La documentation est disponible en anglais.

Governed pipeline

Modules run in a fixed order before the LLM. Policies decide which modules are on for a given API key — the order does not change per client.

Whether the call enters via Native /pipelines/run, OpenAI-compatible egress, Anthropic Messages, SDK ask(), or an MCP tool, it is the same control plane. That is what makes Connect recipes and Build paths interchangeable.

Module order (conceptual)

Cache
RAG
Compression
PII
LLM

Cache hit

Later modules (and the LLM) may be skipped. You still get an audited operation in Flux IA.

Full miss

RAG may inject context, compression may shrink prompts, PII may redact, then the provider runs.