🧠 Local Vision (WIP)

Status: Experimental
Tag: vision

📘 Overview

Local Vision is a simple experimental n8n workflow designed to test and confirm the use of local vision-capable LLMs (such as qwen/qwen2.5-vl-7b) for analyzing images—specifically for tasks like understanding data fields on a receipt.

This proof-of-concept confirms the ability to download an image from the web and process it using a locally-hosted multimodal model, producing structured data that could later be stored or further processed.

---

📌 Use Case

Analyze the content of an image (e.g., a store receipt) using a local Vision LLM to extract visual information in base64 format and interpret it intelligently.

---

🔁 Workflow Structure

Step	Node Name	Description
1	Manual Trigger	Used for testing the workflow manually from the n8n UI.
2	HTTP Request	Downloads an image file from the given URL (https://connect.garmin.com/images/login/desktop/signin-hero-3.jpg).
3	Qwen2.5 Local Vision	Uses a local multimodal LLM model (qwen/qwen2.5-vl-7b) to analyze the downloaded image.
4	Sticky Note (comment)	Describes the experiment, stating that the project attempts to test local vision-based LLM processing.

---

🧠 Technology Stack

• n8n for workflow automation
• Qwen 2.5 VL 7B model (via @n8n/n8n-nodes-langchain.openAi) for local vision AI processing
• HTTP Node to download the image
• Manual Trigger to start the flow

---

⚙️ Credentials Used

• LM Studio (Credential ID: nJxj1tJeQ2RgUtHC)
  Required for the OpenAI-compatible local endpoint to run the Vision LLM.

---

📝 Notes

• The image must be processed as base64 for the model to analyze it.
• Currently, the workflow stops after analysis—no further data extraction or storage is implemented. You can add a Set, Function, or Spreadsheet node to post-process or export the results.
• The workflow uses typeVersion: 4.2 for HTTP and typeVersion: 1.8 for the Langchain/OpenAI node, ensuring compatibility with recent n8n versions.

---

🚀 Next Steps

• Add a Set or Function node to parse the output into structured fields.
• Save the result to Google Sheets, PostgreSQL, or Airtable.
• Enhance the model prompt or context for more accurate receipt parsing.

---

✅ How to Run

1. Make sure your LM Studio or other OpenAI-compatible local inference server is running.
2. Replace the image URL with a picture of a receipt if needed.
3. Open n8n → Import this workflow → Click "Test Workflow".
4. Inspect the output from the Qwen2.5 Vision node.

---

📷 Example Image

Default image:

https://connect.garmin.com/images/login/desktop/signin-hero-3.jpg

Replace it with your own image if needed.

---

📁 Tags

vision, local-llm, image-processing, qwen, multimodal

---

🧩 Requirements

• n8n v1.8+ with latest HTTP & Langchain/OpenAI nodes
• Local LLM setup via LM Studio or compatible API
• Base64 image input format

---