Aurora Café — AI assistant

Aurora Café is an end-to-end AI customer service workflow designed for a fictional coffee shop. The project demonstrates how conversational AI, vector search (RAG), memory, and workflow automation can work together to provide a natural customer experience through WhatsApp.

Customers can ask questions, browse the menu, place takeout orders, make reservations, reschedule or cancel bookings, and seamlessly escalate conversations to a human agent whenever necessary.

Built with n8n, Redis, PGVector, Airtable, Google Calendar, Google Drive, OpenRouter, Ollama, nomic embed text and Evolution API.

1. Business overview

Aurora Café is a fictional premium coffee shop specialized in specialty coffee, artisanal brunch, signature pastries, and warm customer experiences. The business concept combines gastronomy, productivity, meetings, and wellbeing in a modern, sophisticated, and comfortable environment.

In addition to in-store service, Aurora Café provides intelligent WhatsApp support, table reservations, pickup orders, corporate event requests, casual coworking, accessibility, partner parking, customer Wi-Fi, and a pet-friendly outdoor area.

The automation turns WhatsApp into a digital concierge, named Aurora, able to assist customers with an elegant, warm, and objective tone. The assistant must not invent information and must use the official knowledge base whenever the conversation involves opening hours, menu, policies, prices, reservations, payments, accessibility, coworking, events, and other business information.

2. Project goal

Automate a coffee shop customer service operation using n8n as the main orchestration layer. The workflow receives WhatsApp messages, detects customer intent with AI, queries the RAG knowledge base when needed, stores customer memory, creates/reschedule/cancel reservations in Google Calendar, records pickup orders in Airtable, and routes customers to a human attendant whenever manual handling is required.

3. Main features

Area	Feature
WhatsApp	Receives text and voice messages through a webhook.
Conversational AI	Uses an AI Agent to interpret intent and respond naturally.
STT	Transcribes voice notes before processing.
Message debounce	Queues messages in Redis and waits 15 seconds before responding.
Bot/human mode	Controls the service state per customer in Redis.
RAG	Queries a vector knowledge base for business info, menu, FAQ, and policies.
Reservations	Creates reservations in Google Calendar based on operational rules.
Cancellation	Finds a reservation by email and deletes it from Google Calendar.
Rescheduling	Finds a reservation by email and updates its date/time in Google Calendar.
Pickup orders	Records orders in Airtable with name, phone, items, value, and status.
Customer memory	Stores and retrieves customer name/interest in Airtable.
Human handoff	Sets customer mode to human in Redis and alerts an attendant via WhatsApp.
RAG ingestion	Watches Google Drive, downloads files, generates embeddings, and stores vectors in PGVector.

4. Source files analyzed

File	Role in the project
RAG - Aurora Café.pdf	Business knowledge base: description, mission, vision, values, opening hours, menu, FAQ, and policies.
SYSTEM PROMPT — Aurora Café.pdf	Aurora assistant behavior, conversation flows, tools, limits, and governance.
Aurora Café.json	Main n8n workflow: webhook, Redis, AI, RAG, tools, response sending, and document ingestion.
HUMAN_HANDOFF.json	Subworkflow for human handoff.
CANCEL_RESERVATION.json	Subworkflow for Google Calendar reservation cancellation.
RESCHEDULE_RESERVATION.json	Subworkflow for Google Calendar reservation rescheduling.

5. Logical architecture

flowchart LR
    A[WhatsApp / Customer] --> B[Evolution API]
    B --> C[n8n Webhook]
    C --> D[Message normalization]
    D --> E{Customer message?}
    E -- No --> Z[Ignore]
    E -- Yes --> F[Redis: bot/human mode]
    F --> G{Service mode}
    G -- human --> H[Bot paused]
    G -- empty --> I[Set bot mode]
    G -- bot --> J[Classify type]
    I --> J
    J -- voice --> K[OpenAI transcription]
    J -- text --> L[Redis queue]
    K --> L
    L --> M[Wait 15s]
    M --> N[Retrieve queue]
    N --> O[AI Agent]
    O --> P[(Redis Chat Memory)]
    O --> Q[(PGVector RAG)]
    O --> R[Google Calendar]
    O --> S[Airtable]
    O --> T[Subworkflows]
    T --> U[Handoff / Cancellation / Rescheduling]
    O --> V[Send WhatsApp response]
    V --> W[Clear Redis queue]

Loading

6. Stack and integrations

Orchestration

• n8n as the main automation platform.
• AI Agent as the conversation decision center.
• Subworkflows invoked as custom agent tools.

Customer channel

• WhatsApp connected through Evolution API.
• n8n webhook for incoming messages.
• Response sending through Evolution API.

AI and language

• OpenRouter for the chat model.
• OpenAI Whisper for voice transcription.
• Redis Chat Memory for conversation memory by phone number.

RAG and knowledge

• Google Drive as the document source.
• Ollama with nomic-embed-text embeddings.
• PostgreSQL + PGVector as the vector store.

Operations and data

• Google Calendar for reservations.
• Airtable for pickup orders and customer memory.
• Redis for memory, service mode and temporary message queue.

7. n8n workflows

7.1. Main workflow — Aurora Café.json

Responsible for two main areas:

A. AI customer service

1. Receives a message through the webhook.
2. Extracts phone, message, ID, message type, origin, date, and instance.
3. Ignores messages sent by the connected account itself.
4. Reads the customer's service mode from Redis.
5. Sets mode to bot if no mode exists.
6. Stops automatic response if mode is human.
7. Classifies the message as text, voice, or unsupported.
8. Transcribes voice messages when needed.
9. Pushes the message to a Redis queue.
10. Waits 15 seconds to group consecutive messages.
11. Retrieves the accumulated queue.
12. Sends context to the AI Agent.
13. The agent may use RAG, memory, calendar, Airtable, or subworkflows.
14. Sends the answer through WhatsApp.
15. Clears the Redis queue.

