Aerial photo classified by its nearest neighbours in Qdrant

Aerial photo classified by its nearest neighbours in Qdrant

An n8n workflow that another workflow or an AI agent calls with a picture address. It embeds the picture with Voyage AI, finds its nearest neighbours in a Qdrant collection and answers with the class most of them share.

What is inside

• workflow.json: the workflow as n8n exports it, eighteen nodes. Ten do the work and eight are sticky notes that explain the steps. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The Majority Vote step was a Python Code node that counted the class names with Counter. In n8n 2, native Python offers only _items and _item, has no _input, and blocks standard library imports such as collections unless the instance allows them (n8n's docs, read on 12 November 2026). Mindket's copy counts them in JavaScript and returns the same list of the two most common classes.
• The Qdrant cluster address belonged to the original author. It is now the placeholder <YOUR_QDRANT_CLUSTER_URL>.
• The file carried one saved test input, a picture address in the author's own bucket. It is gone, so pin a test input of your own.
• Two nodes were linked to credentials in the author's n8n. The links are gone, so select your own.
• Nothing else. The tie loop, the prompts and the sticky notes are the file's own.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• A Qdrant collection that already holds labelled pictures. Each point needs a named vector called voyage, made with the Voyage model voyage-multimodal-3 and input_type set to document, and a payload field landscape_name that holds its class.
• A Qdrant cluster address and an API key.
• A Voyage AI API key.
• Pictures that anyone can open from a public address, because Voyage downloads them.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add a Qdrant API credential with your cluster address and key. Select it in Query Qdrant.
3. Add a Header Auth credential for Voyage. Name it Authorization and give it the value Bearer <YOUR_VOYAGE_KEY> with your own key. Select it in Embed image.
4. Open Qdrant variables + embedding + KNN neigbours. Replace <YOUR_QDRANT_CLUSTER_URL> with your cluster address, with https:// in front and no slash at the end. Set collectionName to your collection. The file uses land-use.
5. Check limitKNN, the number of neighbours asked for first. The file uses 10.
6. To try it, open Execute Workflow Trigger, pin the data { "query": { "imageURL": "<A_PUBLIC_PICTURE_ADDRESS>" } } with a real address, and press Execute workflow.
7. To use it from another workflow, call it with an Execute Workflow node or a workflow tool, and send an item that has query.imageURL.

How a run goes

1. Execute Workflow Trigger receives the item, and Image Test URL copies query.imageURL into imageURL.
2. Embed image (HTTP Request 4.2) sends the address to Voyage's multimodal embeddings API.
3. Qdrant variables + embedding + KNN neigbours keeps the vector as ImageEmbedding and sets the cluster address, the collection name and limitKNN.
4. Query Qdrant asks the collection's points query for the nearest points to that vector, with the payloads, using the vector name voyage.
5. Propagate loop variables stores how many points came back and keeps the result. Majority Vote counts the landscape_name of those points and returns the two most common classes with their counts.
6. Check tie looks at the two counts. If they are equal and limitKNN is 100 or less, Increase limitKNN adds five neighbours and Query Qdrant runs again.
7. Otherwise Return class returns the most common class as class.

Change the result

• Collection and first number of neighbours: Qdrant variables + embedding + KNN neigbours.
• How far a tie may grow: the 100 in Check tie and the 5 in Increase limitKNN.
• The class field: landscape_name in Majority Vote.
• Embedding model: model in Embed image. It has to match the model that made the collection.

Good to know

• It always answers with a class, even for a picture unlike any class in the collection. The file has no score threshold.
• Filling the collection is not part of this file. Build it with the same steps as any batch upload: embed pictures in document mode, save them as a named vector voyage and put the class in landscape_name.
• A sticky note says the author measured 93.24 percent on the test part of the dataset, with no tuning. Mindket did not run that test.
• Propagate loop variables resets limitKNN to the number of points returned. On a collection of 100 points or fewer, a tie that survives every point can repeat without end.
• Voyage's docs say a picture must be a PNG, JPEG, WEBP or GIF, at most 20 MB and 16 million pixels. Its address must also answer with a content length and respect robots.txt.
• The file uses voyage-multimodal-3, which those docs still list as a recommended model.
• The picture's address goes to Voyage, and the vector goes to Qdrant. Nothing is written to the collection.
• The file holds no credential and no key.