Website AI Assistant: Capability Boundaries & Technical Design

1. 1
   User sends

   The front end collects messages and a sessionId, uploading to /api/chat over encrypted WSS/HTTPS.

2. 2
   Intent routing

   The back end detects intent (price / eye / travel / membership / emergency) via keywords and a classifier, then assembles context and KB snippets.

3. 3
   LLM chat API · generation

   The back end calls the LLM API (with guardrails and RAG) and streams via SSE; escalation conditions route to human agents.

4. 4
   Front-end rendering

   The front end renders replies incrementally and logs sessions and satisfaction for iteration and self-repair.

The R&D assistance pipeline runs independently: a developer files a ticket (new intent / bug) → the agent generates or fixes code → produces a Pull Request → review and merge into production.

The following is a generic example illustrating the integration approach only,not production code of this site, and has not been deployed.

Back-end message service + LLM integration (Node 18+, requires express and openai）：

// server.js
import express from 'express';
import OpenAI from 'openai';
const app = express();
app.use(express.json());
const openai = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });

// Medical compliance guardrails: no diagnosis / prescriptions / efficacy promises
const GUARDRAILS = [
  'You are the AI assistant of a medical travel concierge site; provide objective process and credential information only.',
  'Never diagnose, prescribe, promise outcomes, or rank doctors.',
  'Actual treatment requires an in-person visit with a licensed physician; costs follow hospital pricing.',
  'For emergencies, guide users to local emergency numbers (120 in China / 911 in the US).',
  'When unsure, escalate to a human agent.'
].join(String.fromCharCode(10));

// Intent routing: keyword hit decides the class, otherwise general
const INTENTS = {
  price: /价格|比价|费用|报价|usd|\$/i,
  eye: /眼科|近视|激光|smile|lasik|icl|白内障/i,
  travel: /行程|签证|机票|酒店|陪护/i,
  member: /会员|登录|账号|保险/i,
  emergency: /紧急|急救|意外|危险/i
};
function route(text){ for(const k in INTENTS) if(INTENTS[k].test(text)) return k; return 'general'; }

app.post('/api/chat', async (req, res) => {
  const { sessionId, message, lang = 'zh-CN' } = req.body;
  const intent = route(message);
  const kb = await retrieveKB(intent, message); // vector retrieval placeholder
  const stream = await openai.chat.completions.create({
    model: 'gpt-4o-mini', stream: true,
    messages: [
      { role:'system', content: GUARDRAILS + '\nKB snippets:\n' + kb },
      { role:'user', content: '[' + lang + '] ' + message }
    ]
  });
  res.setHeader('Content-Type','text/event-stream');
  for await (const chunk of stream) {
    const t = (chunk.choices[0] && chunk.choices[0].delta && chunk.choices[0].delta.content) || '';
    if (t) res.write('data: ' + JSON.stringify({ delta: t }) + String.fromCharCode(10) + String.fromCharCode(10));
  }
  res.write('data: [DONE]' + String.fromCharCode(10) + String.fromCharCode(10));
  res.end();
});

async function retrieveKB(intent, q){ return '(Example) public credential and price-range notes for this intent.'; }
app.listen(3000, () => console.log('AI chat service on :3000'));

Front-end streaming call (browser):

// chat-client.js
export async function sendChat(sessionId, message, lang, onDelta) {
  const res = await fetch('/api/chat', {
    method:'POST', headers:{'Content-Type':'application/json'},
    body: JSON.stringify({ sessionId, message, lang })
  });
  const reader = res.body.getReader();
  const dec = new TextDecoder();
  while (true) {
    const { done, value } = await reader.read();
    if (done) break;
    for (const line of dec.decode(value).split(String.fromCharCode(10))) {
      if (line.startsWith('data: ')) {
        const payload = line.slice(6);
        if (payload === '[DONE]') return;
        onDelta(JSON.parse(payload).delta);
      }
    }
  }
}

Code generation / repair module (dev support, planned) (Python, requires openai, replace model when wiring in the R&D assistant agent):

# codex_agent.py
import os, openai, sys
client = openai.OpenAI(api_key=os.environ['OPENAI_API_KEY'])

def generate_or_fix(task, file_path=None, failing_test=None):
    ctx = ''
    if file_path and os.path.exists(file_path):
        ctx += '\n# Existing file ' + file_path + '\n' + open(file_path, encoding='utf-8').read()
    if failing_test:
        ctx += '\n# Failing test output\n' + failing_test
    prompt = ('You are the code engineer for the customer-service system. Task: ' + task +
              '\nOutput only a mergeable code diff or the complete file, with a brief explanation.\n' + ctx)
    r = client.chat.completions.create(
        model='gpt-4o',  # replace model when wiring in the actual code model
        messages=[{'role':'system','content':'You are a senior frontend / Node engineer familiar with this customer-service system.'},
                  {'role':'user','content':prompt}])
    return r.choices[0].message.content

if __name__ == '__main__':
    print(generate_or_fix(sys.argv[1], sys.argv[2] if len(sys.argv) > 2 else None))