5,
  });

  // 2. Feed top docs to Chat for grounded answer
  const docs = ranked.results.map((r, i) => ({
    id: `doc-${i}`,
    data: { text: knowledgeBase[r.index] },
  }));

  const response = await cohere.chat({
    model: 'command-a-03-2025',
    messages: [{ role: 'user', content: query }],
    documents: docs,
  });

  return response.message?.content?.[0]?.text ?? '';
}

Output

• Embedded document vectors (float, int8, or binary)
• Reranked candidates with relevance scores (0.0-1.0)
• Grounded answer with fine-grained citations pointing to source documents

Error Handling

Examples

Index a small approved staging corpus, run a query through embed, rerank, and grounded generation, then inspect citation coverage and retrieval quality without logging raw source documents. If citations are absent or retrieval is irrelevant, return an explicitly uncertain result and improve chunking or document scope before using the workflow for a user-facing decision.

Resources

• RAG Complete Example
• RAG Citations
• Embed API
• Rerank Best Practices

Next Steps

For tool-use and agents workflow, see cohere-core-workflow-b.