Crop image anomaly check against a Qdrant collection

Crop image anomaly check against a Qdrant collection

An n8n workflow that takes the address of one image and says whether it shows a crop your collection already knows. It
turns the picture into a vector with Voyage AI, compares that vector with one centre point per crop in a Qdrant
collection, and answers with a short text. If the picture matches no crop closely enough, the text raises an alert.

What is inside

• workflow.json: the workflow as n8n exports it, 9 nodes and 7 sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The Qdrant address was the author's own cluster. Variables for medoids now holds <YOUR_QDRANT_URL>.
• Compare scores was a Python Code node that read _input and _(). n8n's documentation says Pyodide, which provided
  those names, is removed from n8n 2.0, and that native Python there offers only _items and _item, with no imports by
  default. Mindket's copy holds the same logic written in JavaScript, which runs on every n8n version.
• The node Image URL hardcode hardcoded nothing: it reads query.imageURL from the call. It is now called
  Image URL from input, and Embed image points at the new name.
• Two sticky notes that pointed at the author's other pipelines and pages are gone, and the one beside the three counting
  nodes now says what they do. The others keep their text.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• A Qdrant collection built for this check, described next. The workflow only reads it.
• A Voyage AI API key.
• Your Qdrant address and, if the cluster asks for one, its API key.

The collection it expects

The workflow does not build the collection. The author made it in two earlier pipelines that are not in this file, so you
need your own. It must look like this:

• The collection name is agricultural-crops. Change it in Variables for medoids if yours differs.
• It has a named vector called voyage, filled with Voyage multimodal embeddings of crop pictures made with
  voyage-multimodal-3.
• Every point has a crop_name payload field with the crop's name.
• One point per crop is marked as its centre: is_medoid is true, and is_medoid_cluster_threshold holds a number.
  That number is the lowest similarity score at which a picture still counts as that crop.
• A second set of centres, named by text, may use is_text_anchor_medoid and is_text_anchor_medoid_cluster_threshold.
  To use it, put those two names in Variables for medoids.

The workflow counts the crop names in the collection, and it asks Qdrant for that many centre points.

Set it up

1. In n8n, create a workflow, open its menu, choose Import from File and pick workflow.json. Save it.
2. Add a Header Auth credential for Voyage with the name Authorization and the value Bearer <YOUR_VOYAGE_API_KEY>.
   Select it in Embed image.
3. Add a Qdrant API credential with your address and key. Select it in Total Points in Collection,
   Each Crop Counts and Get similarity of medoids.
4. In Variables for medoids, replace <YOUR_QDRANT_URL> with your cluster's address, without a slash at the end.
5. In the same node, check that the collection name and the centre field names match your collection.
6. To test it by hand, pin this item on Execute Workflow Trigger and press Test workflow:
   { "query": { "imageURL": "<YOUR_IMAGE_URL>" } }
   Use the address of a real picture. To use it as a tool, call this workflow through an agent's Workflow tool, which passes
   the address in the same query field.

How a run goes

1. Execute Workflow Trigger receives the call.
2. Image URL from input copies query.imageURL into a field called imageURL.
3. Variables for medoids sets the Qdrant address, the collection name and the two centre field names.
4. Total Points in Collection asks Qdrant for an exact count of the points. Each Crop Counts asks for the values of
   crop_name with their counts, and Info About Crop Labeled Clusters keeps how many crops there are.
5. Embed image sends the picture's address to Voyage's multimodal embeddings with the model voyage-multimodal-3,
   and gets a vector back.
6. Get similarity of medoids queries Qdrant with that vector against the voyage vector, only among the points marked
   as centres, and asks for one result per crop.
7. Compare scores checks each result against its own threshold. If at least one reaches it, the answer is
   Looks similar to <crop>, with the crop of the highest score. If none does, the answer is
   ALERT, we might have a new undefined crop!.

Change the look

• The wording of the two answers: Compare scores.
• The collection and the centre fields: Variables for medoids.
• The embedding model: Embed image. A new model makes vectors that do not fit the stored ones, so rebuild the
  collection with the same model first.
• Where it gets the address: change Image URL from input to read another field.

Good to know

• The facts below come from pages and n8n documents read on 12 November 2026.
• Voyage's reference lists voyage-multimodal-3.5 and voyage-multimodal-3 as its recommended multimodal models. The
  file keeps the older one, since the collection's vectors must come from the same model.
• Since 8 December 2025 Voyage limits the redirects it follows for an image address, requires a content-length header
  and respects robots.txt. An address that fails any of these makes the embedding call fail. An image can be at most 20
  MB and 16 million pixels.
• With a threshold you set too low, the check calls an odd picture a known crop. Set each threshold against pictures you
  know are not in the collection.
• The answer says only that a picture looks like a known crop or does not. It is a screening aid, not a diagnosis.
• Billing: Voyage bills each embedding, and Qdrant Cloud bills according to your plan.
• Your data: the picture's address goes to Voyage, and Voyage fetches the picture itself. Qdrant receives the vector.
• Credentials: the JSON holds none.