B. RAG knowledge ingestion

1. Watches a Google Drive folder.
2. Finds new files.
3. Processes documents in batches.
4. Downloads each file.
5. Loads document content.
6. Generates embeddings with Ollama.
7. Stores vectors in PGVector.
8. Moves the file to a processed folder.

7.2. Subworkflow — HUMAN_HANDOFF.json

Inputs:

Field	Description
Customer	Customer phone/identifier.
Human	Human attendant phone/identifier.
Instance	Evolution API instance.
Reason	Reason for handoff.

Flow:

1. Triggered by the agent tool.
2. Sets Redis key customer:{phone}:mode to human.
3. Sends an alert to the human attendant through Evolution API with the customer contact and reason.

7.3. Subworkflow — CANCEL_RESERVATION.json

Input:

Field	Description
Email	Email used to find the reservation.

Flow:

1. Searches Google Calendar events using the customer's email.
2. Deletes the matching event.
3. Returns the result to the main flow.

7.4. Subworkflow — RESCHEDULE_RESERVATION.json

Inputs:

Field	Description
Email	Email used to find the reservation.
Date	New reservation date/time.

Flow:

1. Searches Google Calendar events using the customer's email.
2. Updates the reservation start and end time.
3. Sets the end time to new date/time + 1 hour.

8. Implemented business rules

Customer service

• The assistant is named Aurora.
• Tone must be elegant, polite, warm, objective, and professional.
• The AI must query RAG for institutional information, menu, rules, policies, FAQ, and operations.
• The AI must not answer outside the Aurora Café context.
• Severe complaints, corporate events, operational failures, and questions not found in the knowledge base must be routed to a human.

Reservations

• Required data: name, date, time, party size, and email.
• Valid hours in the operational prompt:
  • Monday to Friday: 09:00 to 19:00;
  • Saturday: 09:00 to 20:00;
  • Sunday: unavailable.
• Automatic reservations allow up to 8 people.
• Reservation date/time must be in the future.
• Minimum lead time is 1 hour.
• Late tolerance is 15 minutes.
• The automation uses a 1-hour duration in Google Calendar.

Orders

• The project does not support delivery.
• Orders are for pickup only.
• Required flow: present menu, collect items, confirm order, calculate total, and record the order.

Human support

• Must be triggered when the customer asks for a human, there is a critical complaint, operational failure, booking above the automatic limit, corporate event, financial issue, or a question not found in RAG.
• human mode pauses automatic responses for that customer.
• The current workflow does not automatically switch a customer back to bot; this must be done manually in Redis or through an additional workflow.

9. Data model

Redis

Key	Purpose	Example value
customer:{phone}:mode	Controls whether the customer is handled by bot or human.	bot or human
queue:{phone}	Stores temporary messages before the response.	Message list

Airtable — Customers

Field	Use
Phone	Customer identifier.
Name	Customer name detected or provided.
Interest	Customer's main interest.

Airtable — Orders

Field	Use
Phone	Customer phone.
Name	Full name.
Order	Order summary.
Value	Amount calculated from the menu.
Status	Initial state: In progress.

Google Calendar — Reservation

Field	Use
summary	Customer full name.
attendees	Customer email.
description	Party size / chairs.
start	Reservation start date/time.
end	Reservation end date/time.

PGVector

Table	Use
Aurora_cafe	Stores embeddings from institutional documents used by RAG.

10. How to run/test

Step 1 — Prepare services

Have available:

• running n8n instance;
• Redis;
• PostgreSQL with PGVector;
• Evolution API connected to WhatsApp;
• Google Calendar for reservations;
• Google Drive for RAG documents;
• Airtable with Customers and Orders tables;
• Ollama with nomic-embed-text:latest;
• OpenAI and OpenRouter credentials configured in n8n.

Step 2 — Import workflows

In n8n:

1. Go to Workflows.
2. Click Import from File.
3. Import the four JSON files.
4. Open each workflow and assign the correct credentials.
5. Save each workflow.

Step 3 — Adjust key nodes

In the Aurora Café workflow, review:

• inbound webhook;
• Evolution API instance;
• Redis credentials;
• OpenRouter model;
• OpenAI transcription credentials;
• Google Calendar;
• Airtable base/tables;
• PGVector table;
• Google Drive input and processed folders;
• subworkflow references.

Step 4 — Populate RAG

1. Add institutional documents to the configured Google Drive folder.
2. Run the ingestion block or wait for the trigger.
3. Confirm vectors were stored in Aurora_cafe.
4. Test questions such as:
   • “Do you have vegan options?”
   • “What are your opening hours?”
   • “Are pets allowed?”
   • “Do you have parking?”

Step 5 — Test customer service

Scenario	Expected result
Simple greeting	Aurora replies warmly.
Menu question	Aurora queries RAG and returns existing items/prices.
Delivery request	Aurora says the shop currently supports pickup only.
Valid reservation	Aurora collects data, confirms, and creates a Google Calendar event.
Reservation above 8 people	Aurora transfers to a human.
Cancellation	Aurora requests email and calls CANCEL_RESERVATION.
Rescheduling	Aurora requests email and new date and calls RESCHEDULE_RESERVATION.
Pickup order	Aurora confirms items/value and records the order in Airtable.
Human request	Workflow sets mode to human and alerts an attendant.

11. Disclaimer

Aurora Café is a fictional coffee shop created for educational use, automation demonstration, and technical portfolio purposes. The project must be reviewed and sanitized before public release or production use.