Cohere Rerank 3.5 Pricing &
RAG Context ROI Calculator

Cross-Encoder Architecture, 4096 Token Context & Two-Stage Search TCO Optimization

Cohere Rerank 3.5 represents the benchmark in enterprise semantic re-ranking, expanding document context support to 4,096 tokens per candidate chunk across more than 100 languages. In production search architectures, relying solely on first-stage bi-encoder vector retrieval yields an average Top-5 retrieval precision of only 70-75% due to the loss of granular token-level interaction. Rerank 3.5 evaluates complete cross-attention across the query and candidate documents, lifting Mean Reciprocal Rank (MRR@10) by over 30% and eliminating irrelevant noise before prompt context reaches expensive generation LLMs.

Two-Stage Reranking Pipeline in TypeScript

Filter vector database results through Cohere Rerank 3.5 before constructing LLM context:

import { CohereClientV2 } from "cohere-ai";

const cohere = new CohereClientV2({ token: process.env.COHERE_API_KEY });

export async function rankSearchResults(query: string, rawDocs: string[]) {
  const response = await cohere.rerank({
    model: "rerank-v3.5",
    query: query,
    documents: rawDocs,
    topN: 5,
    returnDocuments: true
  });
  return response.results.map(r => ({
    text: r.document?.text,
    relevanceScore: r.relevanceScore
  }));
}

Downstream LLM Token Cost Reduction ROI

Without reranking, RAG systems must stuff 20 to 30 unverified candidate chunks into prompts to ensure recall, consuming over 15,000 input tokens per query on expensive models like Claude 3.5 Sonnet ($0.045 per prompt). By using Cohere Rerank 3.5 ($0.001 per search) to distill candidates down to the top 3 ultra-relevant chunks, input token usage drops by 85%, slashing net LLM spend by over $0.035 on every single user interaction.