FEVER by Lowdown Labs: Private In-VPC Image Search and RAG

GETTING STARTED WITH FEVER

In ~20 minutes you'll have a private visual search engine inside your own AWS account. Point it at an S3 bucket of images, then search by text or example image and find duplicates - even across resolutions and re-encodes. Your images never leave your account.

Before you start you will need: a VPC with two private subnets in different AZs; optionally an S3 bucket of images.

1. LAUNCH

Open the pre-loaded template (switch to your Region first): https://console.aws.amazon.com/cloudformation/home#/stacks/create/review?templateURL=https://lowdown-marketplace-assets.s3.amazonaws.com/fever/1.0.0/fever-appliance.yaml&stackName=fever 

Under "Start here", set:

VpcId - Your VPC PrivateSubnetIds - Two subnets in different AZs AllowedClientCidr - Your VPC CIDR, e.g. 10.0.0.0/16 (0.0.0.0/0 is rejected) CustomerBucketArn - (optional) arn:aws:s3:::your-images-bucket - grants read access, no IAM later CreateBastion - (optional) true for one-command admin access via bastion

Everything else is prefilled. DbMode: rds (default) suits most; container is a test or trial mode; aurora for mission critical autoscaling.

Click Create stack. The first launch downloads a large image, so the first task takes a few minutes - this is normal. Wait for CREATE_COMPLETE.

2. SIGN IN

The load balancer is internal (private by design). In the Stack Outputs tab, copy the AdminTokenSecretArn value, then:

aws secretsmanager get-secret-value
--secret-id <paste AdminTokenSecretArn here>
--query SecretString --output text

With CreateBastion=true: run the BastionPortForwardCommand output locally, then open http://localhost:8080/admin 

(No bastion? Reach AdminUrl from inside the VPC.)

Sign in with any username and the token as the password.

3. PROVE IT WORKS - 5 MINUTES

In the console's Get Started guide:

1. Create an API key (shown once - copy it).

2. Upload a few images, then Search for something in them.

3. Bulk path: Upload -> Ingest an S3 bucket -> s3://your-bucket/prefix/ - watch progress on the Dashboard, then review Duplicates.

4. USE THE API

curl -X POST http://localhost:8080/v1/search 
-H "Authorization: Bearer $FEVER_KEY"
-H "Content-Type: application/json"
-d '{"text": "red car", "k": 10}'

Interactive docs at /docs. Generate a typed client in one command:

npx @openapitools/openapi-generator-cli generate
-i http://localhost:8080/openapi.json  -g python -o ./fever-client

Swap -g python for typescript-axios, go, java, and more. An MCP server ships in the box for AI agents - mint a scoped key (locked to one customer_id) and the agent can never widen its scope.

CPU OR GPU?

cpu (default) - ~51 img/s indexing, low cost. Trials and regular inference day to day, use where your quota is strained.

gpu - bulk backfills: ~830 img/s per worker node on our 10,495-image demo corpus; scales linearly with workers (up to 4 by default). Runs 2x g6e.xlarge (L40S) minimum.

Indexing is bandwidth-limited - you index, generally, as fast as you can download. On small files the limit is S3 request rate; on multi-MB originals it's network bandwidth.

STACK REMOVAL / DELETION

Simply delete the CloudFormation stack to remove everything.

FAQs

• Scoped API keys isolate downstream customers server-side.
• The appliance verifies your AWS Marketplace entitlement at startup and fails closed. No phone-home, no seller endpoint.
• FEVER stores vectors, metadata, and a reference to each image - your originals stay in your bucket.