夜雨聆风
手撸AI对话助手带上思考过程
之前文章《用 LangChain 驱动本地 Ollama 模型》讲叙了使用 LangChain 进行大模型对话。大模型的响应时间一般都会比较长，那么如何考虑给用户更好的体验呢？
流式输出
类似打字机一样的效果，按token输出。
安装依赖
1 pip install -U uvicorn "fastapi[standard]" "langchain[openai]"
调用流式输出
核心方法：stream/astream
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 import jsonfrom fastapi import FastAPIfrom fastapi.middleware.cors import CORSMiddlewarefrom fastapi.responses import StreamingResponsefrom langchain_openai import ChatOpenAIapp = FastAPI()app.add_middleware( CORSMiddleware, allow_origins=["*"], allow_credentials=True, allow_methods=["*"], allow_headers=["*"],)@app.post("/api/bot/chat")async def bot_chat(request: dict): query = request.get("query", "你好") llm = ChatOpenAI( model="qwen3.5:35b", base_url="http://192.168.31.24:4000", api_key="your api key", temperature=0.7, streaming=True, ) system_prompt = ( "你是一个会展示思考过程的AI。" ) messages = [ {"role": "system", "content": system_prompt}, {"role": "user", "content": query}, ] async def generate(): # 直接用 LLM astream，逐 token 流出 async for chunk in llm.astream(messages): content = chunk.content or "" if not content: continue yield json.dumps({ "type": "chunk", "content": content, }, ensure_ascii=False) + "\n" yield json.dumps({"type": "done"}) + "\n" return StreamingResponse( generate(), media_type="application/x-ndjson", headers={ "Cache-Control": "no-cache", "X-Accel-Buffering": "no", }, )if __name__ == "__main__": import uvicorn uvicorn.run(app, host="0.0.0.0", port=8000)
测试
1 2 3 4 5 6 curl --location --request POST 'http://127.0.0.1:8000/api/bot/chat' \--header 'Accept: application/json' \--header 'Content-Type: application/json' \--data-raw '{ "query": "请一步步思考：2+3等于多少？"}'
思考过程
提示词
在大模型的响应参数里，只有部分模型是带有reasoning，如果要兼容大部分模型，就要换种方式，输出时带标签标识。
1 2 3 4 5 6 7 8 9 10 11 12 #... 其它不变system_prompt = ( "你是一个会展示思考过程的AI。\n" "请先输出你的思考过程（用<THINK>标签包裹），" "然后再输出最终答案（用<FINAL>标签包裹）。\n\n" "示例：\n" "<THINK>这里是推理过程</THINK>\n" "<FINAL>这里是最终答案</FINAL>")#... 其它不变
页面实现
通过fetch实现简单示例效果
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 <!DOCTYPE html><html lang="zh"><head> <meta charset="UTF-8"> <meta name="viewport" content="width=device-width, initial-scale=1.0"> <title>MiMo Chat</title> <style> * { margin: 0; padding: 0; box-sizing: border-box; } :root { --bg: #0a0a0f; --surface: #12121a; --border: #2a2a3a; --think-bg: #13130a; --think-border: #3a3a1a; --think-text: #d4a843; --think-label: #b8941f; --answer-text: #e0e0e0; --accent: #6366f1; --accent-glow: rgba(99, 102, 241, 0.15); --muted: #6b6b80; } body { font-family: 'Inter', -apple-system, sans-serif; background: var(--bg); color: var(--answer-text); min-height: 100vh; display: flex; flex-direction: column; align-items: center; } .chat-container { width: 100%; max-width: 760px; padding: 24px 16px; display: flex; flex-direction: column; gap: 20px; } .input-area { position: sticky; transform: translateY( 0; z-index: 10; background: var(--bg); padding: 16px 0; display: flex; gap: 10px; } .input-area textarea { flex: 1; background: var(--surface); border: 1px solid var(--border); border-radius: 12px; padding: 12px 16px; color: var(--answer-text); font-family: inherit; font-size: 14px; resize: none; outline: none; min-height: 44px; max-height: 120px; transition: border-color 0.2s; } .input-area textarea:focus { border-color: var(--accent); box-shadow: 0 0 0 3px var(--accent-glow); } .send-btn { background: var(--accent); border: none; border-radius: 12px; padding: 0 20px; color: #fff; font-weight: 600; font-size: 14px; cursor: pointer; transition: all 0.2s; white-space: nowrap; } .send-btn:hover { opacity: 0.85; } .send-btn:disabled { opacity: 0.4; cursor: not-allowed; } .message { background: var(--surface); border: 1px solid var(--border); border-radius: 16px; overflow: hidden; animation: