Model Compatibility

April 6, 2026 · View on GitHub

Current Standard

HelpdeskAI is now standardized on a two-model baseline:

  • v1 single-agent route: gpt-5.3-chat
  • v2 multi-agent workflow route: gpt-5.2-chat

This is the pairing currently used across app defaults, deployment defaults, demo collateral, and the Azure environment.


Why This Repo Is Pinned

HelpdeskAI depends on three things from the model layer:

  1. Native OpenAI-compatible tool calling
  2. Reliable multi-step instruction following across long support workflows
  3. Stable behavior when AG-UI streaming, MCP tool results, and workflow handoffs are combined in the same turn

The codebase now assumes the gpt-5.3-chat / gpt-5.2-chat pairing unless explicitly overridden.


Supported Baseline Matrix

RouteModelCurrent roleOperational note
v1gpt-5.3-chatPrimary single-agent routeDefault production baseline
v2gpt-5.2-chatMulti-agent handoff workflowValidate with the regression suite after any workflow or frontend streaming change

Other models are no longer part of the documented baseline for this repo. If you change either deployment, treat that as a regression event and re-run the suite in docs/regression-suite.md.


Known Caveats

  • v2 is still the higher-risk route because it combines orchestration, handoffs, frontend tool forwarding, and AG-UI streaming.
  • Content-safety failures now clear thread-scoped transient state and force a fresh frontend chat instance, so recovery after a blocked turn should be validated as part of any release check.
  • Attachment flows should always be re-tested after workflow or transport changes because they depend on thread continuity and one-shot Redis staging.

Configuration

The active model defaults are configured through AgentHost settings:

  • AzureOpenAI.ChatDeploymentgpt-5.3-chat
  • AzureOpenAI.ChatDeploymentV2gpt-5.2-chat

In Azure Container Apps, the same values are passed through AzureOpenAI__ChatDeployment and AzureOpenAI__ChatDeploymentV2.