Data Pipeline — ManyAI

Request Flow

User sends a message in the ManyAI mobile app.
isImageGenRequest() checks for image-related keywords in the prompt.
pickProvider() selects the first available provider from ROUTING_ORDER.
callProvider() sends the API request with a 30-second timeout.
On failure: add the provider to failedProviders and retry up to MAX_RETRIES = 8 with the next provider in ROUTING_ORDER.
If all retries fail, display an error message to the user.

Build the available provider set: include all providers with valid API keys; always include Pollinations.
Filter to VISION_PROVIDERS if an image is attached.
First pass: iterate ROUTING_ORDER, preferring providers with bestFor matching the task type.
Second pass: return the first capable provider (fallback to any available if no bestFor match).
Final fallback: return Pollinations if no other providers are available.
Return null if all providers are exhausted.

Provider	Method	Endpoint	Auth	Notes
Pollinations	GET	`text.pollinations.ai/{encoded_prompt}`	None	No key required; always available
Gemini	POST	`/models/{model}:generateContent`	API key in URL	`parts` array with `inline_data` for images
Anthropic	POST	`/messages`	`x-api-key` header	Different body structure from OpenAI
Cloudflare	POST	Account ID embedded in URL	Key split on `:`	Key format: `accountId:apiToken`
All others	POST	`/chat/completions`	`Authorization: Bearer`	OpenAI-compatible

Request timeout: 30 seconds per attempt.

Failed providers are tracked in failedProviders using useRef — scoped to the current session only.
Failed providers are skipped on all subsequent attempts within that session.
Conversation context: the last 10 messages are sent as context with each request; system messages and error messages are excluded.
On total exhaustion (all providers failed MAX_RETRIES times), a user-visible error is displayed and the session provider list resets.