Order Accuracy

A suite of Intel® edge AI applications for real-time order validation in Quick Service Restaurant (QSR) environments, powered by Vision Language Models (VLM) and Intel® OpenVINO™ inference.

Order Accuracy automatically detects items in food trays, bags, or containers, compares them against expected order data, and identifies discrepancies before orders reach customers.

Platform Applications

The platform provides two specialized applications optimized for different restaurant scenarios:

Application	Use Case
Dine-In	Restaurant table service validation
Take-Away	Drive-through and counter service

Choosing the Right Application

Criteria	Dine-In	Take-Away
Input Type	Static images	Video streams (RTSP)
Throughput	Low-medium	High (parallel)
Latency Priority	Accuracy over speed	Speed and accuracy
Camera Setup	Fixed position	Multi-station
Typical Use	Table service	Drive-through, counter
Processing	Single request	Batch processing

---

Dine-In Order Accuracy

Image-based order validation for full-service restaurant expo operations. Staff place a plate or tray in the validation station and trigger a check via the Gradio UI or REST API. The system analyzes plate contents with a pre-trained Qwen2.5-VL model and reconciles detected items against the order manifest — no model fine-tuning required.

Best for: table service, fixed-position cameras, accuracy-prioritized single-image validation.

For full details see the Dine-In User Guide. Alternatively, see the Dine-In README for a quick overview of the application.

Take-Away Order Accuracy

Real-time video stream validation for drive-through and counter service. The service ingests RTSP streams from multiple stations, selects relevant frames with a YOLO model, batches VLM requests, and validates orders in parallel across up to 8 workers — with circuit-breaker fault tolerance and automatic recovery for high-throughput deployments.

Best for: drive-through and counter service, multi-station camera setups, speed-and-accuracy continuous validation.

For full details see the Take-Away User Guide. Alternatively, see the Take-Away README for a quick overview of the application.

Documentation

Dine-In	Take-Away
Get Started	Get Started
How It Works	How It Works
How to Use	How to Use
Benchmarking	Benchmarking
API Reference	API Reference
Troubleshooting	Troubleshooting
Release Notes	Release Notes

---

Shared Platform Services

Both applications share a common set of services:

Service	Description
OVMS VLM	OpenVINO™ Model Server running Qwen2.5-VL with an OpenAI-compatible API
Semantic Service	AI-powered item matching (exact → semantic → hybrid) with local fallback
MinIO Storage	S3-compatible object storage for frames, selections, and results
Gradio UI	Web interface for manual validation and demos

The OVMS model files are exported once and shared by both applications. See the OVMS Service README for model setup and KV cache tuning guidance.

Get Started

Choose an application and follow its Get Started guide:

• Dine-In — Get Started
• Take-Away — Get Started

For hardware and software prerequisites, see the system requirements for the application you plan to deploy:

• Dine-In — System Requirements
• Take-Away — System Requirements

⚠ Model Export RAM Requirement: setup_models.sh performs INT8 quantization of Qwen2.5-VL-7B, which temporarily requires up to 40 GB of system RAM (FP16 model ~15 GB + INT8 compressed ~8 GB + calibration buffers ~8–15 GB). On platforms with 32 GB RAM (e.g. Wildcat Lake, Meteor Lake), the export OOMs and writes partial, corrupt XML files, causing the oa_ovms_vlm container to fail at startup with "Unable to read the model" errors. Always run setup_models.sh on a system with at least 48 GB RAM (64 GB recommended). The exported model files can then be copied to lower-memory systems for inference-only deployments.

ℹ OVMS KV Cache (cache_size): The default CACHE_SIZE=4 reserves 4 GB of VRAM for the KV cache. The INT8 model itself uses ~8 GB VRAM, so total VRAM ≈ 12 GB (fits in Intel Arc A770 16 GB). On integrated GPU platforms (Wildcat Lake, Meteor Lake), the KV cache is allocated from system RAM — on a 32 GB system this can exhaust all available memory. Use a smaller value (CACHE_SIZE=2) on iGPU platforms. Set export CACHE_SIZE=<N> before running setup_models.sh, or edit graph.pbtxt directly after export. See OVMS Service README — Tuning the KV Cache Size for a full sizing guide and per-platform recommendations.

---

Project Structure

order-accuracy/
├── dine-in/                     # Dine-In application
│   ├── src/                     # Application source code
│   ├── docker-compose.yaml      # Service orchestration
│   ├── Makefile                 # Build automation
│   └── README.md                # Dine-In documentation
│
├── docs/user-guide
│   ├── dine-in/                 # Dine-In user guide
│   └── take-away/               # Take-Away user guide
│
├── take-away/                   # Take-Away application
│   ├── src/                     # Application source code
│   ├── frame-selector-service/  # YOLO frame selection
│   ├── gradio-ui/               # Web interface
│   ├── docker-compose.yaml      # Service orchestration
│   ├── Makefile                 # Build automation
│   └── README.md                # Take-Away documentation
│
├── ovms-service/                # Shared OVMS configuration
├── performance-tools/           # Benchmarking scripts
└── README.md                    # This file

---

License

Copyright © 2025 Intel Corporation

Licensed under the Apache License, Version 2.0. See LICENSE for details.

Support

For application-specific issues, refer to the respective documentation:

• Dine-In Issues: See Dine-In Troubleshooting
• Take-Away Issues: See Take-Away Troubleshooting

For platform-wide issues or feature requests, submit an issue with:

1. Application name (dine-in / take-away)
2. Steps to reproduce
3. Expected vs. actual behavior
4. Logs (make logs)