Add CLIP embeddings and similar-image search

- add Candle + HF Hub CLIP image embedding pipeline with background embedding worker
- write image embeddings into sqlite-vec and expose similar-image lookup through a new backend command
- surface embedding progress and recovery in the UI, including retries for failed embeddings
- improve gallery/lightbox embedding UX and make similar-image actions directly accessible

Refs: #3, #4

src-tauri/Cargo.toml

@@ -35,3 +35,7 @@ log = "0.4"
 ffmpeg-sidecar = "2.5.0"
 xxhash-rust = { version = "0.8", features = ["xxh3"] }
 sysinfo = "0.38.4"
+candle-core = { version = "0.10.2", features = ["cuda"] }
+candle-nn = { version = "0.10.2", features = ["cuda"] }
+candle-transformers = { version = "0.10.2", features = ["cuda"] }
+hf-hub = { version = "0.5.0", default-features = false, features = ["ureq", "native-tls"] }

src-tauri/src/commands.rs

@@ -1,5 +1,6 @@
-use crate::db::{self, DbPool, Folder, ImageRecord};
+use crate::db::{self, DbPool, Folder, FolderJobProgress, ImageRecord};
 use crate::indexer;
+use crate::vector;
 use serde::{Deserialize, Serialize};
 use std::path::PathBuf;
 use tauri::{AppHandle, State};
@@ -32,6 +33,17 @@ pub struct UpdateImageDetailsParams {
     pub rating: Option<i64>,
 }
 
+#[derive(Deserialize)]
+pub struct FindSimilarImagesParams {
+    pub image_id: i64,
+    pub limit: Option<usize>,
+}
+
+#[derive(Deserialize)]
+pub struct RetryFailedEmbeddingsParams {
+    pub folder_id: i64,
+}
+
 #[tauri::command]
 pub async fn add_folder(
     app: AppHandle,
@@ -75,6 +87,14 @@ pub async fn get_folders(db: State<'_, DbState>) -> Result<Vec<Folder>, String>
     db::get_folders(&conn).map_err(|e| e.to_string())
 }
 
+#[tauri::command]
+pub async fn get_background_job_progress(
+    db: State<'_, DbState>,
+) -> Result<Vec<FolderJobProgress>, String> {
+    let conn = db.get().map_err(|e| e.to_string())?;
+    db::get_all_folder_job_progress(&conn).map_err(|e| e.to_string())
+}
+
 #[tauri::command]
 pub async fn remove_folder(db: State<'_, DbState>, folder_id: i64) -> Result<(), String> {
     let conn = db.get().map_err(|e| e.to_string())?;
@@ -147,3 +167,24 @@ pub async fn reindex_folder(
     indexer::index_folder(app, db.inner().clone(), folder_id, folder_path);
     Ok(())
 }
+
+#[tauri::command]
+pub async fn find_similar_images(
+    db: State<'_, DbState>,
+    params: FindSimilarImagesParams,
+) -> Result<Vec<ImageRecord>, String> {
+    let conn = db.get().map_err(|e| e.to_string())?;
+    let limit = params.limit.unwrap_or(32);
+    let image_ids = vector::find_similar_image_ids(&conn, params.image_id, limit)
+        .map_err(|e| e.to_string())?;
+    db::get_images_by_ids(&conn, &image_ids).map_err(|e| e.to_string())
+}
+
+#[tauri::command]
+pub async fn retry_failed_embeddings(
+    db: State<'_, DbState>,
+    params: RetryFailedEmbeddingsParams,
+) -> Result<usize, String> {
+    let conn = db.get().map_err(|e| e.to_string())?;
+    db::retry_failed_embedding_jobs(&conn, params.folder_id).map_err(|e| e.to_string())
+}

src-tauri/src/db.rs

@@ -63,6 +63,10 @@ pub struct ImageRecord {
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct EmbeddingJob {
     pub image_id: i64,
+    pub folder_id: i64,
+    pub path: String,
+    pub thumbnail_path: Option<String>,
+    pub media_kind: String,
     pub status: String,
     pub attempts: i64,
     pub last_error: Option<String>,
@@ -99,6 +103,9 @@ pub struct FolderJobProgress {
     pub folder_id: i64,
     pub thumbnail_pending: i64,
     pub metadata_pending: i64,
+    pub embedding_pending: i64,
+    pub embedding_ready: i64,
+    pub embedding_failed: i64,
 }
 
 pub fn create_pool(db_path: &Path) -> Result<DbPool> {
@@ -280,6 +287,56 @@ pub fn enqueue_embedding_job(conn: &Connection, image_id: i64) -> Result<()> {
     Ok(())
 }
 
+pub fn backfill_embedding_jobs(conn: &Connection) -> Result<usize> {
+    let inserted = conn.execute(
+        "INSERT INTO embedding_jobs (image_id, status, attempts, last_error, created_at, updated_at)
+         SELECT i.id, 'pending', 0, NULL, datetime('now'), datetime('now')
+         FROM images i
+         LEFT JOIN embedding_jobs j ON j.image_id = i.id
+         WHERE i.embedding_status != 'ready'
+           AND j.image_id IS NULL",
+        [],
+    )?;
+    Ok(inserted)
+}
+
+pub fn retry_failed_embedding_jobs(conn: &Connection, folder_id: i64) -> Result<usize> {
+    let updated = conn.execute(
+        "INSERT INTO embedding_jobs (image_id, status, attempts, last_error, created_at, updated_at)
+         SELECT id, 'pending', 0, NULL, datetime('now'), datetime('now')
+         FROM images
+         WHERE folder_id = ?1 AND embedding_status = 'failed'
+         ON CONFLICT(image_id) DO UPDATE SET
+             status = 'pending',
+             last_error = NULL,
+             updated_at = datetime('now')",
+        [folder_id],
+    )?;
+    conn.execute(
+        "UPDATE images
+         SET embedding_status = 'pending', embedding_error = NULL
+         WHERE folder_id = ?1 AND embedding_status = 'failed'",
+        [folder_id],
+    )?;
+    Ok(updated)
+}
+
+pub fn reset_inflight_jobs(conn: &Connection) -> Result<()> {
+    conn.execute(
+        "UPDATE thumbnail_jobs SET status = 'pending' WHERE status = 'processing'",
+        [],
+    )?;
+    conn.execute(
+        "UPDATE metadata_jobs SET status = 'pending' WHERE status = 'processing'",
+        [],
+    )?;
+    conn.execute(
+        "UPDATE embedding_jobs SET status = 'pending' WHERE status = 'processing'",
+        [],
+    )?;
+    Ok(())
+}
+
 pub fn enqueue_thumbnail_job(conn: &Connection, image_id: i64) -> Result<()> {
     conn.execute(
         "INSERT INTO thumbnail_jobs (image_id, status, attempts, last_error, created_at, updated_at)
@@ -306,40 +363,57 @@ pub fn enqueue_metadata_job(conn: &Connection, image_id: i64) -> Result<()> {
     Ok(())
 }
 
