Add tag cloud feature with k-means clustering and CUDA support

Introduces an Explore view with a tag cloud that clusters image embeddings
using cosine k-means and labels clusters via vocabulary-nearest-neighbour
CLIP matching. Vocabulary embeddings are disk-cached (FNV hash-keyed) to
avoid redundant inference. Enables CUDA for candle dependencies and adds a
build.rs check that surfaces a clear error when the toolkit is missing.

src-tauri/src/commands.rs

@@ -1,10 +1,10 @@
 use crate::db::{self, DbPool, Folder, FolderJobProgress, ImageRecord};
-use crate::embedder::ClipImageEmbedder;
+use crate::embedder;
 use crate::indexer;
 use crate::vector;
 use serde::{Deserialize, Serialize};
 use std::path::PathBuf;
-use tauri::{AppHandle, State};
+use tauri::{AppHandle, Manager, State};
 
 pub type DbState = DbPool;
 
@@ -204,8 +204,7 @@ pub async fn semantic_search_images(
     db: State<'_, DbState>,
     params: SemanticSearchParams,
 ) -> Result<Vec<ImageRecord>, String> {
-    let embedder = ClipImageEmbedder::new().map_err(|e| e.to_string())?;
-    let embedding = embedder.embed_text(&params.query).map_err(|e| e.to_string())?;
+    let embedding = embedder::embed_text_query(&params.query).map_err(|e| e.to_string())?;
 
     let conn = db.get().map_err(|e| e.to_string())?;
     let limit = params.limit.unwrap_or(64);
@@ -225,6 +224,160 @@ pub async fn semantic_search_images(
     Ok(images)
 }
 
