Skip to main content

Changelog

Upgrading from a previous version​

Why reset project settings?

When you install a newer LLMAI plugin in a project that already had LLMAI configured, the plugin and project usually compile and run as before—but provider defaults in your project INI are not upgraded automatically.

At startup, LLMAI loads {Project}/Config/DefaultGame.ini first, then merges Plugins/LLMAI/Config/DefaultGame.ini only into still-empty fields. Anything already saved in your project—Model, KnownModels, KnownVoices, API keys, instructions, and other LLMAI settings—remains unchanged. New shipped defaults in the plugin (including updated model lists) do not replace those entries.

Follow the reset steps below whenever a release changes provider defaults you rely on. Example (v2.1 → v2.2): OpenAI Realtime moved from Beta to GA; preview model names saved in an old project INI will not connect, while the v2.2 plugin defaults use gpt-realtime-2 (OpenAI) and grok-voice-think-fast-1.0 (Grok) with updated KnownModels lists.

Recommended steps

  1. Close the editor and back up {Project}/Config/DefaultGame.ini.
  2. Delete the entire [/Script/LLMAI.LLMAISettings] section before the first open with the new plugin version.
  3. Open the project. LLMAI settings repopulate from the shipped plugin defaults.
  4. Re-enter API keys, custom Instructions, LocalAI URLs, and any other values you had before.
warning

Deleting the section clears all LLMAI Project Settings in that file—not only models.

See also Getting started — Upgrading the plugin? and Configuration reference — Upgrading after a plugin update.

Releases​

Version 2.2​

OpenAI

  • Migrated OpenAI Realtime API from deprecated Beta to GA (removed OpenAI-Beta header; updated session JSON)
  • Removed session temperature (not supported on GA); updated provider profile and defaults
  • Added models gpt-realtime, gpt-realtime-mini, gpt-realtime-1.5, gpt-realtime-2 (default)
  • GA sessions use output_modalities as text or audio only—not both; plugin maps dual requests accordingly (transcripts still available in voice mode)
  • Fixed "Error committing input audio buffer: buffer too small" on voice sessions
  • Fixed empty voice dropdown on fresh projects without a project Config/ folder

Grok (xAI)

  • Added models grok-voice-think-fast-1.0 and grok-voice-latest
  • Moved Grok voices and models to plugin INI (consistent with OpenAI; no longer hard-coded in C++)
  • Fixed duplicate session update sent on both session.created and conversation.created
  • Known issue: complex tool-related requests may leak reasoning into voice output (x.ai server-side; reported)

General

  • Added GetProviderCapabilities / IsProviderFeatureSupported Blueprint API to query per-provider UI features (e.g. session temperature)
  • Runtime settings overlay: Blueprint SetDefault* calls apply for the session by default; optional bPersistToProjectConfig to write project INI
  • Breaking: removed unused legacy delegates OnAITranscript, OnAIResponse, OnAIFunctionCall (use current event delegates instead)
  • Cleaned and refined debug logging (quieter default Output Log)
  • Connect dialog: temperature slider disabled for OpenAI and other providers without session temperature support

Documentation

  • Rebuilt LLMAI documentation: corrected content, new theme, site search, index page, breadcrumbs, and refactored layout

Demos

  • Arcade demo: camera follows player; reduced SFX volume to avoid conflict with voice chat
  • Terminal demo: fixed chat UI artifacts when scrolling or on rapid transcript updates

Version 2.1​

  • Added Grok (xAI) provider support with voice agent integration
  • Added Voice parameter to connection functions - REQUIRED for Grok provider to set voice at connection time
  • Added enhanced rate limit error handling with provider-specific guidance (HTTP 429 detection)
  • Added Grok-specific features : web search, 5 voice options, text-only mode filtering
  • Fixed voice locking behavior for Grok provider (voice must be set at connection)
  • Fixed Audio stream debug messages always displaying even when turned off
  • Improved error messages for connection failures and quota issues
  • Updated LocalAI distribution to 3.9.0 Lite with improved performance

Version 2.0​

  • Added LiveLink voice support to enable lip-sync and function calling for MetaHumans with demo
  • Added support for LocalAI system with Realtime API compatible interface and demo update with providers
  • Added mic gating function allows using voice output without feedback to voice input
  • Added mic mute control in demo projects
  • Updated UI to support multiple providers (LocalAI supported)
  • Added release version of OpenAI Realtime API 5.0
  • Added think and reasoning LLM support with chat panel in demo projects
  • Note: Removed support for Unreal Engine 5.5, since engine has major subsystem differences

Version 1.0​

  • Initial Release

Support​