-#[allow(dead_code)]
+pub fn get_pending_embedding_jobs(conn: &Connection, limit: usize) -> Result<Vec<EmbeddingJob>> {
-pub fn get_next_embedding_job(conn: &Connection) -> Result<Option<EmbeddingJob>> {
     let mut stmt = conn.prepare(
-        "SELECT image_id, status, attempts, last_error, created_at, updated_at
+        "SELECT j.image_id, i.folder_id, i.path, i.thumbnail_path, i.media_kind,
-         FROM embedding_jobs
+                j.status, j.attempts, j.last_error, j.created_at, j.updated_at
+         FROM embedding_jobs j
+         JOIN images i ON i.id = j.image_id
          WHERE status = 'pending'
-         ORDER BY updated_at, image_id
+         ORDER BY j.updated_at, j.image_id
-         LIMIT 1",
+         LIMIT ?1",
     )?;
-
+    let rows = stmt.query_map([limit as i64], |row| {
-    let mut rows = stmt.query([])?;
+        Ok(EmbeddingJob {
-    let Some(row) = rows.next()? else {
-        return Ok(None);
-    };
-
-    Ok(Some(EmbeddingJob {
             image_id: row.get(0)?,
-        status: row.get(1)?,
+            folder_id: row.get(1)?,
-        attempts: row.get(2)?,
+            path: row.get(2)?,
-        last_error: row.get(3)?,
+            thumbnail_path: row.get(3)?,
-        created_at: row.get(4)?,
+            media_kind: row.get(4)?,
-        updated_at: row.get(5)?,
+            status: row.get(5)?,
-    }))
+            attempts: row.get(6)?,
+            last_error: row.get(7)?,
+            created_at: row.get(8)?,
+            updated_at: row.get(9)?,
+        })
+    })?;
+    Ok(rows.collect::<rusqlite::Result<Vec<_>>>()?)
 }
 
-#[allow(dead_code)]
+pub fn claim_embedding_jobs(conn: &mut Connection, limit: usize) -> Result<Vec<EmbeddingJob>> {
-pub fn mark_embedding_job_processing(conn: &Connection, image_id: i64) -> Result<()> {
+    let tx = conn.transaction()?;
-    conn.execute(
+    let candidates = get_pending_embedding_jobs(&tx, limit * 2)?;
+    let mut claimed = Vec::with_capacity(limit);
+
+    for job in candidates {
+        let updated = tx.execute(
             "UPDATE embedding_jobs
              SET status = 'processing', attempts = attempts + 1, updated_at = datetime('now')
-         WHERE image_id = ?1",
+             WHERE image_id = ?1 AND status = 'pending'",
-        [image_id],
+            [job.image_id],
         )?;
-    Ok(())
+
+        if updated == 1 {
+            claimed.push(job);
+        }
+
+        if claimed.len() >= limit {
+            break;
+        }
+    }
+
+    tx.commit()?;
+    Ok(claimed)
 }
 
 #[allow(dead_code)]
@@ -426,13 +500,54 @@ pub fn get_folder_job_progress(conn: &Connection, folder_id: i64) -> Result<Fold
         |row| row.get(0),
     )?;
 
+    let embedding_pending = conn.query_row(
+        "SELECT COUNT(*)
+         FROM embedding_jobs j
+         JOIN images i ON i.id = j.image_id
+         WHERE i.folder_id = ?1 AND j.status IN ('pending', 'processing')",
+        [folder_id],
+        |row| row.get(0),
+    )?;
+
+    let embedding_ready = conn.query_row(
+        "SELECT COUNT(*)
+         FROM images
+         WHERE folder_id = ?1 AND embedding_status = 'ready'",
+        [folder_id],
+        |row| row.get(0),
+    )?;
+
+    let embedding_failed = conn.query_row(
+        "SELECT COUNT(*)
+         FROM images
+         WHERE folder_id = ?1 AND embedding_status = 'failed'",
+        [folder_id],
+        |row| row.get(0),
+    )?;
+
     Ok(FolderJobProgress {
         folder_id,
         thumbnail_pending,
         metadata_pending,
+        embedding_pending,
+        embedding_ready,
+        embedding_failed,
     })
 }
 
+pub fn get_all_folder_job_progress(conn: &Connection) -> Result<Vec<FolderJobProgress>> {
+    let folder_ids = get_folders(conn)?
+        .into_iter()
+        .map(|folder| folder.id)
+        .collect::<Vec<_>>();
+
+    let mut progress = Vec::with_capacity(folder_ids.len());
+    for folder_id in folder_ids {
+        progress.push(get_folder_job_progress(conn, folder_id)?);
+    }
+    Ok(progress)
+}
+
 pub fn get_pending_thumbnail_jobs(conn: &Connection, limit: usize) -> Result<Vec<ThumbnailJob>> {
     let mut stmt = conn.prepare(
         "SELECT j.image_id, i.folder_id, i.path, i.media_kind
@@ -658,6 +773,14 @@ pub fn get_image_by_id(conn: &Connection, image_id: i64) -> Result<ImageRecord>
     .map_err(Into::into)
 }
 