+#[derive(Serialize)]
+pub struct TagCloudEntry {
+    pub label: String,
+    pub count: usize,
+}
+
+/// Clusters the library's image embeddings with k-means, then labels each cluster by
+/// finding the closest word in the vocabulary.  The vocabulary is loaded from
+/// `{app_data_dir}/vocabulary.txt` if present, otherwise from the bundled default.
+/// Vocabulary embeddings are cached to disk — only recomputed when the vocabulary changes.
+#[tauri::command]
+pub async fn get_tag_cloud(
+    app: AppHandle,
+    db: State<'_, DbState>,
+    folder_id: Option<i64>,
+) -> Result<Vec<TagCloudEntry>, String> {
+    let app_data_dir = app.path().app_data_dir().map_err(|e| e.to_string())?;
+
+    let image_embeddings = {
+        let conn = db.get().map_err(|e| e.to_string())?;
+        vector::get_all_image_embeddings(&conn, folder_id).map_err(|e| e.to_string())?
+    };
+
+    let n = image_embeddings.len();
+    if n < 5 {
+        return Ok(vec![]);
+    }
+
+    // Load vocabulary (custom file > bundled default)
+    let vocab = embedder::load_vocabulary(&app_data_dir);
+
+    // Embed vocabulary — disk-cached, only recomputes when vocabulary changes
+    let vocab_refs: Vec<&str> = vocab.iter().map(|s| s.as_str()).collect();
+    let vocab_embeddings = embedder::embed_vocab_cached(&vocab, &app_data_dir)
+        .map_err(|e| e.to_string())?;
+
+    // Choose k proportional to library size, capped at 30
+    let k = (n / 20).clamp(5, 30);
+
+    // Cluster image embeddings
+    let (centroids, cluster_counts) = kmeans_cosine(&image_embeddings, k, 40);
+
+    // Label each cluster with the nearest vocabulary word
+    let mut entries: Vec<TagCloudEntry> = Vec::new();
+    let mut used_labels = std::collections::HashSet::new();
+
+    let mut order: Vec<usize> = (0..k).collect();
+    order.sort_unstable_by(|&a, &b| cluster_counts[b].cmp(&cluster_counts[a]));
+
+    for ci in order {
+        let count = cluster_counts[ci];
+        if count == 0 { continue; }
+
+        let centroid = &centroids[ci];
+        let label = vocab_refs
+            .iter()
+            .zip(vocab_embeddings.iter())
+            .map(|(&word, emb)| {
+                let sim: f32 = centroid.iter().zip(emb.iter()).map(|(a, b)| a * b).sum();
+                (word, sim)
+            })
+            .max_by(|a, b| a.1.partial_cmp(&b.1).unwrap_or(std::cmp::Ordering::Equal))
+            .map(|(word, _)| word.to_string())
+            .unwrap_or_else(|| "other".to_string());
+
+        if used_labels.insert(label.clone()) {
+            entries.push(TagCloudEntry { label, count });
+        }
+    }
+
+    entries.sort_unstable_by(|a, b| b.count.cmp(&a.count));
+    Ok(entries)
+}
+
+// ── k-means with cosine similarity (all vectors assumed to be unit-normalized) ──
+
+fn dot(a: &[f32], b: &[f32]) -> f32 {
+    a.iter().zip(b.iter()).map(|(x, y)| x * y).sum()
+}
+
+fn normalize(v: &mut Vec<f32>) {
+    let norm: f32 = v.iter().map(|x| x * x).sum::<f32>().sqrt();
+    if norm > 1e-10 {
+        v.iter_mut().for_each(|x| *x /= norm);
+    }
+}
+
+fn kmeans_cosine(
+    points: &[Vec<f32>],
+    k: usize,
+    max_iter: usize,
+) -> (Vec<Vec<f32>>, Vec<usize>) {
+    let n = points.len();
+    let dim = points[0].len();
+
+    // Deterministic k-means++ init: spread centroids as far apart as possible
+    let mut centroids: Vec<Vec<f32>> = Vec::with_capacity(k);
+    centroids.push(points[n / 2].clone());
+    for _ in 1..k {
+        let next = points
+            .iter()
+            .map(|p| {
+                let best_sim = centroids.iter().map(|c| dot(p, c)).fold(f32::NEG_INFINITY, f32::max);
+                1.0 - best_sim // distance = 1 - cosine_similarity
+            })
+            .enumerate()
+            .max_by(|(_, a), (_, b)| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal))
+            .map(|(i, _)| i)
+            .unwrap_or(0);
+        centroids.push(points[next].clone());
+    }
+
+    let mut assignments = vec![0usize; n];
+
+    for _ in 0..max_iter {
+        // Assignment step
+        let mut changed = false;
+        for (i, p) in points.iter().enumerate() {
+            let best = centroids
+                .iter()
+                .enumerate()
+                .map(|(j, c)| (j, dot(p, c)))
+                .max_by(|a, b| a.1.partial_cmp(&b.1).unwrap_or(std::cmp::Ordering::Equal))
+                .map(|(j, _)| j)
+                .unwrap_or(0);
+            if assignments[i] != best {
+                assignments[i] = best;
+                changed = true;
+            }
+        }
+        if !changed { break; }
+
+        // Update step: mean of assigned points, then normalize
+        let mut sums = vec![vec![0.0f32; dim]; k];
+        let mut counts = vec![0usize; k];
+        for (p, &c) in points.iter().zip(assignments.iter()) {
+            sums[c].iter_mut().zip(p.iter()).for_each(|(s, v)| *s += v);
+            counts[c] += 1;
+        }
+        for (centroid, (sum, &count)) in centroids.iter_mut().zip(sums.iter_mut().zip(counts.iter())) {
+            if count > 0 {
+                sum.iter_mut().for_each(|v| *v /= count as f32);
+                normalize(sum);
+                *centroid = sum.clone();
+            }
+        }
+    }
+
+    let mut counts = vec![0usize; k];
+    for &a in &assignments { counts[a] += 1; }
+
+    (centroids, counts)
+}
+
 #[derive(Serialize)]
 pub struct WorkerStates {
     pub thumbnail_paused: bool,

src-tauri/src/embedder.rs

@@ -1,11 +1,180 @@
-use anyhow::Result;
+use anyhow::{Context, 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::io::{Read, Write};
 use std::path::{Path, PathBuf};
+use std::sync::{Mutex, OnceLock};
 use tokenizers::Tokenizer;
 
