This article provides a complete, practical walkthrough of the Splitout Code Automation Webhook n8n agent. It connects HTTP Request, Webhook across approximately 1 node(s). Expect a Intermediate setup in 15-45 minutes. One‑time purchase: €29.

What This Agent Does

This agent orchestrates a reliable automation between HTTP Request, Webhook, handling triggers, data enrichment, and delivery with guardrails for errors and rate limits.

It streamlines multi‑step processes that would otherwise require manual exports, spreadsheet cleanup, and repeated API requests. By centralizing logic in n8n, it reduces context switching, lowers error rates, and ensures consistent results across teams.

Typical outcomes include faster lead handoffs, automated notifications, accurate data synchronization, and better visibility via execution logs and optional Slack/Email alerts.

How It Works

The workflow uses standard n8n building blocks like Webhook or Schedule triggers, HTTP Request for API calls, and control nodes (IF, Merge, Set) to validate inputs, branch on conditions, and format outputs. Retries and timeouts improve resilience, while credentials keep secrets safe.

Third‑Party Integrations

• HTTP Request
• Webhook

Import and Use in n8n

1. Open n8n and create a new workflow or collection.
2. Choose Import from File or Paste JSON.
3. Paste the JSON below, then click Import.
4. Show n8n JSON

   **Title:**  
   Setting Up Medoids for Anomaly Detection Using Crops Dataset in Qdrant with n8n
   
   **Meta Description:**  
   Learn how to build a powerful medoid-based anomaly detection system for agricultural crop images using Qdrant, Voyage AI, and n8n. This workflow sets up cluster centers and thresholds using both distance matrices and multimodal embeddings.
   
   **Keywords:**  
   Qdrant, anomaly detection, medoids, n8n workflow, Voyage AI, crop classification, vector search, distance matrix, multimodal embeddings, agricultural dataset, cluster threshold, machine learning, image dataset, cosine similarity, data pipeline
   
   ---
   
   # Setting Up Medoids for Anomaly Detection Using Crops Dataset in Qdrant with n8n
   
   In the expanding world of AI and vector search, anomaly detection has become a widely used technique to identify outliers in data—especially in image-based datasets. One such use case involves agricultural crops, where visual classification and anomaly detection are vital for quality control, fraud detection, and precision farming.
   
   In this article, we break down an n8n-based workflow designed to set up medoids and threshold scores for anomaly detection in a vector database collection hosted on Qdrant. This workflow is the second stage in a three-part pipeline:  
   
   1. Upload crop image dataset to Qdrant  
   2. Set medoids and distance thresholds (this workflow)  
   3. Detect anomalies based on those thresholds  
   
   Our approach sets up cluster centers—called medoids—using two complementary methods:
   - A Distance Matrix method based on Cosine similarity
   - A Multimodal Embedding method using semantic text descriptions
   
   Let’s dive into what this workflow does and how it leverages powerful tools like Qdrant and Voyage AI.
   
   ---
   
   ## Overview of the Workflow
   
   This n8n pipeline is a structured sequence of nodes that interact with third-party APIs, execute Python code through embedded scripts, and handle high-dimensional vector data. Here's how it all comes together:
   
   ### 🔹 Step 1: Initialize Environment & Cluster Info
   The pipeline starts with setting global variables for the Qdrant instance (like URL and collection name), followed by retrieving the total number of points using Qdrant’s points/count API. This count is used to avoid information loss during the facet query that collects crop labels and counts.
   
   ### 🔹 Step 2: Understanding the Clustering Landscape
   A facet count call to Qdrant categorizes all entries under the payload key "crop_name", giving the total distribution of each crop in the collection. The workflow computes:
   - The number of unique crop types
   - The size of the largest crop cluster (maxClusterSize)
   - An array of all crop names
   
   This meta information helps control the logic for both clustering approaches.
   
   ---
   
   ## Method 1: Distance Matrix Approach (Upper Branch)
   
   This method uses Qdrant’s distance matrix API to compute pairwise Cosine distances between all points within a crop cluster.
   
   ### 🧠 Finding the Medoid
   Using Scipy's sparse matrix tools in a Python code node, the workflow constructs a distance matrix and identifies the medoid—a point that is most similar to all other points in the same cluster. This medoid is then stored back in Qdrant as a special payload field `is_medoid: true`.
   
   ### 📉 Determining Anomaly Thresholds
   To find out what constitutes an outlier:
   - The vector of the medoid is retrieved
   - Its direction is inverted (multiplied by -1) to represent the furthest possible point
   - Using this vector, a nearest-neighbor search is performed, effectively locating the most dissimilar image in that cluster
   
   The Cosine similarity score of this point is stored back in the medoid as `is_medoid_cluster_threshold`. This becomes the anomaly detection threshold for that class.
   
   ---
   
   ## Method 2: Multimodal Embedding Approach (Lower Branch)
   
   This technique leverages descriptive text to define "anchor" points in a cluster.
   
