The Architecture Behind Mia’s Automated Habit Summary

Before touching any node, Mia sketched out the high-level architecture. She wanted a clear flow from raw form data to polished weekly insights. The core pieces looked like this:

• Webhook Trigger to receive Habit Form POST payloads
• Text Splitter to chunk long notes
• OpenAI Embeddings to convert text into vectors
• Supabase Insert & Query to store and retrieve relevant context
• Window Memory & Vector Tool to provide context to a RAG agent
• Chat Model + RAG Agent to generate the weekly summary
• Append to Google Sheets to keep a log of all generated summaries
• Slack Alert to notify her if the workflow failed

With the blueprint ready, she started building, one node at a time.

Rising Action: Building the n8n Workflow Step by Step

1. Capturing the Habit Form Data with a Webhook

The first thing Mia needed was a way to get Habit Form submissions into n8n. She created a Webhook Trigger node and configured it to accept a POST request on the path:

/habit-form-weekly-summary

Her habit-tracking form now sent payloads like this directly into n8n:

{  "userId": "123",  "weekStart": "2025-08-25",  "entries": [  {"day": "Monday", "note": "Walked 30 minutes"},  {"day": "Tuesday", "note": "Skipped workout"}  ]
}

Because she cared about security and data integrity, she:

• Validated the incoming JSON payload
• Added authentication using a Bearer token or secret query key
• Restricted which systems were allowed to call the webhook

With that, every new weekly Habit Form submission would trigger the automation.

2. Preparing Text for Embeddings with a Text Splitter

Some of Mia’s clients wrote short, bullet-like notes. Others treated the form like a journal. Long reflections could easily overwhelm a single embedding and hurt retrieval.

To keep things accurate, she added a Text Splitter node. She used a character-based approach with settings like:

• chunkSize = 400
• overlap = 40

This meant that long notes would be broken into overlapping chunks, each small enough to capture local meaning. The result was better embeddings and more precise context when the RAG agent later searched for relevant information.

3. Turning Notes Into Vectors with OpenAI Embeddings

Next, Mia connected those chunks to an OpenAI Embeddings node. She chose a model like text-embedding-3-small to convert each piece of text into a fixed-length vector.

For every chunk, she stored important metadata alongside the vector:

• userId
• weekStart
• day
• Original text content

She knew this metadata would be crucial later. It would let her filter by user and week, making sure the RAG agent only looked at the right entries when generating a weekly summary.

4. Storing and Retrieving Context with Supabase Vector Storage

To persist the embeddings, Mia used Supabase as her vector store. She created a dedicated index in a vector table, for example:

indexName: habit_form_weekly_summary

Her workflow now did two important things with Supabase:

• Insert new embeddings into the vector table after each form submission
• Query Supabase later to retrieve the top-k most relevant chunks for a specific userId and weekStart

She made sure to:

• Use the same embedding model for both insertion and query
• Filter queries by metadata, such as userId and weekStart
• Limit the number of retrieved chunks to keep the prompt efficient

At this point, Mia had a searchable memory of her clients’ weekly habit entries, ready to be used by AI.

5. Supplying Context with Window Memory and a Vector Tool

The next challenge was giving the RAG agent access to both short-lived context and long-term stored data. For that, Mia combined two n8n features:

• Window Memory to preserve recent conversational or operational context
• Vector Tool to wrap the Supabase query results as a tool the agent could call

The Window Memory made sure the agent could “remember” what had already been discussed in the flow, while the Vector Tool gave it direct access to the retrieved habit entries. Together, they provided the RAG agent with everything it needed to produce a coherent, grounded weekly summary.

6. Generating the Weekly Summary with a Chat Model and RAG Agent

Now came the heart of the system. Mia created a Chat Model node using an OpenAI model such as gpt-4o or gpt-4, depending on availability. She then wrapped it in a RAG Agent configuration so it could:

• Use the Vector Tool to fetch relevant chunks
• Leverage Window Memory for short-term context
• Produce a structured, human-friendly summary

She spent time crafting the system prompt, because she knew prompt engineering would make or break the usefulness of the summaries. A version she liked looked something like this:

You are an assistant for Habit Form Weekly Summary. Using the retrieved entries, generate a concise, human-friendly weekly summary with highlights, consistency notes, and an actionable tip.

