Versio 0.2.4

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

.gitignore

@@ -34,3 +34,6 @@ Cargo.lock
 *.pdb
 
 # End of https://www.toptal.com/developers/gitignore/api/rust,linux
+
+# Ajonaikaiset tietokannat
+*.db

network-poc/Dockerfile.native-node

@@ -5,7 +5,8 @@ RUN apt-get update && apt-get install -y \
     && rm -rf /var/lib/apt/lists/*
 
 WORKDIR /app
-COPY Cargo.toml Cargo.lock ./
+COPY Cargo.toml ./
+COPY Cargo.loc[k] ./
 COPY hub/Cargo.toml hub/Cargo.toml
 COPY node/Cargo.toml node/Cargo.toml
 COPY native-node/Cargo.toml native-node/Cargo.toml

network-poc/docker-compose.client.yml

@@ -1,13 +1,12 @@
 services:
-  # NVIDIA GPU -solmu
-  native-node-nvidia:
-    build:
-      context: .
-      dockerfile: Dockerfile.native-node
-    container_name: kipina-node-nvidia
-    environment:
-      - HUB_URL=wss://kipina.studio/ws
-      - ALLOCATED_GB=4
+  # Ollama NVIDIA GPU:lla
+  ollama-nvidia:
+    image: ollama/ollama:latest
+    container_name: kipina-ollama
+    ports:
+      - "11434:11434"
+    volumes:
+      - ollama-models:/root/.ollama
     restart: unless-stopped
     deploy:
       resources:
@@ -16,6 +15,65 @@ services:
             - driver: nvidia
               count: all
               capabilities: [gpu]
+    networks:
+      default:
+        aliases:
+          - ollama
+    profiles:
+      - nvidia
+
+  # Ollama AMD ROCm GPU:lla
+  ollama-amd:
+    image: ollama/ollama:rocm
+    container_name: kipina-ollama
+    ports:
+      - "11434:11434"
+    volumes:
+      - ollama-models:/root/.ollama
+    restart: unless-stopped
+    devices:
+      - /dev/kfd:/dev/kfd
+      - /dev/dri:/dev/dri
+    group_add:
+      - video
+      - render
+    networks:
+      default:
+        aliases:
+          - ollama
+    profiles:
+      - amd
+
+  # Ollama CPU:lla
+  ollama-cpu:
+    image: ollama/ollama:latest
+    container_name: kipina-ollama
+    ports:
+      - "11434:11434"
+    volumes:
+      - ollama-models:/root/.ollama
+    restart: unless-stopped
+    networks:
+      default:
+        aliases:
+          - ollama
+    profiles:
+      - cpu
+
+  # NVIDIA GPU -solmu
+  native-node-nvidia:
+    build:
+      context: .
+      dockerfile: Dockerfile.native-node
+    container_name: kipina-node-nvidia
+    environment:
+      - HUB_URL=wss://kipina.studio/ws
+      - OLLAMA_URL=http://ollama:11434
+      - OLLAMA_MODEL=qwen2.5-coder:7b
+      - ALLOCATED_GB=4
+    restart: unless-stopped
+    depends_on:
+      - ollama-nvidia
     profiles:
       - nvidia
 
@@ -27,14 +85,12 @@ services:
     container_name: kipina-node-amd
     environment:
       - HUB_URL=wss://kipina.studio/ws
+      - OLLAMA_URL=http://ollama:11434
+      - OLLAMA_MODEL=qwen2.5-coder:7b
       - ALLOCATED_GB=4
     restart: unless-stopped
-    devices:
-      - /dev/kfd:/dev/kfd
-      - /dev/dri:/dev/dri
-    group_add:
-      - video
-      - render
+    depends_on:
+      - ollama-amd
     profiles:
       - amd
 
@@ -46,7 +102,14 @@ services:
     container_name: kipina-node-cpu
     environment:
       - HUB_URL=wss://kipina.studio/ws
+      - OLLAMA_URL=http://ollama:11434
+      - OLLAMA_MODEL=qwen2.5-coder:7b
       - ALLOCATED_GB=2
     restart: unless-stopped
+    depends_on:
+      - ollama-cpu
     profiles:
       - cpu
+
+volumes:
+  ollama-models:

network-poc/docker-compose.yml

@@ -11,21 +11,19 @@ services:
     # Käännetään aina käynnistyksen yhteydessä varmuuden vuoksi Wasm uusimmista koodeista, ja päälle pyöräytetään Hub!
     command: bash -c "cd node && wasm-pack build --release --target web --out-dir ../static/pkg && cd ../hub && cargo run"
 
-  # Ollama — LLM-inferenssi GPU:lla (NVIDIA/AMD/Apple)
+  # Ollama — LLM-inferenssi
+  # NVIDIA: vaihda image → ollama/ollama:latest ja lisää deploy.resources (ks. README)
+  # CPU: vaihda image → ollama/ollama:latest ja poista devices
   ollama:
-    image: ollama/ollama:latest
+    image: ollama/ollama:rocm
     container_name: kipina_ollama
     ports:
       - "11434:11434"
     volumes:
       - ollama-models:/root/.ollama
-    deploy:
-      resources:
-        reservations:
-          devices:
-            - driver: nvidia
-              count: all
-              capabilities: [gpu]
+    devices:
+      - /dev/kfd
+      - /dev/dri
     profiles:
       - native
 