+pub fn get_images_by_ids(conn: &Connection, image_ids: &[i64]) -> Result<Vec<ImageRecord>> {
+    let mut images = Vec::with_capacity(image_ids.len());
+    for image_id in image_ids {
+        images.push(get_image_by_id(conn, *image_id)?);
+    }
+    Ok(images)
+}
+
 pub fn get_folders(conn: &Connection) -> Result<Vec<Folder>> {
     let mut stmt =
         conn.prepare("SELECT id, path, name, image_count, indexed_at FROM folders ORDER BY name")?;

src-tauri/src/embedder.rs

@@ -0,0 +1,118 @@
+use anyhow::Result;
+use candle_core::{DType, Device, Tensor};
+use candle_nn::VarBuilder;
+use candle_transformers::models::clip::{self, ClipModel};
+use hf_hub::{api::sync::Api, Repo, RepoType};
+use std::path::{Path, PathBuf};
+
+pub struct ClipImageEmbedder {
+    model: ClipModel,
+    device: Device,
+    image_size: usize,
+}
+
+impl ClipImageEmbedder {
+    pub fn new() -> Result<Self> {
+        println!("Initializing CLIP image embedder...");
+        let api = Api::new()?;
+        let repo = api.repo(Repo::new(
+            "laion/CLIP-ViT-B-32-laion2B-s34B-b79K".to_string(),
+            RepoType::Model,
+        ));
+        println!("Resolving CLIP model weights from Hugging Face cache...");
+        let model_path = repo.get("model.safetensors")?;
+
+        let config = clip::ClipConfig::vit_base_patch32();
+        let device = resolve_device()?;
+        let vb = unsafe {
+            VarBuilder::from_mmaped_safetensors(
+                std::slice::from_ref(&model_path),
+                DType::F32,
+                &device,
+            )?
+        };
+        let model = ClipModel::new(vb, &config)?;
+        println!("CLIP image embedder ready.");
+
+        Ok(Self {
+            model,
+            device,
+            image_size: config.image_size,
+        })
+    }
+
+    pub fn embed_image(&self, path: &Path) -> Result<Vec<f32>> {
+        Ok(self.embed_images(&[path.to_path_buf()])?.remove(0))
+    }
+
+    pub fn embed_images(&self, paths: &[PathBuf]) -> Result<Vec<Vec<f32>>> {
+        let images = load_images(paths, self.image_size)?.to_device(&self.device)?;
+        let features = self.model.get_image_features(&images)?;
+        let normalized = clip::div_l2_norm(&features)?;
+
+        let mut embeddings = Vec::with_capacity(paths.len());
+        for index in 0..paths.len() {
+            embeddings.push(normalized.get(index)?.flatten_all()?.to_vec1::<f32>()?);
+        }
+        Ok(embeddings)
+    }
+}
+
+fn resolve_device() -> Result<Device> {
+    let cuda_device = Device::cuda_if_available(0)?;
+    if cuda_device.is_cuda() {
+        println!("CLIP embedder using CUDA device.");
+        return Ok(cuda_device);
+    }
+
+    #[cfg(target_os = "macos")]
+    {
+        let metal_device = Device::metal_if_available(0)?;
+        if metal_device.is_metal() {
+            println!("CLIP embedder using Metal device.");
+            return Ok(metal_device);
+        }
+    }
+
+    println!("CLIP embedder using CPU device.");
+    Ok(Device::Cpu)
+}
+
+fn load_image(path: &Path, image_size: usize) -> Result<Tensor> {
+    let image = image::ImageReader::open(path)?
+        .with_guessed_format()?
+        .decode()?;
+    let image = image.resize_to_fill(
+        image_size as u32,
+        image_size as u32,
+        image::imageops::FilterType::Triangle,
+    );
+    let image = image.to_rgb8().into_raw();
+    let tensor = Tensor::from_vec(image, (image_size, image_size, 3), &Device::Cpu)?
+        .permute((2, 0, 1))?
+        .to_dtype(DType::F32)?
+        .affine(2.0 / 255.0, -1.0)?;
+    Ok(tensor)
+}
+
+fn load_images(paths: &[PathBuf], image_size: usize) -> Result<Tensor> {
+    let mut images = Vec::with_capacity(paths.len());
+    for path in paths {
+        images.push(load_image(path, image_size)?);
+    }
+    Ok(Tensor::stack(&images, 0)?)
+}
+
+pub fn embedding_source_path(
+    path: &str,
+    thumbnail_path: Option<&str>,
+    media_kind: &str,
+) -> PathBuf {
+    if media_kind == "video" {
+        thumbnail_path
+            .map(PathBuf::from)
+            .unwrap_or_else(|| PathBuf::from(path))
+    } else {
+        PathBuf::from(path)
+    }
+}

src-tauri/src/indexer.rs

@@ -1,4 +1,5 @@
 use crate::db::{self, DbPool, FolderJobProgress, ImageRecord, IndexedMediaEntry};
+use crate::embedder::{embedding_source_path, ClipImageEmbedder};
 use crate::media::{probe_video_metadata, MediaTools};
 use crate::storage::{detect_storage_profile, RuntimeAdaptiveProfile, StorageProfile};
 use crate::thumbnail;
@@ -26,6 +27,7 @@ static ACTIVE_INDEXING_FOLDERS: OnceLock<Mutex<HashSet<i64>>> = OnceLock::new();
 static FOLDER_STORAGE_PROFILES: OnceLock<Mutex<HashMap<i64, RuntimeAdaptiveProfile>>> =
     OnceLock::new();
 static DB_WRITE_LOCK: OnceLock<Mutex<()>> = OnceLock::new();
+const EMBEDDING_BATCH_SIZE: usize = 8;
 
