File size: 10,386 Bytes
e706de2 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 |
# Concept: Streaming & Response Control
## Overview
This example demonstrates **streaming responses** and **token limits**, two essential techniques for building responsive AI agents with controlled output.
## The Streaming Problem
### Traditional (Non-Streaming) Approach
```
User sends prompt
↓
[Wait 10 seconds...]
↓
Complete response appears all at once
```
**Problems:**
- Poor user experience (long wait)
- No progress indication
- Can't interrupt bad responses
- Feels unresponsive
### Streaming Approach (This Example)
```
User sends prompt
↓
"Hoisting" (0.1s) → User sees first word!
↓
"is a" (0.2s) → More text appears
↓
"JavaScript" (0.3s) → Continuous feedback
↓
[Continues token by token...]
```
**Benefits:**
- Immediate feedback
- Progress visible
- Can interrupt early
- Feels interactive
## How Streaming Works
### Token-by-Token Generation
LLMs generate one token at a time internally. Streaming exposes this:
```
Internal LLM Process:
┌─────────────────────────────────────┐
│ Token 1: "Hoisting" │
│ Token 2: "is" │
│ Token 3: "a" │
│ Token 4: "JavaScript" │
│ Token 5: "mechanism" │
│ ... │
└─────────────────────────────────────┘
Without Streaming: With Streaming:
Wait for all tokens Emit each token immediately
└─→ Buffer → Return └─→ Callback → Display
```
### The onTextChunk Callback
```
┌────────────────────────────────────┐
│ Model Generation │
└────────────┬───────────────────────┘
│
┌────────┴─────────┐
│ Each new token │
└────────┬─────────┘
↓
┌────────────────────┐
│ onTextChunk(text) │ ← Your callback
└────────┬───────────┘
↓
Your code processes it:
• Display to user
• Send over network
• Log to file
• Analyze content
```
## Token Limits: maxTokens
### Why Limit Output?
Without limits, models might generate:
```
User: "Explain hoisting"
Model: [Generates 10,000 words including:
- Complete JavaScript history
- Every edge case
- Unrelated examples
- Never stops...]
```
With limits:
```
User: "Explain hoisting"
Model: [Generates ~1500 words
- Core concept
- Key examples
- Stops at 2000 tokens]
```
### Token Budgeting
```
Context Window: 4096 tokens
├─ System Prompt: 200 tokens
├─ User Message: 100 tokens
├─ Response (maxTokens): 2000 tokens
└─ Remaining for history: 1796 tokens
Total used: 2300 tokens
Available: 1796 tokens for future conversation
```
### Cost vs Quality
```
Token Limit Output Quality Use Case
─────────── ─────────────── ─────────────────
100 Brief, may be cut Quick answers
500 Concise but complete Short explanations
2000 (example) Detailed Full explanations
No limit Risk of rambling When length unknown
```
## Real-Time Applications
### Pattern 1: Interactive CLI
```
User: "Explain closures"
↓
Terminal: "A closure is a function..."
(Appears word by word, like typing)
↓
User sees progress, knows it's working
```
### Pattern 2: Web Application
```
Browser Server
│ │
├─── Send prompt ────────→│
│ │
│←── Chunk 1: "Closures"──┤
│ (Display immediately) │
│ │
│←── Chunk 2: "are"───────┤
│ (Append to display) │
│ │
│←── Chunk 3: "functions"─┤
│ (Keep appending...) │
```
Implementation:
- Server-Sent Events (SSE)
- WebSockets
- HTTP streaming
### Pattern 3: Multi-Consumer
```
onTextChunk(text)
│
┌───────┼───────┐
↓ ↓ ↓
Console WebSocket Log File
Display → Client → Storage
```
## Performance Characteristics
### Latency vs Throughput
```
Time to First Token (TTFT):
├─ Small model (1.7B): ~100ms
├─ Medium model (8B): ~200ms
└─ Large model (20B): ~500ms
Tokens Per Second:
├─ Small model: 50-80 tok/s
├─ Medium model: 20-35 tok/s
└─ Large model: 10-15 tok/s
User Experience:
TTFT < 500ms → Feels instant
Tok/s > 20 → Reads naturally
```
### Resource Trade-offs
```
Model Size Memory Speed Quality
────────── ──────── ───── ───────
1.7B ~2GB Fast Good
8B ~6GB Medium Better
20B ~12GB Slower Best
```
## Advanced Concepts
### Buffering Strategies
**No Buffer (Immediate)**
```
Every token → callback → display
└─ Smoothest UX but more overhead
```
**Line Buffer**
```
Accumulate until newline → flush
└─ Better for paragraph-based output
```
**Time Buffer**
```
Accumulate for 50ms → flush batch
└─ Reduces callback frequency
```
### Early Stopping
```
Generation in progress:
"The answer is clearly... wait, actually..."
