tools/prompt-caching.html at main · simonw/tools

History

518 lines (448 loc) · 20.3 KB

Raw

100

101

102

103

104

105

106

107

108

109

110

111

112

113

114

115

116

117

118

119

120

121

122

123

124

125

126

127

128

129

130

131

132

133

134

135

136

137

138

139

140

141

142

143

144

145

146

147

148

149

150

151

152

153

154

155

156

157

158

159

160

161

162

163

164

165

166

167

168

169

170

171

172

173

174

175

176

177

178

179

180

181

182

183

184

185

186

187

188

189

190

191

192

193

194

195

196

197

198

199

200

201

202

203

204

205

206

207

208

209

210

211

212

213

214

215

216

217

218

219

220

221

222

223

224

225

226

227

228

229

230

231

232

233

234

235

236

237

238

239

240

241

242

243

244

245

246

247

248

249

250

251

252

253

254

255

256

257

258

259

260

261

262

263

264

265

266

267

268

269

270

271

272

273

274

275

276

277

278

279

280

281

282

283

284

285

286

287

288

289

290

291

292

293

294

295

296

297

298

299

300

301

302

303

304

305

306

307

308

309

310

311

312

313

314

315

316

317

318

319

320

321

322

323

324

325

326

327

328

329

330

331

332

333

334

335

336

337

338

339

340

341

342

343

344

345

346

347

348

349

350

351

352

353

354

355

356

357

358

359

360

361

362

363

364

365

366

367

368

369

370

371

372

373

374

375

376

377

378

379

380

381

382

383

384

385

386

387

388

389

390

391

392

393

394

395

396

397

398

399

400

401

402

403

404

405

406

407

408

409

410

411

412

413

414

415

416

417

418

419

420

421

422

423

424

425

426

427

428

429

430

431

432

433

434

435

436

437

438

439

440

441

442

443

444

445

446

447

448

449

450

451

452

453

454

455

456

457

458

459

460

461

462

463

464

465

466

467

468

469

470

471

472

473

474

475

476

477

478

479

480

481

482

483

484

485

486

487

488

489

490

491

492

493

494

495

496

497

498

499

500

501

502

503

504

505

506

507

508

509

510

511

512

513

514

515

516

517

518

<!doctype html>

<head>

<title>OpenAI Prompt Caching Playground</title>

<style>

:root {

--bg:#0d1117; --panel:#161b22; --ink:#e6edf3; --muted:#9da7b3; --accent:#3fb950; --warn:#f0883e; --border:#30363d;

}

html,body {background:var(--bg); color:var(--ink); font-family: ui-sans-serif, system-ui, -apple-system, Segoe UI, Roboto, Helvetica, Arial, "Apple Color Emoji", "Segoe UI Emoji"; margin:0; padding:0;}

header {padding:16px 20px; border-bottom:1px solid var(--border); position:sticky; top:0; background:rgba(13,17,23,.9); backdrop-filter: blur(6px);}

h1 {font-size:18px; margin:0 0 6px;}

.sub {color:var(--muted); font-size:13px;}

.wrap {display:grid; grid-template-columns: 360px 1fr; gap:16px; padding:16px;}

.panel {background:var(--panel); border:1px solid var(--border); border-radius:10px; padding:14px;}

.panel h2 {font-size:14px; margin:0 0 10px; color:#c9d1d9; letter-spacing:.2px;}

textarea, input, select {background:#0b0f14; color:var(--ink); border:1px solid var(--border); border-radius:8px; padding:8px; width:100%; box-sizing:border-box; font: inherit;}

textarea {min-height:110px; resize:vertical; line-height:1.35;}

.row {display:flex; gap:8px; align-items:center;}

.row > * {flex:1;}

.row .shrink {flex:0 0 auto;}

.btn {background:#21262d; border:1px solid var(--border); padding:8px 10px; border-radius:8px; cursor:pointer; color:var(--ink);}

.btn:hover {border-color:#6e7681;}

.btn.primary {background:var(--accent); color:#031d0b; border-color:#2ea043; font-weight:600;}

.btn.ghost {background:transparent;}

.grid-2 {display:grid; grid-template-columns:1fr 1fr; gap:8px;}

.stats {display:grid; grid-template-columns: repeat(6, minmax(90px, 1fr)); gap:8px; margin-top:8px;}

