v2.4

EduTelligence v2.4

Backfilled parity release for Artemis 9.4, whose release was published on 2026-06-11 21:42:51 UTC.

Compatibility

This release is compatible with Artemis 9.4.x.

Breaking Changes and Upgrade Notes

• Iris global search now uses the asynchronous pipeline endpoint POST /api/v1/pipelines/global-search/run with status callbacks to Artemis. The legacy /api/v1/search/ask endpoint was removed, and global-search responses may return answer: null for keyword/navigation queries that skip the LLM. This aligns with the Artemis 9.4 inline Iris answer card. #523
• Iris slide-page extraction requires the matching Artemis slide-page mapping support from Artemis #12682, included in Artemis 9.4. #548
• Logos was split into the Python logos-orchestrator plus a Spring logos-webservice for identity/config/statistics operations. Deployments using the repository compose files should update the full compose stack, include the logos-webservice image, run the included database migrations, and set a shared LOGOS_INTERNAL_SECRET if internal orchestrator refresh notifications are required. #597

Artemis 9.4 also introduced a REST path migration for account, course, and admin endpoints. Artemis keeps legacy aliases with Deprecation/Sunset headers until 2026-09-30, but external integrations should move to the new module-prefixed Artemis paths.

Highlights

• Iris: global search gains intent routing, async webhooks, course filters, slide display-page extraction, and verbalized-confidence support. #523, #575, #548, #507
• Logos: host-RAM-aware capacity planning, better sleep/load failure handling, vLLM 0.22 with compile-cache recovery, automatic GPU-wedge rebooting, token costs, benchmarking, and the Spring webservice migration. #577, #583, #580, #593, #572, #597
• Athena: PyTorch security update, updated test URLs, and an LLM_DEFAULT_MODEL override for base LLM model selection. #566, #574, #588

Iris

Features

• Add verbalized confidence scoring for the autonomous tutor pipeline by @toukhi in #507
• Add optional courseIds filtering to lecture search by @Nayer-kotry in #575
• Detect and persist slide display page numbers, plus slide descriptions, during lecture ingestion by @vivienxr in #548
• Add global-search intent routing, HyDE retrieval, async callbacks, and the global-search pipeline endpoint by @Nayer-kotry in #523

Athena

Improvements

• Update torch to v2.8.0 for security fixes by @renovate in #566
• Update Athena test URLs to aet.cit.tum.de by @maximiliansoelch in #574
• Allow overriding the base LLM model with LLM_DEFAULT_MODEL by @maximiliansoelch in #588

Logos

Features

• Make the capacity planner host-RAM-aware and add worker host-memory telemetry by @wasnertobias in #577
• Add token-cost tracking by @alex7sz in #572
• Add GSM8K benchmarking tooling by @wasnertobias in #591
• Add GPU watchdog reboot support for unrecoverable worker-node GPU wedges by @wasnertobias in #593
• Migrate Logos identity/config/statistics operations into the Spring webservice and rename the Python service to logos-orchestrator by @alex7sz in #597

Improvements and Fixes

• Add kv_cache_dtype to vLLM model overrides. direct commit b6e2f2b
• Fix calibrated-profile VRAM double counting and silent cuda:0 fallback. direct commit 31952b0
• Rebuild the classifier when workers announce new models and raise the vLLM readiness timeout. direct commit 0ec269d
• Stabilize worker-node CI tests by @wasnertobias in #578
• Bump vLLM to 0.22.0 and automatically purge poisoned compile caches by @wasnertobias in #580
• Loosen GPU auto-placement headroom by @wasnertobias in #582
• Add load-failure cooldowns, host-RAM-gated sleep escalation, calibrated sleep-transient memory, and quieter tracebacks by @wasnertobias in #583
• Fix calibration orchestration and quiet idle polling. direct commit 148cadb
• Enhance the Logos UI by @alex7sz in #581
• Reorganize the statistics dashboard VRAM panel by @JakubJakubczyk in #586
• Dedupe per-model statistics at the SQL stage by @JakubJakubczyk in #595
• Fix the statistics VRAM count and by-model chart by @JakubJakubczyk in #600
• Add and refine benchmark workflows by @wasnertobias in #596, #602
• Fix VRAM statistics OOMs with DB-side downsampling. direct commit 1e6dda2
• Rename logos-server to logos-orchestrator and fix auto-deploy image tags. direct commit da549c3
• Pin isort first-party packages to fix CI/local disagreement. direct commit 7d3372d

Full Changelog: v2.3...v2.4