 #[derive(Clone, Serialize)]
 pub struct IndexProgress {
@@ -89,6 +91,20 @@ pub fn start_metadata_worker(app: AppHandle, pool: DbPool, media_tools: MediaToo
     });
 }
 
+pub fn start_embedding_worker(app: AppHandle, pool: DbPool) {
+    std::thread::spawn(move || {
+        let mut embedder: Option<ClipImageEmbedder> = None;
+        println!("Embedding worker started.");
+        loop {
+            if let Err(error) = process_embedding_batch(&app, &pool, &mut embedder) {
+                eprintln!("Embedding worker error: {}", error);
+            }
+
+            std::thread::sleep(std::time::Duration::from_millis(500));
+        }
+    });
+}
+
 fn do_index(app: AppHandle, pool: DbPool, folder_id: i64, folder_path: PathBuf) -> Result<()> {
     let existing_entries = {
         let conn = pool.get()?;
@@ -303,6 +319,8 @@ fn process_thumbnail_batch(
         return Ok(());
     }
 
+    println!("Embedding batch claimed: {} items", jobs.len());
+
     let (image_jobs, video_jobs): (Vec<_>, Vec<_>) =
         jobs.into_iter().partition(|job| job.media_kind == "image");
 
@@ -463,6 +481,105 @@ fn process_metadata_batch(app: &AppHandle, pool: &DbPool, media_tools: &MediaToo
     Ok(())
 }
 
+fn process_embedding_batch(
+    app: &AppHandle,
+    pool: &DbPool,
+    embedder: &mut Option<ClipImageEmbedder>,
+) -> Result<()> {
+    let batch_started_at = Instant::now();
+    if embedder.is_none() {
+        *embedder = Some(ClipImageEmbedder::new()?);
+    }
+
+    let claim_started_at = Instant::now();
+    let jobs = with_db_write_lock(|| {
+        let mut conn = pool.get()?;
+        db::claim_embedding_jobs(&mut conn, EMBEDDING_BATCH_SIZE)
+    })?;
+    let claim_elapsed = claim_started_at.elapsed();
+
+    if jobs.is_empty() {
+        return Ok(());
+    }
+
+    println!("Embedding batch claimed: {} items", jobs.len());
+    let folder_ids = jobs.iter().map(|job| job.folder_id).collect::<HashSet<_>>();
+    emit_folder_job_progress(app, pool, &folder_ids.iter().copied().collect::<Vec<_>>());
+    let embedder = embedder.as_ref().expect("embedder should be initialized");
+
+    let infer_started_at = Instant::now();
+    let source_paths = jobs
+        .iter()
+        .map(|job| embedding_source_path(&job.path, job.thumbnail_path.as_deref(), &job.media_kind))
+        .collect::<Vec<_>>();
+
+    let results = match embedder.embed_images(&source_paths) {
+        Ok(embeddings) => jobs
+            .into_iter()
+            .zip(embeddings.into_iter().map(Ok))
+            .collect::<Vec<_>>(),
+        Err(batch_error) => {
+            eprintln!(
+                "Embedding batch fallback to per-image mode: {}",
+                batch_error
+            );
+            jobs.into_iter()
+                .zip(source_paths.into_iter())
+                .map(|(job, source_path)| (job, embedder.embed_image(&source_path)))
+                .collect::<Vec<_>>()
+        }
+    };
+    let infer_elapsed = infer_started_at.elapsed();
+
+    let write_started_at = Instant::now();
+    let updated_images = with_db_write_lock(|| {
+        let mut conn = pool.get()?;
+        let tx = conn.transaction()?;
+        let mut updated_images = Vec::new();
+
+        for (job, embedding_result) in results {
+            match embedding_result {
+                Ok(embedding) => {
+                    vector::upsert_embedding(&tx, job.image_id, &embedding)?;
+                    db::mark_embedding_ready(&tx, job.image_id, vector::CLIP_MODEL_NAME)?;
+                }
+                Err(error) => {
+                    db::mark_embedding_failed(&tx, job.image_id, &error.to_string())?;
+                }
+            }
+
+            updated_images.push(db::get_image_by_id(&tx, job.image_id)?);
+        }
+
+        tx.commit()?;
+        Ok(updated_images)
+    })?;
+
+    if !updated_images.is_empty() {
+        println!("Embedding batch completed: {} items", updated_images.len());
+        let folder_ids = updated_images
+            .iter()
+            .map(|image| image.folder_id)
+            .collect::<HashSet<_>>();
+        emit_media_updates(
+            app,
+            &MediaUpdateBatch {
+                images: updated_images,
+            },
+        );
+        emit_folder_job_progress(app, pool, &folder_ids.into_iter().collect::<Vec<_>>());
+    }
+
+    let write_elapsed = write_started_at.elapsed();
+    let batch_elapsed = batch_started_at.elapsed();
+    println!(
+        "Embedding batch timing: claimed {} in {:?}, infer {:?}, write {:?}, total {:?}",
+        EMBEDDING_BATCH_SIZE, claim_elapsed, infer_elapsed, write_elapsed, batch_elapsed
+    );
+
+    Ok(())
+}
+
 fn active_indexing_folders() -> HashSet<i64> {
     ACTIVE_INDEXING_FOLDERS
         .get_or_init(|| Mutex::new(HashSet::new()))

src-tauri/src/lib.rs

@@ -1,5 +1,6 @@
 mod commands;
 mod db;
+mod embedder;
 mod indexer;
 mod media;
 mod storage;
@@ -32,6 +33,12 @@ pub fn run() {
             {
                 let conn = pool.get().expect("Failed to get connection for migration");
                 db::migrate(&conn).expect("Failed to run migrations");
+                db::reset_inflight_jobs(&conn).expect("Failed to reset inflight jobs");
+                let backfilled = db::backfill_embedding_jobs(&conn)
+                    .expect("Failed to backfill embedding jobs");
+                if backfilled > 0 {
+                    println!("Backfilled {} embedding jobs.", backfilled);
+                }
             }
 
