El Bruno

⚠️ This blog post was created with the help of AI tools. Yes, I used a bit of magic from language models to organize my thoughts and automate the boring parts, but the geeky fun and the 🤖 in C# are 100% mine.

Hi 👋

If you’ve used ElBruno.LocalEmbeddings for text embeddings, you’re going to love the new image capabilities. I asked several friends about this, and they challenge me to give it a try, so here it is:

ElBruno.LocalEmbeddings.ImageEmbeddings a library brings CLIP-based multimodal embeddings to .NET — fully local.

It is powered by ONNX Runtime, and ready for image search and image RAG workflows. In this post, I’ll show you:

• How to download the required CLIP models
• A tiny “hello image embeddings” sample in C#
• The two image samples included in the repo: ImageRagSimple and ImageRagChat

Here is the RAGChat using images as a source:

Let’s dive in! 🚀

Note: Right now, the auto-download feature as part of the library is Work-In-Progress, as these models are big. I’m working on the .NET library that do this (see roadmap), but I think so far with the download scripts we are OK.

📦 The Library: Image Embeddings (CLIP)

The image embedding library is built on top of OpenAI’s CLIP model (Contrastive Language–Image Pretraining). It uses two ONNX models:

• Text encoder → embeds natural language queries
• Vision encoder → embeds images

Both embeddings live in the same vector space, which means text-to-image and image-to-image search works with simple cosine similarity.

⬇️ Download the CLIP Models

CLIP requires four files:

• text_model.onnx
• vision_model.onnx
• vocab.json
• merges.txt

We provide scripts that download the correct files from Hugging Face.

Windows (PowerShell)

./scripts/download_clip_models.ps1

Linux / macOS (Bash)

chmod +x scripts/download_clip_models.sh
./scripts/download_clip_models.sh

These scripts download the models to:

./scripts/clip-models

✅ Basic Usage — Minimal C# Example

Here’s the simplest possible flow using the new library:

using ElBruno.LocalEmbeddings.ImageEmbeddings;
string modelDir = "./scripts/clip-models";
string imageDir = "./samples/images";
string textModelPath = Path.Combine(modelDir, "text_model.onnx");
string visionModelPath = Path.Combine(modelDir, "vision_model.onnx");
string vocabPath = Path.Combine(modelDir, "vocab.json");
string mergesPath = Path.Combine(modelDir, "merges.txt");
using var textEncoder = new ClipTextEncoder(textModelPath, vocabPath, mergesPath);
using var imageEncoder = new ClipImageEncoder(visionModelPath);
var searchEngine = new ImageSearchEngine(imageEncoder, textEncoder);
searchEngine.IndexImages(imageDir);
var results = searchEngine.SearchByText("a cat", topK: 3);
foreach (var (imagePath, score) in results)
{
    Console.WriteLine($"{Path.GetFileName(imagePath)} → {score:F4}");
}

That’s it: index images → run text query → get ranked results.

🧪 Sample 1: ImageRagSimple

ImageRagSimple is the most minimal sample. It demonstrates the core flow:

1. Load CLIP text + vision models
2. Index all images in a folder
3. Run a few hardcoded text queries

This is the best sample to read if you want to understand the library usage with minimal noise.

💬 Sample 2: ImageRagChat

ImageRagChat builds on the same engine but adds a polished CLI experience using Spectre.Console. It supports:

• Live text-to-image search
• Image-to-image search with image:<path>
• A readable, interactive UI

Commands inside the app:

• Type any text → search images
• Type image: path/to/image.jpg → image-to-image search
• Type exit → quit

🧭 Which Sample Should You Start With?

🎬 Video Walkthrough (Coming Soon)

Recorded a short video demo that walks through the library and both samples!

📚 Resources

• Image Embeddings setup guide
• ImageRagSimple sample
• ImageRagChat sample