AI and knowledge

Tasks and providers

Task	Supported providers / purpose
agent	OpenAI, OpenRouter, Anthropic, Google, xAI, DeepSeek; dashboard assistant
prescreening	Those language providers or TypeSafe/Jev; classification, fixed tags and email type choice
spam	TypeSafe/Jev, contextual Hugging Face classification, Decision HTTP; genuine-support admission
prereply	Language providers; suggested agent replies
translation	Language providers; product text, ticket and public-reply translation
embedding	OpenAI, OpenRouter, Qwen, Jina, Cohere, Google
rerank	OpenRouter, Cohere, Jina

Jev returns typed decisions, not summaries or new issue prose. A classifier or Decision HTTP credential cannot generate replies, translations or embeddings. OpenAI supports Responses (default) or Chat for compatible gateways; OpenRouter uses Chat. Model IDs are verified against task capability and runtime output. Custom endpoints require public HTTPS without query strings, fragments or redirects.

Model pickers search IDs and display names without changing the selected model. Use arrow keys and Enter to choose, Escape to close, or refresh the provider list; custom model IDs remain editable. OpenRouter discovery combines its text, embedding and rerank catalogs. Embeddings with incompatible or unverified index dimensions remain visible with an explanation and cannot be selected from the list.

Scoped configuration

Create credentials separately from function routes. A route is an ordered list of credential/model assignments; one credential can serve multiple compatible functions. Product routes inherit tenant, then system settings unless overridden, including explicit disable. System AI is SuperAdmin-only; tenant/product owners manage their permitted scope. Parent credentials can be used in child routes, without revealing saved keys. Secret fields remain encrypted at rest.

An enabled function needs an active assignment. Multi-language products need a translation route. Email AI prescreening needs a prescreening route. Genuine-support filtering is separately enabled in spam settings. This gate runs before customer web/API/email writes; it is not the same as post-creation ticket insights.

Failover and usage

Runtime skips disabled/cooling credentials and tries the next eligible assignment on provider or protocol failure. A failed key enters configurable cooldown only when another candidate is available; the final candidate is not cooled down. A successful response is never repeated just because bookkeeping failed. Metadata edits/disable remain possible during provider outages; changing a key, provider or endpoint can require route compatibility checks.

Each request records credential, provider, actual model where returned, success and token usage at system, tenant and product dimensions. Retention is configurable in days or permanent; cron removes expired detail/rollups. Classification providers may report no tokens; zero reported tokens is not a free-of-charge claim.

Knowledge and coordinate spaces

Structured product knowledge is indexed in product/model-specific namespaces. The embedding dimension comes from the bound Vectorize index (example configuration uses 1024; runtime supports up to 1536). Vectors must have exactly that many finite values; OnFire never pads or truncates an incompatible result.

Embedding fallback keys must use the same provider, endpoint, model and dimensions. Changing the effective embedding space queues a durable, leased rebuild, including inherited changes. The knowledge tab shows progress and can request rebuilding. Pending entries remain queryable through scoped D1 fallback; reranking is bounded and preserves retrieval order when unavailable.

Source document limit

Uploads accept PDF, text, Markdown, DOC and DOCX up to 10 MB. A document marked ready means its original file is safely stored in R2. Automatic document text extraction and indexing are not implemented. Use structured knowledge entries for retrieval; do not assume a successful upload made a file searchable by AI.