Inventory Restock Forecast with n8n & Vector Store

Automating inventory restock decisions helps you avoid stockouts, reduce excess stock, and keep fulfillment running smoothly. This guide walks you through an n8n workflow template called “Inventory Restock Forecast” and explains, step by step, how it works and how to deploy it.

You will learn how the workflow:

• Receives inventory data through a webhook
• Splits and embeds text into vectors
• Stores and queries vectors in a Supabase vector store
• Uses an Anthropic-powered agent to make restock recommendations
• Logs every recommendation into Google Sheets for review

Learning goals

By the end of this tutorial you should be able to:

• Understand the high-level architecture of the Inventory Restock Forecast n8n template
• Configure all required credentials and services (Cohere, Supabase, Anthropic, Google Sheets)
• Customize text splitting, embeddings, and vector queries for your inventory data
• Explain how the agent uses vector search results to produce restock recommendations
• Adapt and troubleshoot the workflow for different business contexts such as ecommerce, wholesale, or retail chains

Concept overview: What this n8n workflow does

This n8n automation combines embeddings, a vector store, and an LLM agent to generate context-aware restock suggestions. It is designed to work with unstructured or semi-structured inventory data such as product descriptions, historical sales notes, and vendor details.

Why use this workflow?

The Inventory Restock Forecast template is useful if you want to:

• Make faster and more consistent restock decisions based on historical context and product signals
• Search large unstructured inventory documents using vector similarity instead of simple keyword search
• Maintain an automated log and audit trail in Google Sheets for reporting and human review
• Have an extensible foundation where you can swap models, adjust chunk sizes, and add your own business rules

High-level architecture of the template

At a high level, the workflow moves through these stages:

1. Webhook – Receives incoming inventory updates or historical documents via HTTP POST.
2. Text Splitter – Breaks long text into smaller chunks using configurable chunkSize and chunkOverlap values.
3. Embeddings – Sends these chunks to Cohere (or another provider) to generate vector embeddings.
4. Supabase Insert – Stores the embedding vectors in a Supabase index called inventory_restock_forecast.
5. Vector Query & Tool – Performs similarity search to retrieve the most relevant context for each query.
6. Memory – Maintains a short conversation buffer so the agent can remember recent context.
7. Chat (LM) & Agent – Uses Anthropic to reason about the data, call the vector tool when needed, and generate recommendations.
8. Google Sheets – Writes the final recommendation and metadata into a sheet as a log entry.

Next, we will walk through how to deploy and configure this template in n8n, then look in more detail at how the agent makes its decisions.

Step-by-step: Deploying the Inventory Restock Forecast template

Step 1 – Import the template into n8n

1. Open your n8n instance and go to the Workflows section.
2. Use the Import option to upload the provided JSON template file.
3. After import, visually inspect the workflow:
   • Confirm all nodes are present (Webhook, Splitter, Embeddings, Supabase, Agent, Google Sheets, etc.).
   • Check that connections between nodes match the intended data flow.

Step 2 – Configure required credentials

The template relies on several external services. In n8n, open each node that uses external APIs and attach or create the correct credentials.

• Cohere (Embeddings)
  • Provide your Cohere API key in the credentials section.
  • Ensure the embeddings node is set to use these credentials.
• Supabase (Vector Store)
  • Enter your Supabase URL and service key.
  • Confirm that the pgvector extension is enabled in your Supabase database.
  • Make sure the table and index for inventory_restock_forecast exist and are configured for vector similarity queries.
• Anthropic (Agent & Chat)
  • Add your Anthropic API credentials.
  • Attach them to the Chat/Agent node that generates recommendations.
• Google Sheets (Logging)
  • Set up OAuth2 credentials for Google Sheets.
  • Create or specify a sheet (for example, a sheet named Log).
  • In the Google Sheets node, point the workflow to this sheet or adjust the node settings to match your preferred sheet name.

Step 3 – Tune text splitting and embeddings

Before sending real data, it is important to understand how text splitting and embeddings work in this template.

Text splitter configuration

The template uses the following default values:

• chunkSize = 400
• chunkOverlap = 40

These values are a good starting point for product descriptions and purchase histories. You can adjust them based on your documents:

• For longer documents, consider a smaller chunkSize or a higher chunkOverlap to capture more context around each sentence.
• For short, focused descriptions, you can often keep the defaults or even reduce overlap to speed up processing.

Choosing an embeddings provider

The template is configured to use Cohere for embeddings, which balances latency and accuracy for many inventory use cases.

You can switch to other providers if needed:

• OpenAI
• Hugging Face
• Other embeddings APIs supported by n8n

To change providers, update the Embeddings node configuration and attach the appropriate credentials.