             let thumb_dir = app_dir.join("thumbnails");
@@ -50,6 +57,7 @@ pub fn run() {
                 );
             }
             indexer::start_metadata_worker(app.handle().clone(), pool.clone(), media_tools.clone());
+            indexer::start_embedding_worker(app.handle().clone(), pool.clone());
 
             app.manage(pool);
             app.manage(media_tools);
@@ -59,10 +67,13 @@ pub fn run() {
         .invoke_handler(tauri::generate_handler![
             commands::add_folder,
             commands::get_folders,
+            commands::get_background_job_progress,
             commands::remove_folder,
             commands::get_images,
             commands::reindex_folder,
             commands::update_image_details,
+            commands::find_similar_images,
+            commands::retry_failed_embeddings,
         ])
         .run(tauri::generate_context!())
         .expect("error while running tauri application");

src-tauri/src/vector.rs

@@ -50,6 +50,34 @@ pub fn upsert_embedding(conn: &Connection, image_id: i64, embedding: &[f32]) ->
     Ok(())
 }
 
+pub fn find_similar_image_ids(conn: &Connection, image_id: i64, limit: usize) -> Result<Vec<i64>> {
+    let embedding: Vec<u8> = conn.query_row(
+        "SELECT embedding FROM image_vec WHERE image_id = ?1",
+        [image_id],
+        |row| row.get(0),
+    )?;
+
+    let mut stmt = conn.prepare(
+        "SELECT image_id
+         FROM image_vec
+         WHERE embedding MATCH vec_f32(?1)
+           AND k = ?2",
+    )?;
+    let rows = stmt.query_map((&embedding, (limit as i64) + 1), |row| row.get::<_, i64>(0))?;
+
+    let mut ids = Vec::new();
+    for row in rows {
+        let candidate_id = row?;
+        if candidate_id != image_id {
+            ids.push(candidate_id);
+        }
+        if ids.len() >= limit {
+            break;
+        }
+    }
+    Ok(ids)
+}
+
 #[allow(dead_code)]
 fn pack_f32(values: &[f32]) -> Vec<u8> {
     let mut out = Vec::with_capacity(values.len() * std::mem::size_of::<f32>());

src/App.tsx

@@ -9,11 +9,15 @@ import { Lightbox } from "./components/Lightbox";
 
 export default function App() {
   const loadFolders = useGalleryStore((state) => state.loadFolders);
+  const loadBackgroundJobProgress = useGalleryStore((state) => state.loadBackgroundJobProgress);
   const loadImages = useGalleryStore((state) => state.loadImages);
   const subscribeToProgress = useGalleryStore((state) => state.subscribeToProgress);
 
   useEffect(() => {
-    loadFolders().then(() => loadImages(true));
+    loadFolders().then(() => {
+      void loadBackgroundJobProgress();
+      return loadImages(true);
+    });
     let unlisten: (() => void) | undefined;
     subscribeToProgress().then((fn) => {
       unlisten = fn;

src/components/BackgroundTasks.tsx

@@ -16,19 +16,27 @@ export function BackgroundTasks() {
   const folders = useGalleryStore((state) => state.folders);
   const indexingProgress = useGalleryStore((state) => state.indexingProgress);
   const mediaJobProgress = useGalleryStore((state) => state.mediaJobProgress);
+  const retryFailedEmbeddings = useGalleryStore((state) => state.retryFailedEmbeddings);
 
   const tasks = useMemo(() => {
     return folders
       .map((folder) => {
         const index = indexingProgress[folder.id];
         const jobs = mediaJobProgress[folder.id];
-        const pendingMediaWork = (jobs?.thumbnail_pending ?? 0) + (jobs?.metadata_pending ?? 0);
+        const pendingMediaWork =
+          (jobs?.thumbnail_pending ?? 0) +
+          (jobs?.metadata_pending ?? 0) +
+          (jobs?.embedding_pending ?? 0);
+        const embeddingProcessed = (jobs?.embedding_ready ?? 0) + (jobs?.embedding_failed ?? 0);
+        const embeddingTotal = embeddingProcessed + (jobs?.embedding_pending ?? 0);
+        const hasFailedEmbeddings = (jobs?.embedding_failed ?? 0) > 0;
 
-        if (!index && pendingMediaWork === 0) {
+        if (!index && pendingMediaWork === 0 && !hasFailedEmbeddings) {
           return null;
         }
 
         const indexPercent = index && index.total > 0 ? (index.indexed / index.total) * 100 : 0;
+        const embeddingPercent = embeddingTotal > 0 ? (embeddingProcessed / embeddingTotal) * 100 : 0;
         return {
           id: folder.id,
           name: folder.name,
@@ -36,6 +44,10 @@ export function BackgroundTasks() {
           jobs,
           pendingMediaWork,
           indexPercent,
+          embeddingProcessed,
+          embeddingTotal,
+          embeddingPercent,
+          hasFailedEmbeddings,
         };
       })
       .filter((task) => task !== null);
@@ -60,12 +72,20 @@ export function BackgroundTasks() {
                 <p className="text-[11px] text-gray-500">
                   {task.index && !task.index.done
                     ? `${task.index.indexed.toLocaleString()} of ${task.index.total.toLocaleString()} scanned`
+                    : task.hasFailedEmbeddings && task.pendingMediaWork === 0
+                      ? `Embedding failures require attention`
                       : `${task.pendingMediaWork.toLocaleString()} media jobs remaining`}
                 </p>
               </div>
               <div className="text-right text-[11px] text-gray-400">
                 {task.jobs?.thumbnail_pending ? <div>{task.jobs.thumbnail_pending.toLocaleString()} thumbnails</div> : null}
                 {task.jobs?.metadata_pending ? <div>{task.jobs.metadata_pending.toLocaleString()} metadata</div> : null}
+                {task.embeddingTotal > 0 ? (
+                  <div>
+                    {task.embeddingProcessed.toLocaleString()} / {task.embeddingTotal.toLocaleString()} embeddings
+                  </div>
+                ) : null}
+                {task.jobs?.embedding_failed ? <div>{task.jobs.embedding_failed.toLocaleString()} failed</div> : null}
               </div>
             </div>
 
