Retail Shelf Inventory Intelligence

Solution Overview

Retailers need accurate visibility into shelf inventory. In large facilities, teams often work across areas with inconsistent connectivity and rely on manual checks to identify products that need restocking.

Learn how to convert a single shelf image into a structured inventory record. A vision-language model identifies and counts products, while MongoDB Enterprise Advanced stores the captured images, item counts, and application context.

MongoDB provides the foundation for distributing inventory state across your network. Change streams deliver real-time updates to PowerSync, which synchronizes selected fields to local SQLite databases on client devices. When connectivity drops, client devices continue working offline with local data and synchronize upon reconnecting.

This MongoDB-centered workflow seamlessly integrates image-based capture, flexible document modeling, and synchronized retail applications. It transforms shelf observations into inventory data that teams can use across client applications.

Reference Architectures

MongoDB Enterprise Advanced serves as the operational data layer and source of truth for shelf inventory.

The browser sends a shelf image to the backend application. The backend passes the image to the configured vision-language model, validates the structured response, and writes the resulting inventory document to MongoDB.

Each document stores the capture context alongside detected items. MongoDB stores metadata such as the capture time, device, operator, store, model, item counts, confidence values, and lifecycle status.

When you enable media persistence, the backend stores the raw image in Amazon S3-compatible object storage and keeps the asset reference in MongoDB. This design separates large media objects from operational inventory records while preserving their relationships.

MongoDB change streams publish events that PowerSync uses to synchronize selected inventory fields to client-side SQLite databases. PowerSync supports local application access and synchronization across connectivity interruptions, while MongoDB remains the source of truth.

This architecture separates responsibilities while maintaining a unified data foundation:

• MongoDB manages the operational inventory record.

• The backend controls write operations.

• The VLM (vision-language model) extracts structured data.

• Object storage retains raw images.

• PowerSync distributes MongoDB-backed state to clients.

Figure 1. From shelf image to synchronized inventory: The backend uses a vision-language model to create a MongoDB inventory document. MongoDB stores the operational record, object storage retains the raw image, and change streams drive PowerSync synchronization to client-side SQLite.

click to enlarge

Data Model Approach

Store each shelf capture as a document in the inventory_captures collection. Keep the capture context and the model-generated inventory result together so your application processes and synchronizes each capture as a single unit.

The document records metadata including:

• Capture timestamp and location

• Submitting device and operator

• Model used for image analysis

• The capture’s lifecycle status

Embed the detected products in the items array. Each item contains a product name, detected count, and model confidence:

{
  "_id": "108d020b-9d3c-4b36-a651-0e1be2b1e899",
  "captured_at": "2026-07-02T17:45:08.661249+00:00",
  "device_id": "web-o0its21v",
  "operator_id": "demo-operator",
  "store_id": "store-unknown",
  "cv_model": "qwen2.5",
  "items": [
    {
      "name": "Nissin Cup Noodles",
      "count": 5,
      "confidence": 0.9
    },
    {
      "name": "Popcorn Bag",
      "count": 1,
      "confidence": 0.9
    },
    {
      "name": "Lays Chips",
      "count": 3,
      "confidence": 0.9
    }
  ],
  "status": "ACTIVE",
  "asset": null
}

Embedding the detected items keeps the model output close to the shelf capture that produced it. The flexible document model of MongoDB enables the application to preserve additional item attributes as the model evolves, eliminating the need for a rigid schema for every possible product.

MongoDB remains the source of truth for the inventory record. PowerSync projects the fields required by client applications into local SQLite databases, while the raw images remain in Amazon S3-compatible object storage when you enable media persistence.

Build the Solution

Use the retail-stock-take GitHub repository to deploy and explore the solution.

./scripts/setup.sh

./scripts/pull-models.sh

./scripts/verify.sh

Key Learnings

• Capture shelf inventory from an image: Use a vision-language model to convert a shelf photograph into structured JSON that MongoDB stores and your application processes.

• Model inventory data as MongoDB documents: Keep capture metadata and detected items together so each document represents a single shelf check.

• Synchronize MongoDB-backed state across client devices: Use MongoDB change streams and PowerSync to distribute selected inventory fields without building custom polling logic.

• Separate operational data from raw media: Store inventory data and image references in MongoDB, while storing the raw frame in Amazon S3-compatible object storage when you enable media persistence.

• Support disconnected retail workflows: Enable client applications to use locally synchronized data and reconcile their state upon reconnecting.

Authors

• Timothy Marland, MongoDB

• Alberto Camarena Rodriguez, MongoDB

• Daniel Jamir, MongoDB

Learn More

• MongoDB change streams documentation

• Sync MongoDB with SQLite

• PowerSync