ScaiPersona

ScaiPersona is the persona product on top of ScaiGrid. You configure a persona — its underlying model, system prompt, RAG sources, default parameters — then publish it. Publishing creates a new frontend model in the catalogue that any caller can target by slug, exactly like a vendor model.

It is built on top of ScaiGrid's inference, model catalogue, and ScaiMatrix layers, so every persona invocation is metered, budgeted, audited, and routed through the same pipeline as any other inference call.

When to use it#

• You want a named, reusable assistant ("Legal Advisor", "Onboarding Coach") that downstream apps can call by model slug.
• You have ScaiMatrix collections or ScaiDrive shares the persona should retrieve from before answering.
• You want one inference call to do RAG against several sources with per-source weighting.
• You want the persona to appear in the standard model catalogue, model groups, and routing policies.

If you only need a one-off chat completion, you don't need ScaiPersona — call ScaiGrid's /v1/inference/chat directly with a vendor model slug.

What you get#

• Persona-as-model. Publish a persona and it shows up as a frontend model with slug tenant/{tenant_slug}/{persona_slug}.
• RAG over multiple sources. Attach ScaiMatrix collections and ScaiDrive shares; results are weighted, merged, and injected into the system prompt at call time.
• Three retrieval strategies. single_step, multi_step, and agentic — pick the trade-off between latency and answer quality.
• Status messages. Streaming clients see "searching…", "found N results…" events while RAG runs.
• Avatars. Upload a persona avatar; the published frontend model carries it.
• Draft / active / archived lifecycle plus an explicit unpublish path.

Two-minute mental model#

You manage two nouns and one verb:

• A Persona is a configured assistant. A tenant owns many personas.
• A Source is a knowledge attachment — a ScaiMatrix collection or a ScaiDrive share — bound to one persona.
• And the verb: you publish a persona, which materialises it as a frontend model that any caller can target.

Once published, callers never hit a ScaiPersona endpoint to use the persona — they call /v1/inference/chat with the persona's structured slug.

Where to go next#

• Quickstart — create, attach a source, publish, invoke. About five minutes.
• Architecture — how the persona, the enricher, and inference fit together.
• RAG strategies — single-step vs multi-step vs agentic.
• API reference — every endpoint, request, response.
• Build a knowledge-grounded persona — full walkthrough.

ScaiPersona's module ID inside ScaiGrid is scaipersona; its API is mounted at /v1/modules/scaipersona/.