ERP/bot

procurement_ai_agent — Bale chatbot for the procurement ERP

A Bale messenger bot that turns Persian chat messages and voice notes into procurement actions in the procurement ERP — draft-first: it captures what it understood, tells you what it did and what is still empty, and commits only after you approve with a ✅ button.

you (Bale, text/voice)                      ┌─────────────────────────────────┐
        │  getUpdates (long poll)           │        procurement_ai_agent     │
        ▼                                   │  ffmpeg → STT (Persian)         │
┌──────────────────┐   POST {base}/bot<T>    │  LLM → intent + entities        │
│   tapi.bale.ai   │◄───────────────────────►│  fuzzy resolve → draft payload  │
└──────────────────┘   sendMessage / buttons │  review + «چه چیزهایی خالی مونده»│
                                             └───────────────┬─────────────────┘
                                                             │ HTTP + X-Actor-Id
                                                             ▼
                                             ┌─────────────────────────────────┐
                                             │   ERP api  (/api/v1)            │
                                             │   quotes · suppliers · materials│
                                             └─────────────────────────────────┘

What it can do

Activity	Example (voice or text)	Committed via
Register a supplier quote	«سازان نو برای پروژه برج آسمان ۵۰۰ کیلو میلگرد A3 سایز ۱۴، هر کیلو ۴۵ تومن، پرداخت ۳ فاز»	POST /quotes
Add a supplier	«تأمین‌کننده جدید پروفیل بر، دسته لوله و اتصالات، تلفن ۰۹۱۲…»	POST /suppliers
Add a material	«کالای جدید میلگرد A3 سایز ۱۸ با واحد کیلوگرم»	POST /materials
Status queries	«استعلام‌های در انتظار؟» / «کدام‌ها به‌زودی منقضی می‌شوند؟»	GET /quotes?…
Accept a quote	«استعلام سازان نو را قبول کن»	PATCH /quotes/{id}/accept
Announce a delivery	«اقلام سازان نو رسید، ۵۰۰ و ۲۰۰»	PATCH /quotes/{id}/arrive-items

The draft → approve loop

1. Send a message or voice note. Voice → ffmpeg → STT; text/STT text → LLM extracts a typed intent + entities (JSON mode, Persian digits normalized).
2. Names are fuzzy-matched against the ERP. Several candidates → inline pick buttons; no match → listed as missing.
3. The bot replies with a review card: everything captured + a ✅/⚠️/❌ list of what is still empty.
4. Keep chatting to refine («قیمت میلگرد ۴۷ شد») — the open draft updates and re-renders. Drafts survive bot restarts (postgres-backed, TTL-swept).
5. Nothing is written to the ERP until you tap ✅. After commit you get the created quote id and the fields left empty.

Team access & attribution

Each member links their Bale chat once: /start → Casdoor username → password (validated via the ERP's real POST /auth/login, used once, never stored). Every mutating call carries X-Actor-Id: <users.id> so quote_events.actor records the human who asked.

Commands: /start · /help · /drafts · /cancel · /whoami · /unlink

Quick start (Docker)

cp .env.example .env          # fill PROCUREMENT_BOT_BOT_TOKEN + AI_API_KEY
docker build -t procurement-ai-agent .
docker run -d --name procurement-bot --env-file .env \
  --add-host host.docker.internal:host-gateway \
  procurement-ai-agent
docker logs -f procurement-bot

Or as part of the ERP stack — see integration/ (compose service + the one tiny ERP-side patch).

Full setup, env-var reference, flows and troubleshooting: docs/SETUP.md.

Architecture

app/
├── main.py          ensure bot tables → sweeper task → getUpdates long-poll loop
├── config.py        pydantic-settings, PROCUREMENT_BOT_ prefix
├── bale/client.py   Telegram-compatible Bale API (poll, send, edit, buttons, files)
├── api/client.py    ERP /api/v1 client (X-Actor-Id attribution, error envelope)
├── stt.py           ffmpeg (OGG/Opus → MP3) → OpenAI-compatible /audio/transcriptions
├── nlu.py           Persian intent+entity extraction (JSON-mode chat completions)
├── resolve.py       fuzzy name→id matching (difflib over ilike search), TTL caches
├── drafts.py        draft store · QuoteCreate payload builder · missing-field logic
├── keyboards.py     inline keyboards (callback_data ≤ 64 B op-codes)
├── texts.py         Persian message templates
├── runtime.py       context, link store, in-memory pending flows
├── models.py        bot_links / bot_drafts / bot_messages / bot_meta (own schema)
└── handlers/        dispatch · commands · callbacks · links ·
                     quote · supplier · material · queries · decisions

Design notes:

• Telegram-compatible Bale API (tapi.bale.ai): getUpdates long polling — no public port, no webhook, resumable offset persisted across restarts.
• Draft-first: bot_drafts.payload holds a talking-friendly JSON that is converted into the exact POST /quotes request body (temp_id cross-references) only on approval; the API deliberately tolerates incomplete quotes, so partial commits are explicit, never accidental.
• Accept flow echoes the termsheet's existing schedule_steps back to PATCH /quotes/{id}/accept (the API replaces steps with what it receives).
• Persian-aware matching: Persian/Arabic digit normalization, ZWNJ folding, Arabic ya/ka folding, difflib ranking on top of the API's ilike search.
• Voice notes arrive as OGG/Opus; ffmpeg converts to mono 16 kHz MP3 before STT. The LLM provider and the STT provider are configured independently (e.g. DeepSeek for NLU + any whisper-compatible endpoint for voice).

Testing

An offline end-to-end harness stubs Bale/ERP/AI and drives the real handler pipeline with fake updates — no token, DB or API key needed:

pip install -r requirements.txt
python tools/smoke_test.py     # → ALL PASSED: 48 passed