.stat {background:#0b0f14; border:1px dashed var(--border); border-radius:8px; padding:8px;}

.stat .k {font-size:11px; color:var(--muted);}

.stat .v {font-size:16px; font-weight:700; margin-top:4px;}

.badge {display:inline-block; padding:2px 8px; border-radius:999px; font-size:12px; border:1px solid var(--border); background:#0b0f14; color:var(--muted);}

.badge.hit {background:rgba(63,185,80,.12); color:#56d364; border-color:#2ea043;}

.badge.miss {background:rgba(240,136,62,.12); color:#f0883e; border-color:#8a5700;}

.log {font-family: ui-monospace, SFMono-Regular, Menlo, Consolas, monospace; font-size:12px; white-space:pre-wrap; background:#0b0f14; border:1px solid var(--border); border-radius:8px; padding:10px; max-height:260px; overflow:auto;}

.out {background:#0b0f14; border:1px solid var(--border); border-radius:8px; padding:10px; min-height:140px; font-family: ui-monospace, Menlo, Consolas, monospace; white-space:pre-wrap;}

/* Raw JSON sections */

.code {

background:#0b0f14; border:1px solid var(--border);

border-radius:8px; padding:10px;

font-family: ui-monospace, SFMono-Regular, Menlo, Consolas, monospace;

font-size:12px; line-height:1.5; white-space:pre-wrap;

max-height:260px; overflow:auto;

}

details summary { cursor:pointer; user-select:none; }

details { margin:8px 0; }

.footer-note {color:var(--muted); font-size:12px;}

.inline {display:inline-flex; align-items:center; gap:8px;}

.sep {height:1px; background:var(--border); margin:10px 0;}

@media (max-width: 980px){ .wrap{grid-template-columns:1fr;} }

</style>

</head>

<body>

<h1>Prompt Caching Playground</h1>

Optimize your prompt structure and observe cache hits. Cache engages at <strong>≥ 1024 tokens</strong>, with cacheable chunks in <strong>128-token</strong> increments (e.g., 1024, 1152, 1280…).

</div>

</header>

<h2>Setup</h2>

<option value="chat">Chat Completions API</option>

<option value="responses">Responses API (supports prompt_cache_key)</option>

</select>

</select>

</div>

</div>

<option value="starter">Load demo scenario…</option>

<option value="A">A) Static instructions + Doc A + Q1</option>

<option value="A2">A2) Same instructions + Doc A + Q2</option>

<option value="B">B) Same instructions + Doc B + Q1</option>

</select>

</div>

Repeat instructions

</label>

Repeat doc

</label>

</div>

Use repeats to push your prompt over 1024 tokens (and observe cached increments).

</div>

<h2>Prompt Parts</h2>

<label class="footer-note">Initial Instructions (static; put these first for best cache hit rate)</label>

<label class="footer-note" style="margin-top:8px; display:block;">Document Context (semi-static; keep above user input)</label>

<label class="footer-note" style="margin-top:8px; display:block;">User Question (dynamic; place last)</label>

<button class="btn" id="sendVariantBtn">Send Variant (new Q)</button>

</div>

<button class="btn ghost" id="clearLogBtn">Clear Log</button>

</div>

Tip: Try A → A2 (same prefix, new question) to see cache hits; then A → B to see instruction-only prefix hits.

</div>

<h2>Result</h2>

<span id="hitBadge" class="badge">cache status</span>

<span id="timeBadge" class="badge">latency</span>

<span id="tsBadge" class="badge">timestamp</span>

</div>

<div class="stat"><div class="k">Prompt Tokens</div><div class="v" id="promptTokens">–</div></div>

<div class="stat"><div class="k">Cached Tokens</div><div class="v" id="cachedTokens">–</div></div>

<div class="stat"><div class="k">Cache Hit %</div><div class="v" id="hitPct">–</div></div>

<div class="stat"><div class="k">Completion Tokens</div><div class="v" id="completionTokens">–</div></div>

<div class="stat"><div class="k">Total Tokens</div><div class="v" id="totalTokens">–</div></div>

<div class="stat"><div class="k">Model</div><div class="v" id="modelEcho">–</div></div>

