Support/Developer Platform/AI ingestion, rebuilds, and multimodal search
/

Developer Platform

AI ingestion, rebuilds, and multimodal search

Use the secured tenant AI surface for authorized record ingestion, rebuild operations, and Qwen3-VL search.

Scopes
ai.read · ai.write
Model
Qwen3-VL-Embedding-2B
Vector lanes
256 · 1024 · 1998

The AI API is app-bound and permission-aware. It does not give an API client a privileged copy of tenant records: runtime identity, DataRole, location, permission keys, and CanAiReadRecords are resolved before context is assembled.

Current routes

GET  /v1/apps/{saasAppId}/ai/status
POST /v1/apps/{saasAppId}/ai/sources/records/{recordId}/ingest
POST /v1/apps/{saasAppId}/ai/rebuilds
GET  /v1/apps/{saasAppId}/ai/rebuilds/{jobId}
POST /v1/apps/{saasAppId}/ai/search

Ingestion and rebuilds are operations

Queue work, retain the returned job/correlation evidence, and poll the documented status resource. Do not repeatedly enqueue a new rebuild because one request timed out. Search is read-only; ingestion and rebuilds require ai.write.

Multimodal evidence

The target embedding model is Qwen/Qwen3-VL-Embedding-2B. Documents preserve visual evidence for tables, charts, diagrams, and screenshots, with supported 256, 1024, and 1998-dimensional lanes. Clients choose documented search modes rather than uploading arbitrary vectors or assuming legacy dimensions.

Permission behavior: AI retrieval uses the normal secured record path. Records outside the resolved identity's access never enter prompt, retrieval, reranking, citation, or response context.