@@ -74,10 +94,32 @@ export function BackgroundTasks() {
                 <ProgressBar value={task.indexPercent} />
                 <p className="truncate text-[11px] text-gray-500">{task.index.current_file || "Scanning..."}</p>
               </div>
+            ) : task.embeddingTotal > 0 && (task.jobs?.embedding_pending ?? 0) > 0 ? (
+              <div className="mt-2 space-y-1">
+                <ProgressBar value={task.embeddingPercent} />
+                <p className="text-[11px] text-gray-500">
+                  {task.embeddingProcessed.toLocaleString()} completed, {task.jobs?.embedding_pending?.toLocaleString() ?? 0} remaining
+                </p>
+              </div>
+            ) : task.hasFailedEmbeddings ? (
+              <div className="mt-2 space-y-1">
+                <ProgressBar value={100} />
+                <div className="flex items-center justify-between gap-3">
+                  <p className="text-[11px] text-amber-300">
+                    {task.jobs?.embedding_failed?.toLocaleString() ?? 0} embedding failures need attention
+                  </p>
+                  <button
+                    className="rounded-full border border-amber-400/20 bg-amber-500/10 px-2.5 py-1 text-[11px] text-amber-200 hover:bg-amber-500/20"
+                    onClick={() => void retryFailedEmbeddings(task.id)}
+                  >
+                    Retry
+                  </button>
+                </div>
+              </div>
             ) : task.pendingMediaWork > 0 ? (
               <div className="mt-2 space-y-1">
                 <ProgressBar value={0} />
-                <p className="text-[11px] text-gray-500">Processing thumbnails and metadata</p>
+                <p className="text-[11px] text-gray-500">Processing thumbnails, metadata, and embeddings</p>
               </div>
             ) : null}
           </div>

src/components/Gallery.tsx

@@ -38,6 +38,19 @@ function formatDuration(durationMs: number | null): string | null {
   return `${minutes}:${seconds.toString().padStart(2, "0")}`;
 }
 
+function embeddingLabel(image: ImageRecord): string {
+  if (image.embedding_status === "ready") {
+    return image.embedding_model ? `Embeddings ready` : "Embeddings ready";
+  }
+  if (image.embedding_status === "failed") {
+    return "Embeddings failed";
+  }
+  if (image.embedding_status === "processing") {
+    return "Embedding...";
+  }
+  return "Embedding queued";
+}
+
 function ContextMenu({
   x,
   y,
@@ -49,10 +62,13 @@ function ContextMenu({
   image: ImageRecord;
   onClose: () => void;
 }) {
-  const { openImage, updateImageDetails } = useGalleryStore();
+  const openImage = useGalleryStore((state) => state.openImage);
+  const updateImageDetails = useGalleryStore((state) => state.updateImageDetails);
+  const loadSimilarImages = useGalleryStore((state) => state.loadSimilarImages);
 
   return (
     <div
+      data-gallery-context-menu
       className="fixed z-40 min-w-56 rounded-2xl border border-white/10 bg-gray-950/95 p-2 shadow-2xl backdrop-blur"
       style={{ left: x, top: y }}
       onClick={(event) => event.stopPropagation()}
@@ -75,21 +91,37 @@ function ContextMenu({
       >
         {image.favorite ? "Remove Favorite" : "Add to Favorites"}
       </button>
+      <button
+        className="w-full rounded-xl px-3 py-2 text-left text-sm text-gray-200 hover:bg-white/5"
+        onClick={async () => {
+          await loadSimilarImages(image.id);
+          onClose();
+        }}
+      >
+        Find Similar
+      </button>
       <div className="my-2 h-px bg-white/5" />
       <div className="px-3 pb-1 pt-1 text-[11px] uppercase tracking-[0.2em] text-gray-500">Rating</div>
-      <div className="flex gap-1 px-2 pb-1">
+      <div className="flex items-center gap-1 px-2 pb-1">
         {Array.from({ length: 5 }, (_, index) => {
           const rating = index + 1;
           return (
             <button
               key={rating}
-              className={`rounded-lg px-2 py-1 text-sm ${rating <= image.rating ? "bg-amber-400/15 text-amber-300" : "bg-white/5 text-gray-400 hover:text-white"}`}
+              className="rounded-md p-1"
               onClick={async () => {
                 await updateImageDetails(image.id, { rating });
                 onClose();
               }}
+              title={`Set ${rating} star rating`}
             >
-              {rating}
+              <svg
+                className={`h-5 w-5 ${rating <= image.rating ? "text-amber-300" : "text-white/20 hover:text-white/50"}`}
+                fill="currentColor"
+                viewBox="0 0 20 20"
+              >
+                <path d="M9.049 2.927c.3-.921 1.603-.921 1.902 0l1.07 3.292a1 1 0 00.95.69h3.462c.969 0 1.371 1.24.588 1.81l-2.8 2.034a1 1 0 00-.364 1.118l1.07 3.292c.3.921-.755 1.688-1.54 1.118l-2.8-2.034a1 1 0 00-1.176 0l-2.8 2.034c-.784.57-1.838-.197-1.539-1.118l1.07-3.292a1 1 0 00-.364-1.118L2.98 8.72c-.783-.57-.38-1.81.588-1.81H7.03a1 1 0 00.951-.69l1.07-3.292z" />
+              </svg>
             </button>
           );
         })}
@@ -123,6 +155,7 @@ function ImageTile({
 }) {
   const [loaded, setLoaded] = useState(false);
   const [errored, setErrored] = useState(false);
+  const loadSimilarImages = useGalleryStore((state) => state.loadSimilarImages);
 