</div>

<div class="k">Endpoint</div>

</div>

<div class="k">prompt_cache_key</div>

</div>

<div class="out" id="output">Response will appear here…</div>

<summary>Most recent request JSON</summary>

</details>

<summary>Most recent response JSON</summary>

</details>

</div>

/** Minimal, in-page "localStorage prompt" API key flow */

function getAPIKey() {

let apiKey = localStorage.getItem('openai_api_key');

if (!apiKey) {

apiKey = prompt('Please enter your OpenAI API Key:');

if (apiKey) {

localStorage.setItem('openai_api_key', apiKey);

}

return apiKey;

}

/** Helpers */

const $ = (id) => document.getElementById(id);

function nowISO() {

const d = new Date();

return d.toLocaleString(undefined, {hour12:false}) + ' (' + d.toISOString() + ')';

}

function logLine(txt) {

const el = $('runLog');

el.textContent += txt + '\n';

el.scrollTop = el.scrollHeight;

}

function setBadge(el, text, kind) {

el.textContent = text;

el.classList.remove('hit','miss');

if (kind) el.classList.add(kind);

}

function pretty(obj) {

try { return JSON.stringify(obj, null, 2); }

catch { return String(obj); }

}

function setReqJson(url, payload) {

// Do NOT include headers / API key. Show only URL + payload.

const snapshot = { url, payload };

$('reqJson').textContent = pretty(snapshot);

}

function setResJson(data) {

$('resJson').textContent = pretty(data);

}

/** Demo scenarios (2× length) */

const demo = {

instructions: `You are a highly concise assistant. Always answer in 1–3 sentences unless explicitly asked for more. Use plain language. If the user asks for code, include only code unless told otherwise.

General rules:

- If asked to compare, list key differences as short bullets.

- If a calculation is needed, show the equation then the result.

- If a definition is requested, give a crisp one-liner first.

- Prefer step-by-step logic but keep it terse; avoid filler.

- Cite assumptions when they influence the answer.

Style guide:

- Avoid hedging like “it seems.” Be direct when evidence supports it.

- Use simple words; avoid jargon unless the question is technical.

- Bullets over paragraphs when listing 3+ items.

- For numbers: include units and orders of magnitude when relevant.

Math & formatting:

- Show one compact formula, then compute.

- Round sensibly; default to 2–3 sig figs unless precision matters.

- Monospace code blocks for code only; no prose inside code fences.

Safety rails:

- Decline disallowed content with a brief reason and a safer alternative.

- Don’t fabricate citations, links, or data you can’t verify.

Edge cases:

- If the question is ambiguous, answer the most common interpretation and note the alternative in one line.

- If insufficient data, state what’s missing and provide a minimal actionable next step.

Examples (for length and reference):

Q: Define latency vs throughput.

A: Latency is per-request delay; throughput is requests/second. They trade off: batching raises throughput but can add latency.

Q: Summarize a spec into 3 bullets.

A: Goal, core objects, critical risks (1 line each).

Q: Show a quick ROI calc.

A: ROI = (benefit − cost)/cost. With $50k benefit and $20k cost, ROI = (50−20)/20 = 1.5 = 150%.

When code is requested:

- Provide a minimal runnable snippet.

- Include a 1-line comment on how to run or integrate.

Tools you could hypothetically use (for realism in token length): web.run (search, open, click), python (analysis), image_query (images for people/places). These tool names are illustrative for prompt length; they don’t execute here.

Glossary (compact):

- CRDT: conflict-free replicated data type.

- Event sourcing: store events then derive state from them.

Extended guidance filler to lift token count:

${'Guideline: Prefer clear, literal phrasing; keep answers scoped; surface assumptions explicitly; show one example when helpful.\n'.repeat(60)}

(End of static instructions.)`,

docA: `Product Spec: Nimbus Notes

- Goal: Lightning-fast note capture with offline-first sync.

- Core objects: Notebook, Note, Tag, Attachment.

- Sync model: CRDT-based; conflict-free merges.

- Pricing: Free; Pro at $4/month with 10GB attachments.

- Roadmap Q4: iOS widgets, Android share sheet revamp, web clipper.

Architecture notes:

- Local-first write path, background sync, deterministic conflict resolution.