   ### 🖼️ Crop Descriptions
   A JSON block contains ChatGPT-generated natural language descriptions for each crop (e.g., “sunflower - Tall plant with large bright yellow flower heads”).
   
   ### 🌐 Embedding & Search
   These descriptions are sent to the Voyage AI’s multimodal embedding model (`voyage-multimodal-3`) for vectorization. The resulting embedding vector is used to query Qdrant for the most visually similar image based on the stored “voyage” vectors.
   
   The returned image point is tagged as the medoid using the payload key `is_text_anchor_medoid: true`.
   
   ### 📏 Setting Text-Based Thresholds
   Just like with the previous approach:
   - The vector of the text-based medoid is inverted
   - A nearest-neighbor search identifies the "most dissimilar" point
   - Its similarity score is saved as `is_text_anchor_medoid_cluster_threshold` in the medoid’s payload
   
   ---
   
   ## Why Both Approaches Matter
   
   Each approach provides a unique advantage:
   - The Distance Matrix method relies strictly on point-to-point relationships in vector space, ensuring mathematical objectivity.
   - The Multimodal Embedding method offers semantic reasoning—it finds the image best matching a human-readable description of a crop, which might capture visual cues that humans deem most important.
   
   Using both methods can help triangulate true patterns and identify anomalies more reliably when combined with real-world test data.
   
   ---
   
   ## Preparing for Anomaly Detection (Step 3)
   
   At this point, both types of medoids and their respective threshold scores are set in Qdrant. The final step (not covered in this workflow) will be an anomaly detection tool that:
   1. Uploads a new crop image
   2. Embeds it using the same vector model
   3. Measures its similarity to class centers (medoids)
   4. Flags it if it falls below the threshold
   
   ---
   
   ## Third-Party APIs Used
   
   1. **Qdrant Cloud API**  
      - CRUD operations on vector points  
      - Distance matrix API  
      - Facet & point count API  
      - Point search & query for payloads
   
      URL: https://qdrant.tech/documentation/
   
   2. **Voyage AI API**  
      - Model: `voyage-multimodal-3`  
      - Performs multimodal embedding of text descriptions into vector embeddings compatible with image vectors
   
      URL: https://voyageai.com/
   
   ---
   
   ## Final Thoughts
   
   This workflow showcases just how accessible and logical it is to set up clustering and anomaly detection pipelines using no-code/low-code tools like n8n and vector databases like Qdrant. By combining two medoid-setting techniques—statistical and semantic—you greatly increase robustness in detecting outliers in any image-based dataset.
   
   Moreover, this method is fully adaptable to other domains like medical imaging, fashion, manufacturing, or facial recognition—anywhere anomaly detection is based on visual or semantic similarity.
   
   Whether you're a machine learning engineer or a no-code enthusiast, this is a practical and scalable approach to powering intelligent detection systems.
   
   ---
   
   Want to test it out yourself? Start by uploading your dataset, set up Qdrant’s free-tier cloud instance, plug into Voyage AI, and build your own AI-powered pipeline—no need to retrain any models!

5. Set credentials for each API node (keys, OAuth) in Credentials.
6. Run a test via Execute Workflow. Inspect Run Data, then adjust parameters.
7. Enable the workflow to run on schedule, webhook, or triggers as configured.

Tips: keep secrets in credentials, add retries and timeouts on HTTP nodes, implement error notifications, and paginate large API fetches.

Validation: use IF/Code nodes to sanitize inputs and guard against empty payloads.

Why Automate This with AI Agents

AI‑assisted automations offload repetitive, error‑prone tasks to a predictable workflow. Instead of manual copy‑paste and ad‑hoc scripts, your team gets a governed pipeline with versioned state, auditability, and observable runs.

n8n’s node graph makes data flow transparent while AI‑powered enrichment (classification, extraction, summarization) boosts throughput and consistency. Teams reclaim time, reduce operational costs, and standardize best practices without sacrificing flexibility.

Compared to one‑off integrations, an AI agent is easier to extend: swap APIs, add filters, or bolt on notifications without rewriting everything. You get reliability, control, and a faster path from idea to production.

Best Practices

• Credentials: restrict scopes and rotate tokens regularly.
• Resilience: configure retries, timeouts, and backoff for API nodes.
• Data Quality: validate inputs; normalize fields early to reduce downstream branching.
• Performance: batch records and paginate for large datasets.
• Observability: add failure alerts (Email/Slack) and persistent logs for auditing.
• Security: avoid sensitive data in logs; use environment variables and n8n credentials.

FAQs

Can I swap integrations later? Yes. Replace or add nodes and re‑map fields without rebuilding the whole flow.

How do I monitor failures? Use Execution logs and add notifications on the Error Trigger path.

Does it scale? Use queues, batching, and sub‑workflows to split responsibilities and control load.

Is my data safe? Keep secrets in Credentials, restrict token scopes, and review access logs.