yw-aida

An AI-powered sales / customer-service training & evaluation platform built on Next.js 15 + React 19 + Prisma + MySQL + Redis + VolcEngine RTC/ASR/TTS.

yw-aida provides real-time voice training rooms (RTC + ASR + TTS), scenario-based AI coaches, batch evaluation, admin workstation and organizational management — all in a single full-stack Next.js application.

• Real-time voice training via VolcEngine RTC, streaming ASR and TTS
• Scenario/agent-based AI coaches with configurable prompts
• Batch evaluation pipeline (BullMQ + Redis + cron scheduler)
• Organization / user / invite / bonus-minute management
• Object storage via VolcEngine TOS (private + public buckets)
• JWT-based auth, phone-number SMS login (Aliyun SMS)
• WebSocket + SSE push, Prisma 7 data layer, Pino structured logs

A full end-to-end product tour, told through one fictional company — YunQi Tech (云启科技), a SaaS vendor whose sales team trains on AIDA. The four scenarios below map to AIDA's closed loop: author → practice → evaluate → manage, each from the point of view of the person who lives that step.

The UI is Simplified Chinese. A shot-by-shot script, the buyer-facing talking points, and a one-command script to generate this exact demo data (scripts/seed-demo-data.ts) all live in docs/DEMO_SCENARIOS.md.

A training lead drops in a plain sales-playbook document; AIDA's LLM auto-extracts the scene setup, customer persona, and scoring rubric and publishes a ready-to-practice scenario in minutes — no scripting, no engineering. Content cost drops from person-days to minutes, and non-technical staff can ship new courses themselves.

Reps rehearse against an AI customer that balks at price, stalls, and probes for specifics — by real-time voice (VolcEngine RTC + streaming ASR + TTS) or text — available 24/7 with zero risk to real leads. Pick a coach (Sam / Anna) and go; the AI plays a budget-sensitive procurement manager who keeps pushing on "too expensive".

Text mode — live LLM conversation	Voice mode — RTC + ASR + TTS

Every session produces a multi-dimension scorecard: an overall score, per-skill ratings (needs discovery, value framing, objection handling, …), turn-by-turn coaching with model answers, and a score trend over time — turning soft skills into measurable, comparable data instead of gut feel.

Bulk-import the org tree, allocate per-team quotas (voice/text minutes, KB & training storage), and track adoption and consumption on a dashboard — internal training at scale that stays deployable, governable, and cost-controlled.

Adoption dashboard	Org tree & quota management

┌────────────────────────────────────────────────┐
│                 Next.js App (SSR)              │
│  Pages / App Router / API Routes / WebSocket   │
└───────────────┬──────────────────┬─────────────┘
                │                  │
        ┌───────▼──────┐    ┌──────▼──────┐
        │    Prisma    │    │   BullMQ    │
        │   (MySQL)    │    │   Workers   │
        └───────┬──────┘    └──────┬──────┘
                │                  │
        ┌───────▼──────┐    ┌──────▼──────┐
        │   MySQL DB   │    │    Redis    │
        └──────────────┘    └─────────────┘

External: VolcEngine RTC/ASR/TTS/TOS/Ark · Aliyun SMS · ipinfo.io · Zhipu LLM

• Node.js 24+ (the project uses Next.js 15 features that assume Node 20+; CI runs on Node 24. Node 22 has been seen to work locally but is not officially supported.)
• MySQL 8.x
• Redis 6+
• (Optional) a private rtc-server repository for RTC signaling, cloned into the project root as rtc-server/

git clone <your-fork-or-repository-url>
cd yw-aida

cp .env.example .env.local
# edit .env.local and fill in real values (see "Environment Variables" below)

npm install
npx prisma generate
npx prisma migrate deploy    # or: npm run migrate:db
npm run dev                  # day-to-day dev
# or, first-time bootstrap (also tries to clone rtc-server if RTC_SERVER_REPO_URL is set):
npm run init

The first render will fail fast if required env vars (JWT keys, DB, Redis, RTC/ASR tokens, SMS keys) are missing. Fill them in .env.local first. This repository is open for GitHub-based source collaboration. The application package remains "private": true, so cloning the source and publishing an npm package are intentionally separate workflows.

