Release Notes: Visual Search and QA#
Version 2026.3.0#
Helm Chart Version: 1.1.0 Release Date: October 9, 2026
Changes:
Data Preparation:
Replaced the Milvus-specific data preparation microservice with the generic Multimodal DataPrep microservice, which is shared with other applications.
Media is embedded in place from the shared data mount, so ingesting a directory no longer stores a second copy of the media on disk.
Re-running an update against the same directory now skips already ingested files instead of duplicating them.
Object detection is disabled by default. Only full frames are indexed, which keeps search results free of crop duplicates. Set
MM_DATAPREP_ENABLE_OBJECT_DETECTIONandDATA_INGEST_WITH_DETECTto enable it.Object detection now uses Hugging Face OpenVINO models. Set
DATAPREP_DETECTION_MODEL(HelmMM_DATAPREP_DETECTION_MODEL_NAME) to anOpenVINO/<repo>object-detection model such asOpenVINO/yolox_s-int8-ov. It has no default and is required whenever object detection is enabled. The optionalDATAPREP_DETECTION_CLASSESkeeps only the listed classes. Downloaded models are cached in thedataprep-detection-modelsvolume; the formerdataprep-yolox-modelsvolume is no longer used and can be removed.Added
DATAPREP_PIPELINE_STALL_TIMEOUT_S(default 300s), which fails an ingest job whose pipeline stops making progress instead of leaving it running indefinitely. Set it to 0 to disable the watchdog.
Helm Chart:
The chart is no longer tied to a namespace named
vsqa. In-cluster service addresses are derived from the release namespace, so it can be installed into any namespace.Added
resourceson the VLM, embedding, and data preparation services to request Intel device plugin resources when running on GPU or NPU.The render device group is now applied to the pods, so a non-root container user can open
/dev/dri/renderD*.Added
global.nodeSelectorto pin the pods to the node that holds the host data path and the accelerator devices.
Deployment:
Inference devices are now selected per microservice in Docker Compose instead of being derived from
EMBEDDING_DEVICE. Setting a GPU for query embedding no longer moves the ingest pipeline onto the same device, where two copies of a large embedding model could exhaust it and make queries fail withCL_OUT_OF_RESOURCES.The dataprep devices are now set with
DATAPREP_EMBEDDING_DEVICEandDATAPREP_DETECTION_DEVICE, matching the otherDATAPREP_*settings. The previousMM_DATAPREP_*names are still honoured.Device defaults are now
CPUfor query embedding andGPUfor the VLM and the ingest pipeline. The previousGPU.1default did not exist on platforms with a single GPU.
Documentation:
Documented device selection, device plugin resources, device group ids, node pinning, and the corresponding troubleshooting steps for Helm.
Documented the per-service device variables, their defaults, and the single-GPU contention caveat for Docker Compose.
Documented
HOST_DATA_PATH,HF_ENDPOINTandVLM_COMPRESSION_WEIGHT_FORMATin the Compose guide.HF_ENDPOINTdefaults to a mirror intended for mainland China and has to be overridden to download models from Hugging Face directly.VLM OpenVINO Serving is now consumed as a published image instead of being built from source.
Hardware Used for Validation:
Intel® Core™ processors (13th Gen, i7 recommended)
Intel® Arc™ A-Series Graphics (Intel® Arc™ A770 recommended)
Version 2025.2.0#
Helm Chart Version: 1.0.0 Release Date: December 10, 2025
Features:
Visual Search:
Search images and videos using natural language queries.
Question Answering:
Ask questions about selected media and receive context-aware answers.
Microservices Architecture:
Includes data preparation, retriever, multimodal embedding, VLM serving, and a Streamlit-based web UI.
Hardware Used for Validation:
Intel® Core™ processors (13th Gen, i7 recommended)
Intel® Arc™ A-Series Graphics (Intel® Arc™ A770 recommended)