slideUp 0.3s ease; } @keyframes slideUp { from { opacity: 0; transform: translateY(12px); } to { opacity: 1; transform: translateY(0); } } .message-header { padding: 12px 16px; border-bottom: 1px solid var(--border); display: flex; justify-content: space-between; font-size: 12px; color: var(--muted); } .message-body { padding: 16px; } .think-section { background: var(--think-bg); border: 1px solid var(--think-border); border-radius: 10px; margin-bottom: 12px; overflow: hidden; display: none; } .think-section.visible { display: block; } .think-header { padding: 10px 14px; display: flex; align-item)s: center; gap: 8px; cursor: pointer; user-select: none; font-size: 13px; color: var(--think-label); font-weight: 500; } .think-arrow { display: inline-block; transition: transform 0.25s; font-size: 10px; } .think-arrow.collapsed { transform: rotate(-90deg); } .think-content { padding: 0 14px 12px; font-family: 'JetBrains Mono', monospace; font-size: 13px; line-height: 1.75; color: var(--think-text); white-space: pre-wrap; word-break: break-word; max-height: 800px; opacity: 1; overflow: hidden; transition: max-height 0.3s, padding 0.3s, opacity 0.3s; } .think-content.collapsed { max-height: 0; padding: 0 14px; opacity: 0; } .answer-content { font-size: 15px; line-height: 1.8; white-space: pre-wrap; word-break: break-word; } .cursor { display: inline-block; width: 2px; height: 1.1em; background: var(--accent); margin-left: 1px; animation: blink 0.7s step-end infinite; vertical-align: text-bottom; } @keyframes blink { 50% { opacity: 0; } } .loading-dots { display: inline-flex; gap: 3px; margin-left: auto; } .loading-dots span { width: 5px; height: 5px; background: var(--think-label); border-radius: 50%; animation: bounce 1.2s infinite; } .loading-dots span:nth-child(2) { animation-delay: 0.15s; } .loading-dots span:nth-child(3) { animation-delay: 0.3s; } @keyframes bounce { 0%, 80%, 100% { transform: scale(0.5); opacity: 0.3; } 40% { transform: scale(1); opacity: 1; } } </style></head><body><div class="chat-container"> <div class="input-area"> <textarea id="queryInput" placeholder="输入问题..." rows="1" onkeydown="if(event.key==='Enter'&&!event.shiftKey){event.preventDefault();sendMessage();}"> 请一步步思考：2+3等于多少？ </textarea> <button class="send-btn" id="sendBtn" onclick="sendMessage()">发送</button> </div> <div id="messages"></div></div><script> let isSending = false; async function sendMessage() { if (isSending) return; const query = document.getElementById('queryInput').value.trim(); if (!query) return; isSending = true; document.getElementById('sendBtn').disabled = true; const msgId = crypto.randomUUID().replace(/-/g, ''); const conversationId = localStorage.getItem('convId') || crypto.randomUUID().replace(/-/g, ''); localStorage.setItem('convId', conversationId); const msgEl = createMessage(msgId); document.getElementById('messages').prepend(msgEl); const state = { raw: '', lastThinkLen: 0, lastAnswerLen: 0, thinkCollapsed: false, }; try { const resp = await fetch('http://127.0.0.1:8000/api/bot/chat', { method: 'POST', headers: {'Content-Type': 'application/json'}, body: JSON.stringify({ msgId, conversationId, query, isEditedQuery: false, modelConfig: {enableThinking: true, webSearchStatus: "disabled", model: ""}, multiMedias: [], }), }); const reader = resp.body.getReader(); const decoder = new TextDecoder(); let buffer = ''; while (true) { const {done, value} = await reader.read(); if (done) break; buffer += decoder.decode(value, {stream: true}); const lines = buffer.split('\n'); buffer = lines.pop(); for (const line of lines) { if (!line.trim()) continue; let data; try { data = JSON.parse(line); } catch (e) { continue; } if (data.type === 'chunk') { state.raw += data.content; render(msgEl, state); } else if (data.type === 'done') { render(msgEl, state); removeCursor(msgEl); } } } if (buffer.trim()) { try { const d = JSON.parse(buffer); if (d.type === 'chunk') state.raw += d.content; } catch (e) { } } render(msgEl, state); removeCursor(msgEl); } catch (err) { msgEl.querySelector('.answer-content').textContent = '错误: ' + err.message; } finally { isSending = false; document.getElementById('sendBtn').disabled = false; } } /** * 当标签被 chunk 切断时（如 </THI → NK>），正则的 |$ 兜底 * 会把部分标签（</THI）当成内容显示。 * * 示例： * raw = "<THINK>2+3等于5</THI" * thinkMatch[1] = "2+3等于5</THI" ← 正则兜底，残留标签混入 * raw.includes('</THINK>') = false ← 完整标签还没到 * stripPartial(thinkText, '</THI') ← 从末尾剥掉残留 * → "2+3等于5" ← 干净 * * raw = "<THINK>2+3等于5</THINK>" * thinkMatch[1] = "2+3等于5" ← 正则精确匹配，无残留 * raw.includes('</THINK>') = true * → 不需要剥 */ function render(msgEl, state) { const raw = state.raw; const thinkMatch = raw.match(/<THINK>([\s\S]*?)(?:<\/THINK>|$)/); const answerMatch = raw.match(/<FINAL>([\s\S]*?)(?:<\/FINAL>|$)/); let thinkText = thinkMatch ? thinkMatch[1] : ''; let answerText = answerMatch ? answerMatch[1] : ''; // 标签未完整时，剥掉末尾的残留片段 if (!raw.includes('</THINK>')) { thinkText = stripPartial(thinkText, '</THINK>'); } if (!raw.includes('</FINAL>')) { answerText = stripPartial(answerText, '</FINAL>'); } // Think 区域 if (thinkText.length > 0) { const section = msgEl.querySelector('.think-section'); const content = msgEl.querySelector('.think-content'); section.classList.add('visible'); if (thinkText.length > state.lastThinkLen) { const delta = thinkText.substring(state.lastThinkLen); content.appendChild(document.createTextNode(delta)); content.scrollTop = content.scrollHeight; state.lastThinkLen = thinkText.length; } if (raw.includes('</THINK>') && !state.thinkCollapsed) { state.thinkCollapsed = true; const dots = section.querySelector('.loading-dots'); if (dots) dots.style.display = 'none'; section.querySelector('.think-arrow').classList.add('collapsed'); content.classList.add('collapsed'); } } // Answer 区域 if (answerText.length > 0) { const content = msgEl.querySelector('.answer-content'); if (answerText.length > state.lastAnswerLen) { const delta = answerText.substring(state.lastAnswerLen); const oldCursor = content.querySelector('.cursor'); if (oldCursor) oldCursor.remove(); content.appendChild(document.createTextNode(delta)); const cursor = document.createElement('span'); cursor.className = 'cursor'; content.appendChild(cursor); state.lastAnswerLen = answerText.length; } } } /** * 从文本末尾剥掉部分标签 * 比如 stripPartial("2+3等于5</THI", "</THINK") → "2+3等于5" * 逐个尝试 "</", "</T", "</TH", ... 直到完整标签 */ function stripPartial(text, fullTag) { for (let i = 2; i <= fullTag.length; i++) { const suffix = fullTag.substring(0, i); if (text.endsWith(suffix)) { return text.slice(0, -suffix.length); } } return text; } function createMessage(msgId) { const el = document.createElement('div'); el.className = 'message'; el.id = `msg-${msgId}`; el.innerHTML = ` <div class="message-header"> <span>AI 助手</span> <span class="timestamp">${new Date().toLocaleTimeString()}</span> </div> <div class="message-body"> <div class="think-section"> <div class="think-header" onclick="toggleThink(this)"> <span class="think-arrow">▼</span> <span>思考过程</span> <div class="loading-dots"><span></span><span></span><span></span></div> </div> <div class="think-content"></div> </div> <div class="answer-content"></div> </div> `; return el; } function removeCursor(msgEl) { msgEl.querySelectorAll('.cursor').forEach(c => c.remove()); } function toggleThink(header) { header.querySelector('.think-arrow').classList.toggle('collapsed'); header.parentElement.querySelector('.think-content').classList.toggle('collapsed'); }</script></body></html>
思考模型
如果业务需求需要大模型输出JSON格式，可以考虑使用仅思考模型，在回复前总会思考，后续有时间说这个方案