Support Bot (Maximilian)

Customer support automation for Studyflash using Gemini AI with KB-grounded responses.

Quick Reference

# Local dev
uvicorn main:app --reload --port 8000

# Deploy
# Auto-deploys on push to `main` of studyflash-ai/studyflash via Dokploy's
# GitHub integration. No manual deploy step.

# Adopt / drift-check the Dokploy resources via Pulumi
pnpm preview   # pulumi preview --diff
pnpm run pulumi:up # pulumi up

# Check logs
# Open the Customer Support project in Dokploy UI → Customer Support Bot →
# Logs tab. Same source-of-truth used by the Dokploy auto-deploy pipeline.

Architecture

Context Stuffing: Entire KB (~10 pages) fed to AI per request, no vector DB
Gemini 3 Flash: Used for classification and draft generation
Chatwoot: Ticket management integration
Dokploy: Deployed at https://dokploy.studyflash.ch (project: Customer Support). Public at https://support-bot.studyflash.dev.

Key Files

File	Purpose
`main.py`	FastAPI app, `/ask` endpoint
`services/draft_agent_service.py`	Gemini draft generation
`services/category_classifier_service.py`	Query classification
`services/kb_service.py`	KB loading from `kb/` directory
`models/draft.py`	Response schemas (no defaults in `DraftResponseSchema` - Gemini limitation)
`kb/*.md`	Knowledge base articles by category

API Endpoints

POST /ask - Main support query endpoint
POST /analyze - Weekly gap analyzer
POST /log-correction - Log human corrections

Secrets

Sourced from Infisical at /internal/support-bot/ (prod env), written to Dokploy's native env at pulumi up time. Live keys (per application.one):

AUTOMATIONS_ADMIN_KEY
AUTO_REPLY_ENABLED
CHATWOOT_ACCOUNT_ID
CHATWOOT_API_TOKEN
CHATWOOT_API_URL
GEMINI_API_KEY
METABASE_EMBEDDING_SECRET_KEY
PORT

Deployment Notes

KB files must be in .dockerignore whitelist (!kb/*.md).
Public URL: https://support-bot.studyflash.dev (port 8080 in-container, Let's Encrypt via Dokploy).
Pulumi adoption stack lives next to the source — index.ts, Pulumi.yaml, package.json in this directory.