↑
onTextChunk detects issue
↓
Stop generation
↓
"Let me reconsider"
```
Useful for:
- Detecting off-topic responses
- Safety filters
- Relevance checking
### Progressive Enhancement
```
Partial Response Analysis:
┌─────────────────────────────────┐
│ "To implement this feature..." │
│ │
│ ← Already useful information │
│ │
│ "...you'll need: 1) Node.js" │
│ │
│ ← Can start acting on this │
│ │
│ "2) Express framework" │
└─────────────────────────────────┘
Agent can begin working before response completes!
```
## Context Size Awareness
### Why It Matters
```
┌────────────────────────────────┐
│ Context Window (4096) │
├────────────────────────────────┤
│ System Prompt 200 tokens │
│ Conversation History 1000 │
│ Current Prompt 100 │
│ Response Space 2796 │
└────────────────────────────────┘
If maxTokens > 2796:
└─→ Error or truncation!
```
### Dynamic Adjustment
```
Available = contextSize - (prompt + history)
if (maxTokens > available) {
maxTokens = available;
// or clear old history
}
```
## Streaming in Agent Architectures
### Simple Agent
```
User → LLM (streaming) → Display
└─ onTextChunk shows progress
```
### Multi-Step Agent
```
Step 1: Plan (stream) → Show thinking
Step 2: Act (stream) → Show action
Step 3: Result (stream) → Show outcome
└─ User sees agent's process
```
### Collaborative Agents
```
Agent A (streaming) ──┐
├─→ Coordinator → User
Agent B (streaming) ──┘
└─ Both stream simultaneously
```
## Best Practices
### 1. Always Set maxTokens
```
✓ Good:
session.prompt(query, { maxTokens: 2000 })
✗ Risky:
session.prompt(query)
└─ May use entire context!
```
### 2. Handle Partial Updates
```
let fullResponse = '';
onTextChunk: (chunk) => {
fullResponse += chunk;
display(chunk); // Show immediately
logComplete = false; // Mark incomplete
}
// After completion:
saveToDatabase(fullResponse);
```
### 3. Provide Feedback
```
onTextChunk: (chunk) => {
if (firstChunk) {
showLoadingDone();
firstChunk = false;
}
appendToDisplay(chunk);
}
```
### 4. Monitor Performance
```
const startTime = Date.now();
let tokenCount = 0;
onTextChunk: (chunk) => {
tokenCount += estimateTokens(chunk);
const elapsed = (Date.now() - startTime) / 1000;
const tokensPerSecond = tokenCount / elapsed;
updateMetrics(tokensPerSecond);
}
```
## Key Takeaways
1. **Streaming improves UX**: Users see progress immediately
2. **maxTokens controls cost**: Prevents runaway generation
3. **Token-by-token generation**: LLMs produce one token at a time
4. **onTextChunk callback**: Your hook into the generation process
5. **Context awareness matters**: Monitor available space
6. **Essential for production**: Real-time systems need streaming
## Comparison
```
Feature intro.js coding.js (this)
──────────────── ───────── ─────────────────
Streaming ✗ ✓
Token limit ✗ ✓ (2000)
Real-time output ✗ ✓
Progress visible ✗ ✓
User control ✗ ✓
```
This pattern is foundational for building responsive, user-friendly AI agent interfaces.
|