+static TEXT_SEARCH_EMBEDDER: OnceLock<Mutex<Option<ClipImageEmbedder>>> = OnceLock::new();
+/// In-process cache so the disk file is only read/written once per session.
+static VOCAB_EMBED_CACHE: OnceLock<Mutex<Option<Vec<Vec<f32>>>>> = OnceLock::new();
+
+/// Default vocabulary bundled with the binary.
+pub const DEFAULT_VOCABULARY: &str = include_str!("../data/vocabulary.txt");
+
+/// Embed a text query using a lazily-initialized, cached CLIP embedder.
+pub fn embed_text_query(query: &str) -> Result<Vec<f32>> {
+    with_text_embedder(|e| e.embed_text(query))
+}
+
+/// Parse vocabulary text: strip comment lines (starting with `#`) and blank lines.
+pub fn parse_vocabulary(text: &str) -> Vec<String> {
+    text.lines()
+        .map(|l| l.trim())
+        .filter(|l| !l.is_empty() && !l.starts_with('#'))
+        .map(|l| l.to_string())
+        .collect()
+}
+
+/// Load the vocabulary from `{cache_dir}/vocabulary.txt` if it exists,
+/// otherwise fall back to the binary-bundled default.
+pub fn load_vocabulary(cache_dir: &Path) -> Vec<String> {
+    let custom_path = cache_dir.join("vocabulary.txt");
+    if custom_path.exists() {
+        if let Ok(text) = std::fs::read_to_string(&custom_path) {
+            let words = parse_vocabulary(&text);
+            if !words.is_empty() {
+                println!("Using custom vocabulary from {:?} ({} words)", custom_path, words.len());
+                return words;
+            }
+        }
+    }
+    parse_vocabulary(DEFAULT_VOCABULARY)
+}
+
+/// Embed the vocabulary, using a disk cache so CLIP is only called when the vocabulary
+/// actually changes.  Cache file: `{cache_dir}/vocab_embeddings.bin`.
+///
+/// Format: `[u64 hash][u32 n_words][u32 dim][(n_words × dim) × f32 LE]`
+pub fn embed_vocab_cached(vocab: &[String], cache_dir: &Path) -> Result<Vec<Vec<f32>>> {
+    // In-process cache hit
+    {
+        let guard = VOCAB_EMBED_CACHE
+            .get_or_init(|| Mutex::new(None))
+            .lock()
+            .map_err(|_| anyhow::anyhow!("Vocab cache lock poisoned"))?;
+        if let Some(cached) = guard.as_ref() {
+            return Ok(cached.clone());
+        }
+    }
+
+    let vocab_hash = fnv_hash(vocab);
+    let disk_path = cache_dir.join("vocab_embeddings.bin");
+
+    // Try loading from disk
+    if let Ok(embeddings) = load_disk_cache(&disk_path, vocab_hash) {
+        let mut guard = VOCAB_EMBED_CACHE
+            .get_or_init(|| Mutex::new(None))
+            .lock()
+            .map_err(|_| anyhow::anyhow!("Vocab cache lock poisoned"))?;
+        *guard = Some(embeddings.clone());
+        println!("Vocabulary embeddings loaded from disk cache ({} words).", embeddings.len());
+        return Ok(embeddings);
+    }
+
+    // Compute embeddings
+    println!("Computing vocabulary embeddings ({} words) — this is cached after the first run.", vocab.len());
+    let prompts: Vec<String> = vocab.iter().map(|w| format!("a photo of {}", w)).collect();
+    let prompt_refs: Vec<&str> = prompts.iter().map(|s| s.as_str()).collect();
+
+    let embeddings = with_text_embedder(|e| {
+        let mut all = Vec::with_capacity(vocab.len());
+        for chunk in prompt_refs.chunks(64) {
+            all.extend(e.embed_texts_batch(chunk)?);
+        }
+        Ok(all)
+    })?;
+
+    // Save to disk
+    if let Err(e) = save_disk_cache(&disk_path, vocab_hash, &embeddings) {
+        eprintln!("Warning: could not write vocab cache: {e}");
+    }
+
+    let mut guard = VOCAB_EMBED_CACHE
+        .get_or_init(|| Mutex::new(None))
+        .lock()
+        .map_err(|_| anyhow::anyhow!("Vocab cache lock poisoned"))?;
+    *guard = Some(embeddings.clone());
+    Ok(embeddings)
+}
+
+fn fnv_hash(words: &[String]) -> u64 {
+    let mut h: u64 = 0xcbf29ce484222325;
+    for w in words {
+        for b in w.bytes() {
+            h = h.wrapping_mul(0x100000001b3) ^ (b as u64);
+        }
+        h = h.wrapping_mul(0x100000001b3) ^ 0x0A; // newline separator
+    }
+    h
+}
+
+fn load_disk_cache(path: &Path, expected_hash: u64) -> Result<Vec<Vec<f32>>> {
+    let mut f = std::fs::File::open(path).context("no cache file")?;
+    let mut buf = Vec::new();
+    f.read_to_end(&mut buf)?;
+
+    if buf.len() < 16 {
+        anyhow::bail!("cache too short");
+    }
+
+    let hash = u64::from_le_bytes(buf[0..8].try_into()?);
+    if hash != expected_hash {
+        anyhow::bail!("vocab hash mismatch — recomputing");
+    }
+    let n = u32::from_le_bytes(buf[8..12].try_into()?) as usize;
+    let dim = u32::from_le_bytes(buf[12..16].try_into()?) as usize;
+
+    let expected_len = 16 + n * dim * 4;
+    if buf.len() != expected_len {
+        anyhow::bail!("cache size mismatch");
+    }
+
+    let mut embeddings = Vec::with_capacity(n);
+    let data = &buf[16..];
+    for i in 0..n {
+        let start = i * dim * 4;
+        let emb: Vec<f32> = data[start..start + dim * 4]
+            .chunks_exact(4)
+            .map(|b| f32::from_le_bytes([b[0], b[1], b[2], b[3]]))
+            .collect();
+        embeddings.push(emb);
+    }
+    Ok(embeddings)
+}
+
+fn save_disk_cache(path: &Path, hash: u64, embeddings: &[Vec<f32>]) -> Result<()> {
+    let n = embeddings.len();
+    let dim = embeddings.first().map(|e| e.len()).unwrap_or(0);
+
+    let mut buf = Vec::with_capacity(16 + n * dim * 4);
+    buf.extend_from_slice(&hash.to_le_bytes());
+    buf.extend_from_slice(&(n as u32).to_le_bytes());
+    buf.extend_from_slice(&(dim as u32).to_le_bytes());
+    for emb in embeddings {
+        for &v in emb {
+            buf.extend_from_slice(&v.to_le_bytes());
+        }
+    }
+
+    std::fs::write(path, buf)?;
+    println!("Vocabulary embeddings cached to disk ({n} words, {dim} dims).");
+    Ok(())
+}
+
+fn with_text_embedder<T>(f: impl FnOnce(&ClipImageEmbedder) -> Result<T>) -> Result<T> {
+    let lock = TEXT_SEARCH_EMBEDDER.get_or_init(|| Mutex::new(None));
+    let mut guard = lock.lock().map_err(|_| anyhow::anyhow!("Text embedder lock poisoned"))?;
+    if guard.is_none() {
+        println!("Initializing CLIP text embedder...");
+        *guard = Some(ClipImageEmbedder::new()?);
+    }
+    f(guard.as_ref().unwrap())
+}
+
 pub struct ClipImageEmbedder {
     model: ClipModel,
     tokenizer: Tokenizer,
@@ -67,36 +236,61 @@ impl ClipImageEmbedder {
     }
 
