Concepts
AI Fallback
Hybrid mode uses bounded fallback only when local graph-based resolution is insufficient.
Need the shortest path?
Use Quick Start when you want the smallest working route → scan → graph → run flow first.
Something feels off?
Inspect scan output, graph state, active route/context, and dispatch in a practical order.
Need the full surface?
Browse all runtime methods, hooks, extension points, and lifecycle APIs.
Fallback exists to support requests the local runtime cannot confidently complete on its own.
Local-first rule
The runtime should always attempt local resolution first.
Fallback is only relevant when:
- AI is enabled
- the mode allows fallback
- an endpoint exists
- the local result is unknown or low-confidence
What fallback is for
Typical fallback cases include:
- weak local confidence
- analytical or open-ended questions
- phrasing that the local planners do not cover
- requests that still need to normalize back into the existing app graph
What fallback is not for
Fallback is not the product core. It is not there to hide weak tagging, broken graph structure, or ambiguous duplicates that should be fixed locally.
Fallback flow
- local resolution runs first
- the client builds a normalized request with runtime context and graph summary
- the server or provider returns a structured result
- the client normalizes that result against the real graph
- only usable, non-ambiguous normalized results continue to dispatch
What the browser now sends
Fallback requests can include:
- input text
- active route and active context
- graph summaries
- runtime selections
- injected runtime context
What fallback now guarantees
The fallback path is now bounded and observable.
That includes:
- timeout
- provider metadata
- auth mode
- error classification
- normalization success/failure state
- whether the result was actually used
Runtime events for fallback
Use:
fallback:startfallback:end
These events are the best way to inspect:
- why fallback was attempted
- how long it took
- whether it timed out
- whether the provider response normalized successfully
- whether a usable intent actually came back
Ambiguity rule still applies
A cloud result should not silently choose the wrong duplicate target.
If normalization cannot safely ground the provider output to one real node, the runtime should reject the fallback result instead of guessing.
Tier 3 is Relay
config.ai.endpoint points at Cinder Relay - a hosted, metered tier-3 proxy: tier 3 is usage-based, and you choose whether your own provider key or your app's end user's key is used per call.
Related pages
Continue reading