- Indexing: inverted index on device; server compaction nightly.

- Export: Markdown + attachments in a flat bundle.

Risks:

- Complex merge semantics for rich text.

- Battery usage on large notebooks during background sync.

Extra long filler to increase tokens:

${'• Feature: ' + 'rich text, backlinks, slash commands, export to Markdown.\n'.repeat(80)}`,

docB: `Product Spec: Zephyr Tasks

- Goal: Kanban-like tasking for small teams, built on ActivityPub.

- Core objects: Board, Column, Card, Checklist, Comment.

- Sync model: Event-sourced with snapshotting.

- Pricing: Free; Team at $6/user/month, custom SSO.

- Roadmap Q4: Gantt view, email in, calendar sync.

Architecture notes:

- Federated updates via ActivityPub; per-tenant queues.

- Snapshot intervals tuned to active column size.

- Automation hooks for webhooks and email ingestion.

Risks:

- Federation consistency and moderation boundaries.

- Notification overload without good defaults.

Extra long filler to increase tokens:

${'• Capability: ' + 'labels, due dates, swimlanes, WIP limits, automations.\n'.repeat(80)}`,

q1A: `What are the minimum moving parts I need to build first MVP?`,

q2A: `What are the top three risks for this approach?`,

q1B: `What metrics should we track in the first month?`,

};

function loadScenario(which) {

if (which === 'A') {

$('instructions').value = demo.instructions;

$('doc').value = demo.docA;

$('question').value = demo.q1A;

$('scenarioName').value = 'A (Instr + Doc A + Q1)';

} else if (which === 'A2') {

$('instructions').value = demo.instructions;

$('doc').value = demo.docA;

$('question').value = demo.q2A;

$('scenarioName').value = 'A2 (Instr + Doc A + Q2)';

} else if (which === 'B') {

$('instructions').value = demo.instructions;

$('doc').value = demo.docB;

$('question').value = demo.q1B;

$('scenarioName').value = 'B (Instr + Doc B + Q1)';

} else {

$('instructions').value = demo.instructions;

$('doc').value = demo.docA;

$('question').value = demo.q1A;

$('scenarioName').value = 'Starter';

}

$('loadScenarioBtn').addEventListener('click', () => loadScenario($('scenarioPicker').value));

/** Build messages with static-first ordering (to maximize cache hits) */

function buildMessages() {

const instr = $('instructions').value;

const doc = $('doc').value;

const question = $('question').value;

const repI = Math.max(1, parseInt($('repeatInstructions').value || '1', 10));

const repD = Math.max(1, parseInt($('repeatDoc').value || '1', 10));

const instrRepeated = Array.from({length:repI}).map(() => instr).join('\n\n');

const docRepeated = Array.from({length:repD}).map((_,i) => `# Document Copy ${i+1}\n` + doc).join('\n\n');

const system = instrRepeated.trim();

const user = (`Document Context:\n${docRepeated}\n\nUser Question:\n${question}`).trim();

const messages = [

{role:'system', content: system},

{role:'user', content: user},

];

return {messages, system, user};

}

/** Core send function (supports Chat Completions and Responses API) */

async function send(kind) {

const apiKey = getAPIKey();

if (!apiKey) return;

// Optionally mutate the user question for the "variant" button

if (kind === 'variant') {

$('question').value = $('question').value + ' (Please keep it to 3 bullets.)';

}

const endpointSel = $('endpoint').value;

const model = $('model').value;

const pck = $('cacheKey').value.trim() || null;

const label = $('scenarioName').value.trim() || '(unnamed)';

const {messages} = buildMessages();

let url, payload, headers = {

'Content-Type': 'application/json',

'Authorization': `Bearer ${apiKey}`,

};

if (endpointSel === 'responses') {

url = 'https://api.openai.com/v1/responses';

payload = {

model,

messages,

prompt_cache_key: pck || undefined,

temperature: 0.2,

};

} else {

url = 'https://api.openai.com/v1/chat/completions';

payload = {

model,

messages,

temperature: 0.2,

};

}

// Show outgoing request snapshot (no headers / key) and reset response pane

setReqJson(url, payload);

$('resJson').textContent = '—';