Step 4 – Create and test the webhook

The workflow begins with a Webhook node that listens for POST requests. In the template, this is typically configured with a path such as:

/inventory_restock_forecast

Send a test POST request with a JSON payload that includes product details and recent activity. For example:

{  "product_id": "SKU-12345",  "description": "Red cotton t-shirts - monthly sales history and reorder points...",  "last_30d_sales": 150
}

After sending the request:

• Check that the Splitter node receives the payload and creates chunks.
• Verify that the Embeddings node processes each chunk without errors.
• Open your Supabase dashboard to confirm that vectors are inserted into the inventory_restock_forecast table.

Once this basic data flow is working, you are ready to explore how the agent uses these vectors to make restock decisions.

How the n8n agent generates restock recommendations

From vector search to decision

When you trigger a restock query, the workflow performs the following logical steps:

1. Vector similarity search
   • The Query node calls the Supabase vector index inventory_restock_forecast.
   • It retrieves the most relevant vectors based on the incoming product data, such as sales history, notes, or vendor lead times.
2. Context assembly
   • The retrieved chunks are combined into a context package for the agent.
   • This context may include past sales, reorder patterns, supplier performance, and other product signals.
3. Agent reasoning with Anthropic
   • The Anthropic-powered agent receives:
     • New input data (current stock, recent sales velocity, etc.)
     • Retrieved context from the vector store
     • Business rules defined in the prompt
   • Using this information, the agent calculates:
     • A recommended order quantity
     • A priority level (for example, urgent vs normal restock)
     • A textual rationale explaining the decision
4. Memory buffer
   • A short-term memory node keeps track of recent interactions.
   • This allows the agent to maintain continuity, especially if you query multiple related products in sequence.

Designing the agent prompt

The quality of recommendations depends heavily on the prompt attached to the Agent node. A clear prompt should define:

Inputs available to the agent

• Current stock level
• Sales in the last 30 days or another relevant period
• Vendor lead times and variability
• Any notes or constraints stored in the vector database

Business rules to apply

Examples of rules you might include in the prompt:

• Maintain at least 14 days of safety stock.
• Respect minimum order quantities (for example, at least 50 units per order).
• Take into account supplier lead time and known delays.

Output format for easier downstream use

To make the output machine readable and easy to log, define a structured format such as JSON. For example, instruct the agent to return:

• product_id
• recommended_qty
• priority (for example, “high”, “medium”, “low”)
• rationale explaining the reasoning in plain language

By standardizing this format, you make it easier to integrate the recommendations with other systems such as ERP or purchasing tools.

Logging recommendations in Google Sheets

After the agent produces a recommendation, the workflow sends the result to a Google Sheets node. This creates a running log of all restock decisions.

What gets logged

Each row in the sheet (for example, in a sheet named Log) can include:

• Timestamp of the recommendation
• Product ID
• Recommended order quantity
• Priority level and rationale
• A context snippet or reference to the vectors/IDs used

This log acts as an audit trail. Buyers and managers can:

• Review past recommendations
• Compare them with actual purchase decisions and outcomes
• Export the data into procurement or BI systems for analysis

Best practices for scaling and reliability

As you move from testing to production, consider the following best practices.

Indexing strategy in Supabase

• Use product-level metadata such as product_id, category, and vendor to tag your vectors.
• Filter queries by these tags when possible to narrow the search space and improve performance.

Data housekeeping

• Re-embed documents after significant changes such as updated pricing, changed lead times, or revised product descriptions.
• Remove or archive outdated vectors that no longer reflect current business reality.

Handling rate limits and throughput

• Batch incoming webhook submissions if you expect high volume.
• Use n8n concurrency and rate limit controls on the Embeddings and Agent nodes to avoid provider throttling.

Security considerations

• Protect your webhook with authentication or IP allowlists if possible.
• Use service keys for Supabase and never expose them in public workflows.
• Store API keys and secrets only in n8n credentials, not in plain text fields.

Monitoring and error handling

• Add error handling branches to catch failures in:
  • Embeddings generation
  • Supabase inserts
  • Agent responses
• Send alerts to a notification channel such as Slack or email when errors occur.

Adapting the workflow to different business types

The core template is generic, but small adjustments can make it more effective for specific industries.

Ecommerce stores

For ecommerce, consider:

• Integrating detailed order history and SKU-level unit economics.
• Using shorter chunk sizes to capture line-item details and recent customer reviews or comments.
• Emphasizing seasonality and promotions in your agent prompt.

Wholesale and B2B suppliers

For wholesale or B2B operations, it is often important to:

• Incorporate minimum order quantities and vendor-specific lead time variability.
• Store supply contracts and terms as vectors so the agent can reference them.
• Include large account commitments and contract dates in the context.

Retail chains and multi-location businesses

Retail chains may want to:

• Include location-level sales velocity and stock levels.
• Model stock transfers between locations as an alternative to new purchase orders.
• Use the memory buffer to keep track of recent interactions for each location-product combination.

Troubleshooting: Common issues and quick checks

• No vectors