Organizations get account authorizations and resource quotas (message / voice training minutes, KB storage, training storage) through a super-admin only recharge flow. Two entry points:

1. Platform-admin permission on the caller user:

   UPDATE users SET platform_admin_type = 1 WHERE phone = '<your-phone>';

2. Auth token (JWT) — log in as that user in the browser, then copy the auth_token_admin cookie value (DevTools → Application → Cookies).
3. A running app — npm run dev on https://localhost:3000/ (default), or a deployed instance.

# 1. Put the token in .env (or export in shell). Do NOT wrap the value in
#    non-ASCII quotes — only the raw JWT, no spaces:
echo 'RECHARGE_AUTH_TOKEN=eyJhbGciOi...your-jwt...' >> .env
# Optional: override the target host (defaults to https://localhost:3000/)
# echo 'RECHARGE_API_BASE_URL=https://staging.example.com/' >> .env

# 2. Edit the RECHARGE_LIST array at the top of the script with your orgIds
#    and the amounts you want to top up:
#    scripts/recharge-authorization.ts

# 3. Run it:
npx tsx scripts/recharge-authorization.ts

The API (POST /api/org/authorization/recharge) accepts user-facing units (minutes / GB / days); the backend handles the minute→second conversion before persistence. Never pre-multiply by 60 on the client side.

The script validates the HTTP status and the business-layer success / code fields, so a silent failure on a 200 response will still be surfaced as 业务失败 code=400: ....

-- Authorization coupons
SELECT * FROM b_org_authorization_transaction
WHERE transaction_type = 1 ORDER BY create_time DESC LIMIT 20;

-- Quota top-ups
SELECT * FROM b_org_quota_transaction
WHERE transaction_type = 1 ORDER BY create_time DESC LIMIT 20;

See docs/RECHARGE_GUIDE.md for the full permission model, Web-UI walkthrough, warning thresholds and operational safety notes.

All configuration lives in .env.local (loaded by Next.js + dotenv). Start from .env.example — it lists every variable used across the codebase with <your-xxx> placeholders.

Major groups:

• App / runtime: APP_ENV, NODE_ENV, PORT, NEXT_PUBLIC_BASE_PATH, …
• MySQL / Prisma: DATABASE_URL, MYSQL_HOST, MYSQL_USER, …
• Redis: REDIS_URL or REDIS_HOST / REDIS_PORT / REDIS_PASSWORD
• JWT: JWT_PRIVATE_KEY, JWT_PUBLIC_KEY (base64)
• Aliyun SMS: ALIYUN_ACCESS_KEY_ID, ALIYUN_ACCESS_KEY_SECRET, ALIYUN_SMS_SIGN_NAME, ALIYUN_SMS_TEMPLATE_CODE
• VolcEngine (RTC / ASR / TTS / TOS / Ark): NEXT_PUBLIC_RTC_APP_ID, NEXT_PUBLIC_ASR_APP_ID, NEXT_PUBLIC_ASR_TOKEN, VOLC_TOS_*, NEXT_PUBLIC_VOLC_ARK_ENDPOINT_ID, …
    TOS 上传/CORS 排障与桶配置见 docs/TOS对象存储实现总结.md
• LLM: ZHIPU_KEY, ZHIPU_MODEL, OPENAI_API_KEY, …
• IP geolocation: IPINFO_TOKEN
• Batch evaluation: BATCH_EVALUATION_API_KEY, DIFY_EVALUATION_BASE_URL

The RTC signaling server lives in a separate private repository. Set RTC_SERVER_REPO_URL in your environment before running npm run init, or clone it yourself into ./rtc-server and run npm ci inside. If the directory is absent at runtime the main app will still build and start, but real-time voice training will be degraded.

• Docker: build from the included Dockerfile (node:24-alpine)
• CI: GitHub Actions — see .github/workflows/ci.yml

This project was extracted from an internal codebase. A few rough edges remain — contributions that progressively clean them up are very welcome:

• ESLint warnings/errors: npm run lint currently reports a number of pre-existing violations (no-require-imports, no-empty-object-type, prefer-const, a few hook-dependency issues). They do not affect npm run build. CI runs lint for visibility but does not fail the build on them.
• Node version: CI pins Node 24. Older versions may work but are untested.
• rtc-server: lives in a separate private repo. Without it the app still builds and non-RTC features work, but real-time voice training is degraded.
• Translations / copy: some user-facing strings are Simplified Chinese only.

See CONTRIBUTING.md if you want to help.

See CONTRIBUTING.md and our Code of Conduct.

Please do not file public GitHub issues for security vulnerabilities. See SECURITY.md for our responsible disclosure policy.

MIT © yw-aida contributors