Text and images, one vector space.

using LMKit.Model;
using LMKit.Embeddings;
using LMKit.Media.Image;

// One model, one vector space for text and images.
using LM model = LM.LoadFromModelID("embeddinggemma-2");
var embedder = new Embedder(model);

// Embed every photo once and keep the vectors.
var index = new List<(string Path, float[] Vector)>();
foreach (string path in Directory.EnumerateFiles("./photos", "*.jpg"))
{
    using var image = ImageBuffer.LoadAsRGB(path);
    index.Add((path, embedder.GetEmbeddings(image)));
}

// Query with a sentence, in any of 100+ languages.
float[] query = embedder.GetEmbeddings("forklift on a warehouse floor");

foreach (var hit in index
    .OrderByDescending(e => Embedder.GetCosineSimilarity(query, e.Vector))
    .Take(5))
{
    Console.WriteLine(hit.Path);
}

// Embed the new upload with the same model.
using var upload = ImageBuffer.LoadAsRGB("upload.jpg");
float[] probe = embedder.GetEmbeddings(upload);

// The closest vectors are the closest images.
var closest = index
    .Select(e => (e.Path, Score: Embedder.GetCosineSimilarity(probe, e.Vector)))
    .OrderByDescending(m => m.Score)
    .Take(3);

foreach (var (path, score) in closest)
    Console.WriteLine($"{score:F3}  {path}");

float[] full  = embedder.GetEmbeddings(image);   // 768 dimensions
float[] small = Normalize(full[..256]);       // 3x smaller index

// Compare truncated vectors only with vectors truncated to the same size.
static float[] Normalize(float[] v)
{
    float norm = MathF.Sqrt(v.Sum(x => x * x));
    return v.Select(x => x / norm).ToArray();
}