     pub fn embed_text(&self, query: &str) -> Result<Vec<f32>> {
-        let encoding = self
-            .tokenizer
-            .encode(query, true)
-            .map_err(anyhow::Error::msg)?;
-        let token_ids = encoding
-            .get_ids()
+        Ok(self.embed_texts_batch(&[query])?.remove(0))
+    }
+
+    /// Embed multiple text queries in a single CLIP forward pass.
+    /// All sequences are padded to the same length (max 77 tokens).
+    pub fn embed_texts_batch(&self, queries: &[&str]) -> Result<Vec<Vec<f32>>> {
+        if queries.is_empty() {
+            return Ok(vec![]);
+        }
+
+        let encodings: Vec<_> = queries
             .iter()
-            .map(|token| *token as u32)
-            .collect::<Vec<_>>();
-        let input_ids = Tensor::new(vec![token_ids], &self.device)?;
+            .map(|q| self.tokenizer.encode(*q, true).map_err(anyhow::Error::msg))
+            .collect::<Result<_>>()?;
+
+        let max_len = encodings
+            .iter()
+            .map(|e| e.get_ids().len())
+            .max()
+            .unwrap_or(10)
+            .min(77);
+
+        let n = queries.len();
+        let mut flat = vec![0u32; n * max_len];
+        for (i, enc) in encodings.iter().enumerate() {
+            let ids = enc.get_ids();
+            let len = ids.len().min(max_len);
+            for j in 0..len {
+                flat[i * max_len + j] = ids[j] as u32;
+            }
+        }
+
+        let input_ids = Tensor::new(flat, &self.device)?.reshape((n, max_len))?;
         let features = self.model.get_text_features(&input_ids)?;
         let normalized = clip::div_l2_norm(&features)?;
-        Ok(normalized.flatten_all()?.to_vec1::<f32>()?)
+
+        (0..n)
+            .map(|i| Ok(normalized.get(i)?.flatten_all()?.to_vec1::<f32>()?))
+            .collect()
     }
 }
 
