Turn a folder of crop images in cloud storage into a searchable image index. The flow reads image files, creates embeddings, and stores them in a vector database for fast visual search and anomaly checks. Ideal for teams that manage large photo sets and need quick quality control.

The run begins by setting cluster details and checking if a collection already exists. If it does not, the flow creates a Qdrant collection with a named vector called voyage and cosine distance, then builds an index on the crop_name field for fast filtering and counting. Images are listed from Google Cloud Storage by prefix, each file is turned into a public URL, and the crop name is taken from the folder path. Tomato images are removed to support a clean anomaly test. Items are split into batches, unique IDs are generated, and the batch is sent to Voyage AI for multimodal embeddings. The results are uploaded to Qdrant in one request with vectors and payloads aligned to point IDs.

Prepare a Google Cloud Storage bucket and prefix, a Voyage AI key, and a Qdrant cluster URL and API key. Tune batch size and the embedding dimension to match the chosen model. Expect a faster setup for image search, more consistent data, and a structure ready for counts and filters by crop type.

• Manual start for safe, on demand runs during setup or testing.
• Checks if the Qdrant collection exists and creates it only when needed.
• Sets up a named vector voyage with cosine similarity to match the embedding model.
• Builds a payload index on crop_name to enable fast counts and filters.
• Lists images from Google Cloud Storage by prefix and constructs public URLs.
• Extracts the crop label from folder names and removes a chosen class for anomaly tests.
• Splits records into fixed size batches and generates UUIDs for reliable point IDs.
• Calls Voyage AI multimodal embeddings and uploads vectors and payloads to Qdrant in batches.

• Reduce manual file handling from hours to minutes by automating URL building, embedding, and upload.
• Connect Google Cloud Storage, Voyage AI, and Qdrant in one flow to avoid context switching.
• Eliminate ID collisions by generating UUIDs for every image point.
• Handle large image sets with batching to keep API calls stable and predictable.
• Speed up search and counts with an index on crop_name for quick filtering.
• Prevent duplicate collection errors by checking existence before create.

1. Import the template into n8n: Create a new workflow in n8n > Click the three dots menu > Select 'Import from File' > Choose the downloaded JSON file.
2. You'll need accounts with Google Cloud Storage, Qdrant and Voyage AI. See the Tools Required section above for links to create accounts with these services.
3. Open the Google Cloud Storage node and set the bucket name and prefix that contain your images. If credentials are not set, in the Credential dropdown click Create new credential and follow the on screen steps to connect your Google account with the correct project and permissions to list and read objects.
4. Open the node that sets Qdrant cluster variables and fill in your Qdrant Cloud URL, collection name, embedding dimension that matches your Voyage model, and a safe batch size.
5. Double click the Check Qdrant Collection Existence node. In the credential field, click Create new credential for Qdrant API, then paste your Qdrant API key and cluster URL. Save and test the connection.
6. Open the Create Qdrant Collection node and use the same Qdrant credential. Confirm the vectors section uses the same dimension as your embedding model and cosine distance.
7. Open the Payload index on crop_name node and use your Qdrant credential. This step creates an index on the crop_name payload field to improve filtering and counts.
8. Open the Embed crop image node. For authentication, choose HTTP Header Auth, click Create new credential, then set the header Authorization with value Bearer YOUR_VOYAGE_API_KEY from your Voyage AI dashboard.
9. In the Get fields for Qdrant node, confirm the expression builds a public image URL and captures the crop label from the folder path. Adjust the expression if your folder structure differs.
10. In the Filtering out tomato to test anomalies node, update the filter rule if you want to include all classes or remove a different class.
11. Review the Split in batches, generate uuids for Qdrant points code node. Adjust batch size by changing the variable you set earlier and confirm UUID generation is enabled.
12. Click Execute test. Verify logs in the Embed crop image node to confirm embeddings return with the expected dimension. Check the Batch Upload to Qdrant node for a successful upsert response.
13. In your Qdrant dashboard, open the collection, inspect a few points, and confirm vectors and payloads include crop_name and the public image URL. If you see dimension mismatch errors, update the embedding dimension in the collection variables to match the chosen model.
14. If you hit rate limits or payload too large errors, lower the batch size, then run again. For permission errors on Google Cloud Storage, grant Storage Object Viewer to the connected account.