Rex ChangGenAI & Cloud Solution Architect / Program LeadView portfolio ↗
RC

FAQ surge assistant

Instant answers to repeat customer questions on peak sale days.

Support spikes bury agents in the same questions.

  1. Sale days flood the support queue.A sale or product launch brings thousands of near-identical questions at once.Fix: Repeat questions are answered almost instantly, so queues don’t build up.
  2. Every answer costs time and money.Generating each reply from scratch keeps customers waiting and runs up model costs at peak load.Fix: Reusing saved answers cuts model calls, so a spike doesn’t multiply your costs.
  3. Answers pull from the wrong market.One shared search across languages can mix another market’s policies or prices into an answer.Fix: Each language searches only its own knowledge, and every answer cites its sources.
Watch it workHow it's built

See it work

Watch the assistant answer a support question end to end.

Watch on YouTube ↗
Full case studyPDF · 11 pages Executive briefPDF · 6 pages

Engineering note on Medium: Why “Just RAG” Breaks During Multilingual Support Surges ↗ · 8 min read

Code: faq-surge-assistant on GitHub ↗ · technical annex with routing, caching and observability

Under the hood

Search each language separately; reuse what’s already been answered.

  1. Customer questionIn the customer’s own language
  2. Language laneSeparate knowledge for each languagethis lab: English, Chinese, Thai
  3. Saved-answer checkReuse only on an exact match
  4. SearchPrimary path, backup when needed
  5. AnswerWith source linkse.g. FastAPI service, Gradio UI
// a saved answer is reused only when all of these match
reuse_key = {
  language,      // which language lane
  search_route,  // primary or backup path
  question,
  versions       // changed content → fresh answer
}
  • Repeat questions skip generation.A saved answer is served when the question, language, route and versions match.
  • Each language has its own lane.Retrieval stays aligned with the selected support content.
  • A backup search for harder cases.It can surface different evidence for complex or disputed questions.

Trade-off: reuse saved answers for speed; generate a fresh answer when the information needs updating.

Tested

Benchmarked across languages and search paths.

108–169×primary-path median speedup for repeated answers
12–14 mscached p95 response time
62–86×primary-path throughput gain (backup path 3.7–4.2×)

In context: best-performing lab setup, after warm-up, 20 requests per setup, compared with the corresponding uncached AI paths. The gains come from warmed repeat requests; the slower backup path can surface different evidence for complex or disputed questions.

Source links, routing details, traces and benchmark files support review of the answer path.

Next

Cut support wait times and costs on your busiest days.

I can scope a pilot around your knowledge base and channels. Hiring? Happy to walk you through how this was built.

Let’s talk →rex@rexchang.dev