   const src = image.thumbnail_path
     ? convertFileSrc(image.thumbnail_path)
@@ -207,7 +240,19 @@ function ImageTile({
         <p className="truncate text-sm font-medium text-white">{image.filename}</p>
         <div className="mt-1 flex items-center justify-between gap-2 text-xs text-white/70">
           <RatingStars rating={image.rating} />
-          <span>{image.embedding_status === "ready" ? "indexed" : image.embedding_status}</span>
+          <span>{embeddingLabel(image)}</span>
+        </div>
+        <div className="mt-2 flex items-center gap-2 opacity-0 transition-opacity group-hover:opacity-100">
+          <button
+            className="rounded-full border border-white/10 bg-black/40 px-2.5 py-1 text-[11px] text-white/85 hover:bg-black/60"
+            onClick={(event) => {
+              event.stopPropagation();
+              void loadSimilarImages(image.id);
+            }}
+          >
+            Find Similar
+          </button>
+          <span className="text-[11px] text-white/50">Right-click for more</span>
         </div>
       </div>
     </button>
@@ -242,18 +287,21 @@ export function Gallery() {
   }, [handleScroll]);
 
   useEffect(() => {
-    const close = () => setContextMenu(null);
+    const close = (event: PointerEvent) => {
+      if ((event.target as HTMLElement | null)?.closest("[data-gallery-context-menu]")) {
+        return;
+      }
+      setContextMenu(null);
+    };
     const handleKeyDown = (event: KeyboardEvent) => {
       if (event.key === "Escape") {
         setContextMenu(null);
       }
     };
-    window.addEventListener("click", close);
+    window.addEventListener("pointerdown", close);
-    window.addEventListener("contextmenu", close);
     window.addEventListener("keydown", handleKeyDown);
     return () => {
-      window.removeEventListener("click", close);
+      window.removeEventListener("pointerdown", close);
-      window.removeEventListener("contextmenu", close);
       window.removeEventListener("keydown", handleKeyDown);
     };
   }, []);

src/components/Lightbox.tsx

@@ -32,11 +32,25 @@ function formatDuration(durationMs: number | null): string {
   return `${minutes}:${seconds.toString().padStart(2, "0")}`;
 }
 
+function embeddingLabel(status: string, model: string | null): string {
+  if (status === "ready") {
+    return model ? `Ready (${model})` : "Ready";
+  }
+  if (status === "failed") {
+    return "Failed";
+  }
+  if (status === "processing") {
+    return "Processing";
+  }
+  return "Queued";
+}
+
 export function Lightbox() {
   const selectedImage = useGalleryStore((state) => state.selectedImage);
   const closeImage = useGalleryStore((state) => state.closeImage);
   const images = useGalleryStore((state) => state.images);
   const openImage = useGalleryStore((state) => state.openImage);
+  const loadSimilarImages = useGalleryStore((state) => state.loadSimilarImages);
   const updateImageDetails = useGalleryStore((state) => state.updateImageDetails);
   const [zoom, setZoom] = useState(1);
   const imageViewportRef = useRef<HTMLDivElement>(null);
@@ -184,6 +198,12 @@ export function Lightbox() {
                         <path d="M3.172 5.172a4 4 0 015.656 0L10 6.343l1.172-1.171a4 4 0 115.656 5.656L10 17.657l-6.828-6.829a4 4 0 010-5.656z" />
                       </svg>
                     </button>
+                    <button
+                      className="rounded-full border border-white/10 bg-white/5 px-3 py-1.5 text-xs text-gray-300 hover:text-white"
+                      onClick={() => void loadSimilarImages(selectedImage.id)}
+                    >
+                      Similar
+                    </button>
                   </div>
                   <button className="rounded p-1 text-gray-400 hover:text-white" onClick={closeImage}>
                     <svg className="h-4 w-4" fill="none" viewBox="0 0 24 24" stroke="currentColor">
@@ -281,7 +301,10 @@ export function Lightbox() {
 
                   <div>
                     <p className="mb-1 text-xs uppercase tracking-wider text-gray-500">Embedding</p>
-                    <p className="text-white">{selectedImage.embedding_status}</p>
+                    <p className="text-white">{embeddingLabel(selectedImage.embedding_status, selectedImage.embedding_model)}</p>
+                    {selectedImage.embedding_error ? (
+                      <p className="mt-1 text-xs text-amber-300">{selectedImage.embedding_error}</p>
+                    ) : null}
                   </div>
 
                   <div>

src/components/Toolbar.tsx

@@ -41,6 +41,7 @@ export function Toolbar() {
   const totalImages = useGalleryStore((state) => state.totalImages);
   const loadedCount = useGalleryStore((state) => state.loadedCount);
   const selectedFolderId = useGalleryStore((state) => state.selectedFolderId);
+  const collectionTitle = useGalleryStore((state) => state.collectionTitle);
   const folders = useGalleryStore((state) => state.folders);
   const mediaFilter = useGalleryStore((state) => state.mediaFilter);
   const setMediaFilter = useGalleryStore((state) => state.setMediaFilter);
@@ -53,7 +54,7 @@ export function Toolbar() {
   const debounceRef = useRef<ReturnType<typeof setTimeout> | null>(null);
 
   const selectedFolder = folders.find((folder) => folder.id === selectedFolderId);
-  const title = selectedFolder ? selectedFolder.name : "All Media";
+  const title = collectionTitle ?? (selectedFolder ? selectedFolder.name : "All Media");
   const tileSize = tileSizeForZoom(zoomPreset);
 
   useEffect(() => {

src/store.ts

@@ -53,6 +53,9 @@ export interface FolderJobProgress {
   folder_id: number;
   thumbnail_pending: number;
   metadata_pending: number;
+  embedding_pending: number;
+  embedding_ready: number;
+  embedding_failed: number;
 }
 