// UI pre-state

setBadge($('hitBadge'), '…', null);

setBadge($('timeBadge'), '–', null);

setBadge($('tsBadge'), new Date().toLocaleTimeString(), null);

$('output').textContent = 'Waiting for response…';

$('promptTokens').textContent = '–';

$('cachedTokens').textContent = '–';

$('hitPct').textContent = '–';

$('completionTokens').textContent = '–';

$('totalTokens').textContent = '–';

$('modelEcho').textContent = model;

$('endpointEcho').textContent = endpointSel;

$('pckEcho').textContent = pck || '—';

const t0 = performance.now();

let resp, data, ok = false, textOut = '', usage = {}, cached = 0;

try {

resp = await fetch(url, {method:'POST', headers, body: JSON.stringify(payload)});

const t1 = performance.now();

const latencyMs = Math.round(t1 - t0);

ok = resp.ok;

data = await resp.json();

setResJson(data);

// Extract output text + usage across both APIs

if (endpointSel === 'responses') {

textOut = data.output_text ?? '';

if (!textOut && Array.isArray(data.output)) {

textOut = data.output.map(o => {

if (o?.content) {

return o.content.map(c => (c.type === 'output_text' && c.text) ? c.text : '').join('');

}

return '';

}).join('').trim();

}

usage = data.usage || {};

} else {

textOut = (data.choices && data.choices[0] && data.choices[0].message && data.choices[0].message.content) || '';

usage = data.usage || {};

}

// Pull cached_tokens if present

const promptDetails = usage.prompt_tokens_details || {};

cached = (typeof promptDetails.cached_tokens === 'number') ? promptDetails.cached_tokens : 0;

// Display stats

const pt = usage.prompt_tokens ?? '—';

const ct = usage.completion_tokens ?? '—';

const tt = usage.total_tokens ?? '—';

$('promptTokens').textContent = pt;

$('cachedTokens').textContent = cached;

$('completionTokens').textContent = ct;

$('totalTokens').textContent = tt;

// Cache hit badge logic

let hitBadgeText = (cached && cached > 0) ? `cache hit: ${cached}` : 'cache miss';

let hitKind = (cached && cached > 0) ? 'hit' : 'miss';

setBadge($('hitBadge'), hitBadgeText, hitKind);

// Cache hit %

let hitPct = (typeof pt === 'number' && pt > 0 && typeof cached === 'number') ? ((cached/pt)*100).toFixed(1) + '%' : '—';

$('hitPct').textContent = hitPct;

setBadge($('timeBadge'), `${latencyMs} ms`, null);

setBadge($('tsBadge'), new Date().toLocaleTimeString(), null);

// Show response text

$('output').textContent = textOut || JSON.stringify(data, null, 2);

// Log line (concise)

const ts = nowISO();

logLine(`[${ts}] label="${label}" endpoint=${endpointSel} model=${model}` +

(pck ? ` pck="${pck}"` : '') +

` | latency=${latencyMs}ms tokens: prompt=${pt} cached=${cached} completion=${ct} total=${tt} | ` +

`hit=${cached>0} | Q="${$('question').value.trim().slice(0,80)}"`);

} catch (e) {

$('output').textContent = 'Error: ' + (e?.message || e);

setResJson({ error: String(e) });

setBadge($('hitBadge'), 'error', 'miss');

logLine(`[${nowISO()}] ERROR ${e?.message || e}`);

}

/** UI wires */

$('sendBtn').addEventListener('click', () => send('normal'));

$('sendVariantBtn').addEventListener('click', () => send('variant'));

$('clearLogBtn').addEventListener('click', () => { $('runLog').textContent=''; });

$('copyLogBtn').addEventListener('click', async () => {

await navigator.clipboard.writeText($('runLog').textContent);

alert('Log copied to clipboard.');

});

/** Initialize with starter scenario + session log */

function init() {

loadScenario('starter');

logLine(`[${nowISO()}] Session started. Try A → A2 for a cache hit (same prefix, new question). Then A → B to see instruction-only prefix hits. Caches typically persist ~5–10 minutes idle (sometimes up to ~1 hour off-peak).`);

}

init();

</script>

</body>

</html>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

FilesExpand file tree

prompt-caching.html

Latest commit

History

prompt-caching.html

File metadata and controls