Release Notes#

Version 2026.0.0 (March 2026)#

General Availability Release

Promoted from 2026.0-rc2 with no code changes. All functionality is identical to 2026.0-rc2.

Published Images#

Image

Tag

intel/order-accuracy-dine-in

2026.0.0

Version 2026.0-rc2 (March 2026)#

What’s New#

  • Image tag updated to 2026.0-rc2; TAG Makefile variable updated to 2026.0-rc2

  • FastAPI and Starlette version updates in the application image

  • setup_models.sh updated to OVMS 2026.0 export branch (releases/2026/0); openvino and openvino-tokenizers updated to 2026.0.0rc3

  • init-env target added to Makefile for creating .env from .env.example

Published Images#

Image

Tag

intel/order-accuracy-dine-in

2026.0-rc2

Version 2026.0-rc1 (March 2026)#

New Features#

  • Flexible Image Deployment: New REGISTRY flag for build and deployment control

    • make build REGISTRY=false - Build image locally from source

    • make build - Pull pre-built image from registry (default)

    • make up REGISTRY=false - Run with locally built image

    • make up - Run with registry image (default)

  • Docker Image Tagging: Standardized image naming

    • Image: intel/order-accuracy-dine-in:2026.0-rc1

    • Configurable via TAG and DINEIN_IMAGE_NAME variables

Configuration Changes#

Variable

Old Default

New Default

TAG

1.0.0

2026.0-rc1

REGISTRY

false

true

DINEIN_IMAGE

-

intel/order-accuracy-dine-in:2026.0-rc1

Migration Notes#

  • Update deployment scripts to use new REGISTRY flag

  • Image name changed from dine-in-dine-in to intel/order-accuracy-dine-in

  • Use REGISTRY=false for local development workflows

Version 2.0.0 (February 2026)#

New Features#

  • Circuit Breaker Pattern: Added fault tolerance for VLM and Semantic services

    • 5 consecutive failures trigger circuit OPEN state

    • 30s recovery timeout for VLM, 15s for Semantic service

    • Automatic recovery with half-open state testing

  • Connection Pooling: Shared HTTP clients with optimized settings

    • Up to 50 concurrent connections for VLM client

    • HTTP/2 support enabled for improved performance

    • Keepalive connections (30s expiry)

  • Bounded Validation Cache: LRU cache prevents memory exhaustion

    • Maximum 10,000 entries

    • Thread-safe operations

    • Automatic eviction of oldest entries

  • Image Preprocessing Pipeline: Optimized images for faster VLM inference

    • Smart resizing (672px max dimension)

    • Adaptive contrast enhancement

    • Light sharpening for food detail

    • JPEG compression (82% quality)

  • Stream Density Benchmark: New testing mode for concurrent validation

    • Automatic density scaling

    • Latency-based pass/fail criteria

    • Comprehensive results export (JSON/CSV)

Improvements#

  • Thread-safe Singleton: Config manager uses double-checked locking

  • Async Metrics Collection: Non-blocking system stats retrieval

  • Token Usage Logging: Detailed TPS and token metrics

  • Enhanced VLM Prompts: Inventory-aware prompts for better accuracy

Bug Fixes#

  • Fixed race condition in service initialization

  • Fixed unbounded Dict causing OOM under load

  • Fixed blocking psutil.cpu_percent() call

  • Fixed sync HTTP in Gradio callback

Configuration Changes#

Variable

Old Default

New Default

API_TIMEOUT

60

300

BENCHMARK_TARGET_LATENCY_MS

1000

2000

Dependencies#

  • Added: httpx[http2]>=0.25.0 (HTTP/2 support)

  • Added: aiofiles>=23.0.0 (async file operations)

  • Removed: requests (replaced by httpx)

Known Issues#

  • GPU memory utilization metric always reports 0.0 (metrics collector limitation)

  • First psutil.cpu_percent() call returns 0.0 (expected behavior)

Version 1.0.0 (January 2026)#

Initial Release#

  • Gradio UI for interactive validation

  • FastAPI REST endpoints

  • OVMS VLM integration (Qwen2.5-VL-7B)

  • Semantic matching service

  • Basic metrics collection

  • Docker Compose deployment

Features#

  • Single image validation

  • Batch validation endpoint

  • Order manifest comparison

  • Accuracy scoring

  • Performance metrics