cohere-performance-tuning

v2026.09.24

Measure and improve Cohere latency, throughput, retrieval quality, streaming, batching, and vector representation without fabricated benchmarks. Use when optimizing a Cohere workload. Trigger with "Cohere performance", "Cohere latency", or "optimize Cohere".

GitHub
Install command
npx skhub add jeremylongshore/cohere-performance-tuning
Markdown
SKILL.md

Cohere Performance Tuning

Overview

Tune from workload-specific measurements and representative quality thresholds rather than copying universal latency numbers or synthetic benchmark claims.

Prerequisites

  • The target repository, runtime, environment, and accountable owner
  • An approved Cohere team and key for any live verification
  • Current quality, security, privacy, capacity, and change-control requirements

Tool Discipline

Use Read, Glob, and Grep to inspect code, configuration, and evidence. Use WebFetch only for current Cohere primary documentation. Use Write or Edit only when the user requested implementation and the exact target files are known; never write credentials or customer content.

Current Contract

  • Resolve models and their current context and output limits from the live catalog.
  • Measure Chat time to first event and completion separately from application queue and retrieval time.
  • Batch Embed inputs within documented request constraints and track inputs per minute.
  • Choose Rerank v4 pro or fast from measured relevance and latency on representative queries.

Authentication

Use an environment-specific key injected from an approved secret manager. Never print, persist, commit, or place CO_API_KEY in an example. Confirm access with the least costly bounded operation appropriate to the task, and treat key creation, rotation, revocation, role changes, and production-capacity requests as owner-approved actions.

Instructions

  1. Define a representative workload, quality floor, latency objective, concurrency, and cost ceiling.
  2. Instrument queue, retrieval, provider, first-event, completion, and post-processing durations.
  3. Establish a cold and warm baseline with fixed inputs and resolved model IDs.
  4. Test one variable at a time: model, context size, output bound, stream mode, Embed batch size, Rerank candidate count, or cache.
  5. Reject changes that improve speed while violating quality, citation, safety, or cost thresholds.
  6. Record the selected configuration, confidence interval, capacity headroom, and rollback trigger.

Approval Boundaries

Do not expose or rotate keys, change Cohere Team roles, accept commercial terms, enable sensitive production data, increase spend or capacity, switch production models, send a support bundle, or execute model-proposed side effects without the accountable owner's approval. Keep diagnosis read-only unless implementation was requested.

Output

Return the resolved API and model contract, files or settings inspected, evidence collected, validation result, remaining risk, owner, and rollback or next action. Redact keys, authorization headers, prompts, retrieved documents, embeddings, customer identifiers, and unrestricted environment output.

Error Handling

ConditionResponse
Benchmark varianceIncrease repetitions and separate cold, warm, and throttled samples.
Quality lossRevert the optimization even if latency improves.
429 during testLower concurrency and exclude throttled samples from normal latency claims.
Cache leakInclude tenant, model, input type, policy, and version in cache keys.

Examples

Use this compact handoff shape to keep the selected scope, validation evidence, and operational result reviewable.

Input:

endpoint=rerank; candidates=100; objective=p95; quality=ndcg-threshold

Expected handoff:

variant=rerank-v4-fast; quality=pass; p95=measured; headroom=recorded

Resources

Discovery
Tags

No tags published for this skill.

Version
Latest version metadata

Version

v2026.09.24

Published

Sep 24, 2026

Category

Uncategorized

License

MIT

Source path

skills/.curated/cohere-performance-tuning

Default branch

main

Latest commit

e5a6c3b

Tree SHA

c2dc8e8