Insurance Claims Image Vector Search with MongoDB

Tech stack tags: mongodb mongodb-atlas mongodb-vector-search python jupyter-notebook pymongo torchvision computer-vision image-similarity insurance-claims

This repository demonstrates how to build an insurance claims image similarity workflow with MongoDB Vector Search. It embeds vehicle damage photos with a pretrained computer vision model, stores the image vectors and metadata in MongoDB, and retrieves visually similar claims images for faster triage and review.

Capabilities

• Generate image embeddings from vehicle damage photos with a pretrained TorchVision model.
• Store image binaries, filenames, and embedding vectors in MongoDB Atlas.
• Query similar insurance claim photos with MongoDB Vector Search.
• Visualize the query image and top matches directly in a Jupyter notebook.
• Use a documented MongoDB data model for agent-friendly maintenance.

Why use MongoDB for image similarity search?

• Vector storage and similarity search: MongoDB Vector Search stores image embeddings with the source image documents and retrieves visually similar vehicle damage photos with cosine similarity.
• Flexible document model: MongoDB stores image metadata, binary image payloads and model-generated vectors in a single collection without requiring a rigid relational schema.

Tech Stack

• Notebook interface: Jupyter Notebook for interactive data loading, embedding generation, querying, and visualization.
• Language: Python for data processing, model inference, and MongoDB access.
• Machine learning: TorchVision SqueezeNet for image embedding generation.
• Database: MongoDB Atlas for storing image documents and vectors.
• Search: MongoDB Vector Search for nearest-neighbor retrieval over image embeddings.
• Driver: PyMongo for connecting to MongoDB Atlas.

Architecture Overview

flowchart LR
    A["Vehicle damage image dataset"] --> B["Jupyter notebook"]
    B --> C["TorchVision SqueezeNet image embedder"]
    C --> D["1000-dimension image embedding"]
    B --> E["MongoDB Atlas: claim_resolution.car_damage_photos"]
    D --> E
    E --> F["MongoDB Vector Search index on embedding"]
    G["Query image"] --> C
    C --> H["$vectorSearch aggregation"]
    F --> H
    H --> I["Top similar claim images"]

Loading

Prerequisites

Before running this demo, install or configure:

• Python 3.10 or above
• MongoDB Atlas cluster
• Atlas database user credentials with read/write access
• Network access from your local environment to the Atlas cluster

Quick Start

1. Clone the repository

git clone https://github.com/mongodb-industry-solutions/Insurance-image-search.git
cd Insurance-image-search

2. Create a Python environment

python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.txt

3. Configure MongoDB Atlas

Create a free or dedicated MongoDB Atlas cluster, then set your connection string as an environment variable:

export MONGODB_URI="mongodb+srv://<username>:<password>@<cluster-name>/?retryWrites=true&w=majority"

4. Run the notebook

jupyter notebook image_similarity.ipynb

Run the cells in order. The notebook downloads a sample car damage image dataset, writes images to car_damage/, inserts image documents into MongoDB Atlas, builds a vector search index and queries for similar images.

Data Model

The notebook writes documents to:

Database: claim_resolution
Collection: car_damage_photos

Each document stores:

• filename: source image filename.
• data: image binary data.
• embedding: 1000-dimension image vector generated by TorchVision SqueezeNet.

See EDD.md for the full entity document diagram, field definitions, index contract, and Mermaid schema diagram.

Testing

Run the default notebook contract tests:

python -m pip install -r requirements-dev.txt
pytest -m "not integration"

Run the Atlas-backed end-to-end notebook test when you have a test cluster and matching Vector Search index:

export MONGODB_URI="mongodb+srv://<username>:<password>@<cluster-name>/?retryWrites=true&w=majority"
export NOTEBOOK_MAX_DATASET_IMAGES=5
export NOTEBOOK_CLEAR_COLLECTION=true
python -m pip install -r requirements.txt -r requirements-dev.txt
pytest -m integration

NOTEBOOK_CLEAR_COLLECTION=true clears claim_resolution.car_damage_photos before loading the test data. Use it only with a disposable test database.

Example

Example query image:

Example top-5 similar image output:

Additional Resources

• MongoDB Vector Search documentation
• PyMongo driver documentation
• TorchVision model documentation
• MongoDB aggregation documentation

License

Apache 2.0

Disclaimer

This repository is for educational use and is not a supported MongoDB product.