# Release Notes: Smart Classroom ## Deprecation Notice - **Browser UI on Windows**: the desktop app is now the default way to run Smart Classroom on Windows, and `start-smart-classroom.ps1` launches it unless `-WebUI` is passed. The browser UI path (`setup-smart-classroom.ps1`, `start-smart-classroom.ps1 -WebUI`, and the Vite UI at ) is deprecated and will be removed after release 2027.0.0. See [Get Started on Windows](./get-started/get-started-windows.md). The Linux setup is not affected. ## Version 2026.2 **Release Date**: September 10, 2026 Smart Classroom 2026.2 refactors the backend around a **modular feature-module architecture** and adds two new classroom AI capabilities — **VLM-based exam grading with graded reports** and **Board (content-screen) OCR**. It also introduces a cross-platform **Flutter application** for Content Search. All new capabilities run locally on Intel hardware using OpenVINO™. **New**: - **Flutter Content Search application** — a cross-platform (Windows Desktop and Web) Flutter + Riverpod client for the Content Search backend, delivering file ingestion, RAG-powered Q&A with cited sources, multi-turn conversations, and file management. - Runs from its own dedicated `config.yaml` so it can be deployed independently of the main Smart Classroom application. - Talks to a lightweight **standalone VLM service** (Qwen3-VL-8B) instead of requiring the full application stack, for faster startup and lower resource use. - Adds an agentic **Coding Companion** mode with skills that drive setup, ingestion, Q&A, file management, and health diagnostics from natural-language commands. - **VLM-based exam grading and graded reports** — a new grading service that scores student papers against a teacher-provided rubric and produces per-student graded reports. - Runs as its own feature/service, gated by the `grading` feature flag. - Supports single-column and two-column exam page layouts. - Accepts student papers organized by directory (one folder per student). - Adds **Qwen3.5 9B** VLM support with configurable INT8/INT4 quantization. - Ships sample rubrics and sample exams for a guided quick test. - **Board OCR** — extracts text written on the classroom content/whiteboard screen from an RTSP stream or recorded video. - Extracts frames using FFmpeg + Intel QSV and feeds them to an OCR reader with text normalization and deduplication. - Integrated into the audio-summary pipeline so recognized board text enriches the generated class summary. **Improved**: - **Modular feature-module architecture** — the backend is now composed of self-contained feature modules (`asr`, `summary`, `mindmap`, `topic_segmentation`, `video_analytics`, `board_ocr`, `content_search`, `qa`, `grading`, `report`), each exposing a common interface (router, `build`/`teardown`, and a UI descriptor). - Features are toggled independently via the `features` block in `config.yaml`. - A dependency-ordered registry bootstraps only the enabled features and detects dependency cycles at startup. - **Unified Content Search Python environment** so that Content Search shares a single managed virtual environment with the rest of the application. - **Externalized LLM prompts** to make prompt tuning easier without code changes. - **Health checks** hardened against proxy interference. ## Version 2026.1 **Release Date**: June 17, 2026 Smart Classroom 2026.1, a modular, extensible framework for the Windows OS, adds a Content Search subsystem, document upload, text/image retrieval, OCR, QnA, and multilingual processing including Mandarin/Chinese. This release also adds WebRTC WHEP streaming, Intel® Core™ Series 3 (codenamed Wildcat Lake) platform support, and updates to audio transcription. **New**: - **Content Search module** for uploading documents and media, indexing them with OpenVINO™-accelerated embedding models, and retrieving results by text or image query. - **Q&A support** in Content Search for querying uploaded content with locally running LLMs. - **OCR** in Content Search with OpenVINO™ and PaddleOCR for printed and handwritten documents. - **Mandarin/Chinese language support** for QnA and transcription pipelines. - **File listing and file removal endpoints** in the Content Search API. - **Configuration API endpoint** in Content Search for runtime search and embedding settings. - **WebRTC WHEP streaming** for low-latency live video delivery. - **Intel® Core™ Series 3** processor support. - **Video start and end timestamps** in Content Search results for precise navigation. - **Resource utilization monitoring** to cap pipeline resource usage. **Improved**: - Long audio transcription is now supported with and without speaker diarization. - RTSP playback mode and file duration validation in the video ingestion pipeline. - Content Search embedding models upgraded to multilingual and cross-lingual models. - Batch embedding insertion for improved indexing throughput on large document sets. - Local PyAnnote audio model caching for offline diarization loading. - Pipeline startup reliability under concurrent load. - YOLO model configuration and quantization setting updates. - CMake configuration and build scripts for the gvasmartclassroom GStreamer plugin. - Error tracking for pipeline status reporting. - Video summarization output consistency. - UI behavior to stop active streaming sessions automatically after 10 minutes. - Support for Deep Learning Streamer (DL Streamer) 2026.1. - Summary scoring formula updates and reranker configuration refactoring. - Content Search download streaming optimization to reduce memory overhead for large file transfers. **Fixed**: - Crash in the video analytics pipeline. - Noise in per-class attendance statistics in the video pipeline. - Content Search cleanup when no task ID was present. - Duplicate file entries caused by missing task IDs during upload. - Inference request conflicts when embedding multiple documents concurrently. - Incorrect video scaling in the HLS player at some resolutions. - Encoding errors in video analytics output. - Incorrect YOLO model quantization results. - Content Search upload and delete operations under some conditions. - Corrupted file handling in Content Search that could cause indexing failures. - Hardcoded model name in the startup script; model name is now read from configuration. ## Version 2026.0 **Release Date**: April 1, 2026 The Smart Classroom application now offers after-class summary enhancements through next‑generation real-time audio and visual analytics. Teachers and schools gain a better understanding of classroom dynamics through AI‑driven summaries and engagement metrics. The Education AI Suite now also includes built-in telemetry hooks and benchmarking. **New**: - **Speaker Diarization** (via the Audio Pipeline): - identifies teacher and student speakers using NPU-accelerated diarization - generates an interactive audio timeline for replay and analysis - enables time-coded navigation within class video recordings - **Class Engagement Metrics – Audio**: - measure teacher and student speech duration - track questions asked and answered - track student-teacher interaction frequency - **Class Engagement Metrics – Video**: - track student hand raises - track posture changes (stand up/sit down) - track teacher movement - **Built‑in telemetry** to measure classroom workloads across Intel platforms (CPU core utilization, iGPU load, NPU load, memory usage, and workload-specific performance counters) - **Benchmarking scripts** to reproduce Intel internal performance measurements and validate XPU performance **Improved**: - **Knowledge Graph UI** readability and formatting, and increased clarity when visualizing topic relationships ## Documentation and Source Code - [GitHub](https://github.com/open-edge-platform/education-ai-suite/tree/release-2026.3.0/smart-classroom) ## Previous Releases - [Release notes 2025](./release-notes/release-notes-2025.md) :::{toctree} :hidden: Release Notes 2025 <./release-notes/release-notes-2025.md> :::