 export interface MediaJobProgressEvent {
@@ -89,11 +92,13 @@ interface GalleryState {
   favoritesOnly: boolean;
   zoomPreset: ZoomPreset;
   selectedImage: ImageRecord | null;
+  collectionTitle: string | null;
   indexingProgress: Record<number, IndexProgress>;
   mediaJobProgress: Record<number, FolderJobProgress>;
   cacheDir: string;
 
   loadFolders: () => Promise<void>;
+  loadBackgroundJobProgress: () => Promise<void>;
   addFolder: (path: string) => Promise<void>;
   removeFolder: (folderId: number) => Promise<void>;
   reindexFolder: (folderId: number) => Promise<void>;
@@ -107,6 +112,8 @@ interface GalleryState {
   setZoomPreset: (zoomPreset: ZoomPreset) => void;
   openImage: (image: ImageRecord) => void;
   closeImage: () => void;
+  loadSimilarImages: (imageId: number) => Promise<void>;
+  retryFailedEmbeddings: (folderId: number) => Promise<void>;
   updateImageDetails: (imageId: number, updates: { favorite?: boolean; rating?: number }) => Promise<void>;
   setCacheDir: (dir: string) => void;
   subscribeToProgress: () => Promise<UnlistenFn>;
@@ -235,6 +242,7 @@ export const useGalleryStore = create<GalleryState>((set, get) => ({
   favoritesOnly: false,
   zoomPreset: "comfortable",
   selectedImage: null,
+  collectionTitle: null,
   indexingProgress: {},
   mediaJobProgress: {},
   cacheDir: "",
@@ -246,16 +254,25 @@ export const useGalleryStore = create<GalleryState>((set, get) => ({
     set({ folders });
   },
 
+  loadBackgroundJobProgress: async () => {
+    const progress = await invoke<FolderJobProgress[]>("get_background_job_progress");
+    set(() => ({
+      mediaJobProgress: Object.fromEntries(progress.map((entry) => [entry.folder_id, entry])),
+    }));
+  },
+
   addFolder: async (path) => {
-    const { loadFolders } = get();
+    const { loadFolders, loadBackgroundJobProgress } = get();
     await invoke("add_folder", { path });
     await loadFolders();
+    await loadBackgroundJobProgress();
   },
 
   removeFolder: async (folderId) => {
     await invoke("remove_folder", { folderId });
-    const { selectedFolderId, loadFolders, loadImages } = get();
+    const { selectedFolderId, loadFolders, loadImages, loadBackgroundJobProgress } = get();
     await loadFolders();
+    await loadBackgroundJobProgress();
     if (selectedFolderId === folderId) {
       set({ selectedFolderId: null });
       await loadImages(true);
@@ -263,13 +280,14 @@ export const useGalleryStore = create<GalleryState>((set, get) => ({
   },
 
   reindexFolder: async (folderId) => {
-    const { loadFolders } = get();
+    const { loadFolders, loadBackgroundJobProgress } = get();
     await invoke("reindex_folder", { folderId });
     await loadFolders();
+    await loadBackgroundJobProgress();
   },
 
   selectFolder: (folderId) => {
-    set({ selectedFolderId: folderId, images: [], loadedCount: 0 });
+    set({ selectedFolderId: folderId, images: [], loadedCount: 0, collectionTitle: null });
     void get().loadImages(true);
   },
 
@@ -301,6 +319,7 @@ export const useGalleryStore = create<GalleryState>((set, get) => ({
         totalImages: result.total,
         loadedCount: reset ? result.images.length : state.loadedCount + result.images.length,
         loadingImages: false,
+        collectionTitle: reset ? null : state.collectionTitle,
       }));
     } catch (error) {
       console.error("Failed to load media:", error);
@@ -315,22 +334,22 @@ export const useGalleryStore = create<GalleryState>((set, get) => ({
   },
 
   setSearch: (search) => {
-    set({ search, images: [], loadedCount: 0 });
+    set({ search, images: [], loadedCount: 0, collectionTitle: null });
     void get().loadImages(true);
   },
 
   setSort: (sort) => {
-    set({ sort, images: [], loadedCount: 0 });
+    set({ sort, images: [], loadedCount: 0, collectionTitle: null });
     void get().loadImages(true);
   },
 
   setMediaFilter: (mediaFilter) => {
-    set({ mediaFilter, images: [], loadedCount: 0 });
+    set({ mediaFilter, images: [], loadedCount: 0, collectionTitle: null });
     void get().loadImages(true);
   },
 
   setFavoritesOnly: (favoritesOnly) => {
-    set({ favoritesOnly, images: [], loadedCount: 0 });
+    set({ favoritesOnly, images: [], loadedCount: 0, collectionTitle: null });
     void get().loadImages(true);
   },
 
@@ -339,6 +358,25 @@ export const useGalleryStore = create<GalleryState>((set, get) => ({
   openImage: (image) => set({ selectedImage: image }),
   closeImage: () => set({ selectedImage: null }),
 
+  loadSimilarImages: async (imageId) => {
+    const images = await invoke<ImageRecord[]>("find_similar_images", {
+      params: { image_id: imageId, limit: PAGE_SIZE },
+    });
+    set({
+      images,
+      totalImages: images.length,
+      loadedCount: images.length,
+      loadingImages: false,
+      collectionTitle: "Similar Images",
+      selectedFolderId: null,
+    });
+  },
+
+  retryFailedEmbeddings: async (folderId) => {
+    await invoke("retry_failed_embeddings", { params: { folder_id: folderId } });
+    await get().loadBackgroundJobProgress();
+  },
+
   updateImageDetails: async (imageId, updates) => {
     const updatedImage = await invoke<ImageRecord>("update_image_details", {
       params: {
@@ -366,6 +404,7 @@ export const useGalleryStore = create<GalleryState>((set, get) => ({
 
       if (progress.done) {
         void get().loadFolders();
+        void get().loadBackgroundJobProgress();
         void get().loadImages(true);
 
         setTimeout(() => {