Dattavani Backend: Ask Tataji Chat Diagram

What it is for

POST /api/chat is the one Ask Tataji endpoint for text, audio and video answers.
CHRAG, reached at AI_SERVICE_URL, writes the answer. core-service owns the gates, credits, XP and storage.
Both turns are saved in messages under a conversations row owned by the kid.
GET /api/conversations/list and /api/conversations/{id} read the history back.

Gates before CHRAG

The kid JWT is checked first (get_active_kid_context). Then the handle_chat_setup RPC checks membership, fair usage and daily minutes.
Fair usage defaults: 20 questions in a 4-hour window, then a 4-hour block (app_config).
Audio and video need the audio_video flag, a premium parent, room under the kid's daily credit cap and a balance (audio 1, video 2).
A home curated prompt whose text matches is free and served from its cached row.

Gotchas

AI_GATEWAY_MAX_CONCURRENT builds app.state.ai_semaphore, but no code uses it: CHRAG calls are not capped.
The CHRAG circuit breaker lives in each process, so every API instance trips it on its own.
Credit deduction reads a cached balance and writes balance minus cost, so two paid turns at once can under-charge.
The XP task starts before CHRAG answers, so a failed answer can still award XP.

Where to change it

Flow: routers/ask_tataji.py chat() and the _handle_audio_chat / _handle_video_chat helpers.
Costs and timeouts: app_config keys tataji_ai_timeout_sec, credit_cost_per_tataji_audio, tataji_video_min_credits_before_start.
Circuit: CHRAG_CIRCUIT_* settings in app/config.py. Curated cache: curated_prompt_cache_service.py.
Safety: SENSITIVE_RESPONSE_REASONS in tataji_sensitive_email_service.py.