 fn resolve_device() -> Result<Device> {
-    #[cfg(feature = "candle-cuda")]
-    {
-        let cuda_device = Device::cuda_if_available(0)?;
-        if cuda_device.is_cuda() {
-            println!("CLIP embedder using CUDA device.");
-            return Ok(cuda_device);
+    match Device::cuda_if_available(0) {
+        Ok(device) if device.is_cuda() => {
+            println!("CLIP embedder: using CUDA GPU (device 0).");
+            return Ok(device);
+        }
+        Ok(_) => {
+            println!("CLIP embedder: no compatible CUDA GPU found — using CPU.");
+        }
+        Err(e) => {
+            println!("CLIP embedder: CUDA init failed ({e}) — using CPU.");
         }
     }
-
-    #[cfg(not(feature = "candle-cuda"))]
-    println!("CLIP embedder built without CUDA support.");
-
-    println!("CLIP embedder using CPU device.");
     Ok(Device::Cpu)
 }
 

src-tauri/src/lib.rs

@@ -77,6 +77,7 @@ pub fn run() {
             commands::semantic_search_images,
             commands::set_worker_paused,
             commands::get_worker_states,
+            commands::get_tag_cloud,
         ])
         .run(tauri::generate_context!())
         .expect("error while running tauri application");

src-tauri/src/vector.rs

@@ -78,6 +78,41 @@ pub fn find_similar_image_ids(conn: &Connection, image_id: i64, limit: usize) ->
     Ok(ids)
 }
 
+/// Returns all stored image embeddings, optionally filtered to one folder.
+/// Each embedding is returned as a normalized f32 vector.
+pub fn get_all_image_embeddings(conn: &Connection, folder_id: Option<i64>) -> Result<Vec<Vec<f32>>> {
+    let packed_rows: Vec<Vec<u8>> = match folder_id {
+        Some(fid) => {
+            let mut stmt = conn.prepare(
+                "SELECT embedding FROM image_vec
+                 WHERE image_id IN (SELECT id FROM images WHERE folder_id = ?1)",
+            )?;
+            let rows: Vec<Vec<u8>> = stmt
+                .query_map([fid], |row| row.get::<_, Vec<u8>>(0))?
+                .filter_map(|r| r.ok())
+                .collect();
+            rows
+        }
+        None => {
+            let mut stmt = conn.prepare("SELECT embedding FROM image_vec")?;
+            let rows: Vec<Vec<u8>> = stmt
+                .query_map([], |row| row.get::<_, Vec<u8>>(0))?
+                .filter_map(|r| r.ok())
+                .collect();
+            rows
+        }
+    };
+
+    Ok(packed_rows.iter().map(|b| unpack_f32(b)).collect())
+}
+
+fn unpack_f32(bytes: &[u8]) -> Vec<f32> {
+    bytes
+        .chunks_exact(4)
+        .map(|b| f32::from_le_bytes([b[0], b[1], b[2], b[3]]))
+        .collect()
+}
+
 pub fn search_image_ids_by_embedding(
     conn: &Connection,
     embedding: &[f32],