Later, she refined it using best practices:

• Defined role and output format clearly, such as:
  “Produce a 5-sentence weekly summary with 1-2 actionable tips.”
• Restricted retrieval to the same user and week using metadata filters (userId + weekStart)
• Optionally asked the agent to return source excerpts or chunk IDs for auditability
• Set a maximum token or character limit so the summary would fit cleanly into a Google Sheets cell

She also experimented with a 3-part output structure:

• Highlights
• Issues
• Recommendations

When she finally hit “Execute” on the workflow, the RAG agent read the retrieved entries and produced something like this:

Highlights: 5/7 walks completed; notable consistency on mornings. Missed Tuesday and Thursday workouts. Suggestion: schedule a 20-minute morning walk alarm and set an accountability reminder on mid-week. Consider replacing one long session with two shorter sessions if time is constrained.

Mia smiled. It was exactly the kind of feedback she used to write by hand.

7. Logging Everything in Google Sheets

She did not want these summaries to disappear into email threads. She wanted a simple, auditable log of every weekly summary, accessible at a glance.

So she added an Append to Google Sheets node. Each time the RAG agent produced a summary, n8n would append a new row to her “Habit Weekly Log” sheet with columns such as:

• userId
• weekStart
• generatedSummary
• timestamp

Google Sheets became her lightweight analytics and audit layer. She could filter by client, compare weeks, and quickly spot patterns across her entire coaching practice.

8. Staying in Control with Slack Alerts on Error

Automation is only helpful when you know it is working. To avoid silent failures, Mia created an onError path in her n8n workflow.

If anything went wrong at runtime, an Slack Alert node would send a message to her #alerts channel with:

• A short description of the error
• A reference ID or execution URL
• Enough context to find the problematic payload

This gave her peace of mind. She could let the system run in the background, knowing she would be notified if a summary failed to generate.

Behind the Scenes: Prompt Engineering and RAG Best Practices

As Mia refined her n8n workflow, she realized that small changes in prompt design or retrieval setup had a big impact on summary quality. She adopted a few core best practices:

• Clear system messages
  She always defined the agent’s role and output format. For example:
  “You are a habit coach assistant. Produce a concise weekly summary with 1 short highlights paragraph, 1 issues paragraph, and 1-2 actionable tips.”
• Strict metadata filters
  She limited retrieval to the same userId and weekStart so the agent never mixed up different clients or weeks.
• Optional source excerpts
  When auditability mattered, she had the agent return chunk IDs or brief excerpts, so she could trace which notes influenced each summary.
• Length control
  She set a max token or character limit, keeping summaries readable and avoiding oversized Google Sheets cells.

These adjustments turned the RAG agent from “pretty good” into “reliably useful” for her coaching workflow.

Security, Monitoring, and Scaling as the Client List Grows

As more clients joined, Mia started thinking like a systems architect as much as a coach. She made sure her habit form weekly summary automation was secure and scalable.

Securing the Workflow

• Protected the webhook with authentication, rate limiting, and IP restrictions where possible
• Used Supabase Row-Level Security (RLS) to ensure that vector records were isolated per user
• Relied on scoped API keys that only had access to the specific embeddings table

Monitoring and Cost Control

• Enabled n8n execution logs to inspect runs and debug issues
• Set up alerts for failed runs so she would never miss a broken summary
• Kept an eye on OpenAI usage and defined cost thresholds

Scaling Vector Storage

When she thought about future growth, she prepared strategies such as:

• Sharding vector indexes by user cohort or date range
• Archiving older embeddings to reduce index size and query latency

Her workflow was no longer just a clever shortcut. It was an automation she trusted to grow with her business.

Testing, Edge Cases, and Getting the Details Right

Before fully handing over weekly summaries to automation, Mia stress-tested the workflow. She sent a variety of payloads through the webhook:

• Very short notes and very long reflections
• Mixed-length inputs with some days empty and others verbose
• Edge cases like empty entries, unusual characters, or malicious text
• High concurrency tests simulating multiple clients submitting at once

For each run, she compared the AI-generated summary to a human-written version. When the AI missed nuances or overemphasized minor details, she adjusted:

• The prompt wording
• The number of retrieved chunks
• The structure of the summary output

After a few iterations, she reached a point where the AI’s summaries were so close to her own that she often only made small tweaks before sending them to clients.

The Turning Point: How Mia’s Week Changed

Two weeks after deploying the n8n Habit Form Weekly Summary workflow, Mia opened her calendar and noticed something unusual. She had free time on Friday afternoons.

Before, Fridays were for catching up on habit summaries she had not written yet. Now, by the time Friday rolled around, n8n had already:

• Received each client’s Habit Form payload via webhook
• Generated embeddings with OpenAI
• Stored and queried context in Supabase using vector search
• Run a RAG agent to produce a concise weekly summary
• Appended the result to her Google Sheets log
• Kept her posted via Slack if anything broke