Latency Benchmarks
| API | Typical Latency | Notes |
|---|---|---|
| Writing Score | 1-3s | Depends on text length |
| AI Detection | 1-2s | Fast for short text |
| Plagiarism | 10-60s | Async, requires polling |
Instructions
Cache Score Results
import { LRUCache } from 'lru-cache';
import { createHash } from 'crypto';
const scoreCache = new LRUCache<string, any>({ max: 500, ttl: 3600000 });
async function cachedScore(text: string, token: string) { const key = createHash('sha256').update(text).digest('hex'); const cached = scoreCache.get(key); if (cached) return cached; const score = await grammarlyClient.score(text); scoreCache.set(key, score); return score; }
### Parallel API Calls
```typescript
// Score + AI detect in parallel (they're independent)
async function fullAudit(text: string, token: string) {
const [score, ai] = await Promise.all([
grammarlyClient.score(text),
grammarlyClient.detectAI(text),
]);
return { score, ai };
}
Overview
Tune latency and throughput using bounded synthetic fixtures and aggregate metrics. A performance gain is invalid if it expands text retention, access scope, error rate, or duplication risk.
Prerequisites
- Baseline latency percentiles, error/quota rates, synthetic fixture revision, and an approved error budget.
- A rollback revision for cache, concurrency, chunking, and retry settings, plus retention-safe telemetry.
Output
Return a tuning receipt with baseline/canary percentile bands, cache/concurrency/chunking revisions, quota and error outcomes, fixture result, retention check, owner approval, and rollback reference. Use aggregates only.
Error Handling
Roll back for quota saturation, increased errors, retained text in telemetry, duplicated submissions, or incorrect synthetic results. Do not raise concurrency or cache duration to hide a failure.
Examples
env=sandbox; p95=420ms->310ms; concurrency=2; cache=r4; quota=within-budget; fixture=pass; rollback=perf-r3 documents a safe canary.
Resources
Next Steps
For cost optimization, see grammarly-cost-tuning.