Route and analyze customer feedback with Qwen3-VL, Tally, PostgreSQL

Description

Self-Hosted

This workflow provides a complete end-to-end system for capturing, analyzing, and routing customer feedback. By combining local multimodal AI processing with structured data storage, it allows teams to respond to customer needs in real-time without compromising data privacy.

Who is this for?

This is designed for Customer Success Managers, Product Teams, and Community Leads who need to automate the triage of high-volume feedback. It is particularly useful for organizations that handle sensitive customer data and prefer local AI processing over cloud-based API calls.

🛠️ Tech Stack

Tally.so**: For front-end feedback collection.
LM Studio**: To host the local AI models (Qwen3-VL).
PostgreSQL**: For persistent data storage and reporting.
Discord**: For real-time team notifications.

✨ How it works

Form Submission: The workflow triggers when a new submission is received from Tally.so.
Multimodal Analysis: The OpenAI node (pointing to LM Studio) processes the input using the Qwen3-VL model across three specific layers:
Sentiment Analysis: Evaluates the text to determine if the customer is Positive, Negative, or Neutral.
Zero-Shot Classification: Categorizes the feedback into pre-defined labels based on instructions in the prompt.
Vision Processing: Analyzes any attached images to extract descriptive keywords or identify UI elements mentioned in the feedback.
Data Storage: The PostgreSQL node logs the user's details, the original message, and all AI-generated insights.
AI-Driven Routing: The same Qwen3-VL model makes the routing decision by evaluating the classification results and determining the appropriate path for the data to follow.
Discord Notification: The Discord node sends a formatted message to the corresponding channel, ensuring the support team sees urgent issues while the marketing team sees positive testimonials.

📋 Requirements

LM Studio** running a local server on port 1234.
Qwen3-VL-4B** (GGUF) model loaded in LM Studio.
PostgreSQL** instance with a table configured for feedback data.
Discord Bot Token** and specific Channel IDs.

🚀 How to set up

Prepare your Local AI:
Open LM Studio and download the Qwen3-VL-4B model.
Start the Local Server on port 1234 and ensure CORS is enabled.
Disable the Require Authentication setting in the Local Server tab.
Configure PostgreSQL:
Ensure your database is running. Create a table named customer_feedback with columns for name, email_address, feedback_message, image_url, sentiment, category, and img_keywords.
Import the Workflow:
Import the JSON file into your n8n instance.
Link Services:
Update the Webhook node with your Tally.so URL.
In the Discord nodes, paste the relevant Channel IDs for your #support, #feedback, and #general channels.
Test and Activate:
Toggle the workflow to Active.
Send a test submission through your Tally form and verify the data appears in PostgreSQL and Discord.

🔑 Credential Setup

To run this workflow, you must configure the following credentials in n8n:

OpenAI API (Local)**:
Create a new OpenAI API credential.
API Key: Enter any placeholder text (e.g., lm-studio).
Base URL: Set this to your machine's local IP address (e.g., http://192.168.1.10:1234/v1) to ensure n8n can connect to the local AI server, especially if running within a Docker container.
PostgreSQL**:
Create a new PostgreSQL credential.
Enter your database Host, Database Name, User, and Password. If using the provided Docker setup, the host is usually db.
Discord Bot**:
Create a new Discord Bot API credential.
Paste your Bot Token obtained from the Discord Developer Portal.
Tally**:
Create a new Tally API credential.
Enter your API Key, which you can find in your Tally.so account settings.

⚙️ How to customize

Refine AI Logic**: Update the System Message in the AI node to change classification categories or sentiment sensitivity.
Switch to Cloud AI: If you prefer not to use a local model, you can swap the local **LM Studio connection for any 3rd party API, such as OpenAI (GPT-4o), Anthropic (Claude), or Google Gemini, by updating the node credentials and Base URL.
Expand Destinations: Add more **Discord nodes or integrate Slack to notify different departments based on the AI's routing decision.
Custom Triggers: Replace the Tally webhook with a **Typeform, Google Forms, or a custom Webhook trigger if your collection stack differs.