@@ -36,12 +34,11 @@ services:
       dockerfile: Dockerfile.native-node
     container_name: kipina_native_node
     environment:
-      - HUB_URL=ws://agentic-poc:3000/ws
+      - HUB_URL=wss://kipina.studio/ws
       - OLLAMA_URL=http://ollama:11434
       - OLLAMA_MODEL=qwen2.5-coder:7b
       - ALLOCATED_GB=4
     depends_on:
-      - agentic-poc
       - ollama
     profiles:
       - native

network-poc/hub/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "hub"
-version = "0.2.0"
+version = "0.2.4"
 edition = "2024"
 
 [dependencies]
@@ -16,3 +16,4 @@ futures = "0.3"
 rusqlite = { version = "0.31", features = ["bundled"] }
 chrono = "0.4"
 base64 = "0.22"
+reqwest = { version = "0.12", features = ["json"] }

network-poc/hub/src/main.rs

@@ -386,6 +386,7 @@ async fn main() {
         .route("/api/v1/chat/completions", axum::routing::post(api_chat_completions))
         .route("/api/v1/model", axum::routing::post(api_change_model))
         .route("/api/v1/hardware", get(api_hardware))
+        .route("/api/v1/ollama/tags", get(api_ollama_tags))
         .route("/admin", get(admin_page))
         .nest_service("/", {
             let static_dir = std::env::var("STATIC_DIR").unwrap_or_else(|_| "../static".to_string());
@@ -960,6 +961,20 @@ struct ChatCompletionResponse {
     tokens_generated: u64,
 }
 
+async fn api_ollama_tags() -> axum::response::Response {
+    let ollama_url = std::env::var("OLLAMA_URL").unwrap_or_else(|_| "http://ollama:11434".to_string());
+    match reqwest::get(format!("{}/api/tags", ollama_url)).await {
+        Ok(resp) => {
+            if let Ok(body) = resp.json::<serde_json::Value>().await {
+                axum::Json(body).into_response()
+            } else {
+                axum::Json(serde_json::json!({ "models": [] })).into_response()
+            }
+        }
+        Err(_) => axum::Json(serde_json::json!({ "models": [] })).into_response(),
+    }
+}
+
 async fn api_hardware(
     axum::extract::State(state): axum::extract::State<Arc<AppState>>,
 ) -> axum::response::Response {
@@ -969,17 +984,34 @@ async fn api_hardware(
         s.get("node_type").and_then(|v| v.as_str()) == Some("native")
     });
 
-    let (vram_mb, gpu_name, ram_mb) = if let Some(s) = native {
+    let (mut vram_mb, mut gpu_name, ram_mb) = if let Some(s) = native {
         let gpus = s.get("gpus").and_then(|v| v.as_array());
         let gpu = gpus.and_then(|g| g.first());
         let vram = gpu.and_then(|g| g.get("vram_total_mb")).and_then(|v| v.as_u64()).unwrap_or(0);
-        let name = gpu.and_then(|g| g.get("name")).and_then(|v| v.as_str()).unwrap_or("?");
+        let name = gpu.and_then(|g| g.get("name")).and_then(|v| v.as_str()).unwrap_or("").to_string();
         let ram = s.get("system").and_then(|v| v.get("ram_total_mb")).and_then(|v| v.as_u64()).unwrap_or(0);
-        (vram, name.to_string(), ram)
+        (vram, name, ram)
     } else {
-        (0, "ei natiivisolmua".to_string(), 0)
+        (0, String::new(), 0)
     };
 
+    // Fallback: kysytään Ollamalta onko malleja ladattu (= Ollama on käynnissä)
+    if vram_mb == 0 {
+        let ollama_url = std::env::var("OLLAMA_URL").unwrap_or_else(|_| "http://ollama:11434".to_string());
+        if let Ok(resp) = reqwest::get(format!("{}/api/tags", ollama_url)).await {
+            if let Ok(body) = resp.json::<serde_json::Value>().await {
+                let models = body["models"].as_array().map(|a| a.len()).unwrap_or(0);
+                if models > 0 {
+                    gpu_name = "Ollama (GPU/CPU)".to_string();
+                    // Natiivisolmun RAM fallbackina
+                    vram_mb = if ram_mb > 0 { ram_mb } else { 0 };
+                }
+            }
+        }
+    }
+
+    if gpu_name.is_empty() { gpu_name = "ei natiivisolmua".to_string(); }
+
     axum::Json(serde_json::json!({
         "gpu_name": gpu_name,
         "vram_mb": vram_mb,

network-poc/native-node/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "native-node"
-version = "0.2.0"
+version = "0.2.2"
 edition = "2024"
 
 [dependencies]

network-poc/native-node/src/inference.rs

@@ -8,17 +8,47 @@ pub struct LlmEngine {
 }
 
 impl LlmEngine {
-    pub fn load() -> Result<Self, String> {
-        let ollama_url = std::env::var("OLLAMA_URL").unwrap_or_else(|_| "http://localhost:11434".to_string());
+    pub async fn load() -> Result<Self, String> {
         let model = std::env::var("OLLAMA_MODEL").unwrap_or_else(|_| "qwen2.5-coder:7b".to_string());
 
-        tracing::info!("Ollama backend: {} | malli: {}", ollama_url, model);
-
         let client = reqwest::Client::builder()
             .timeout(std::time::Duration::from_secs(600))
+            .connect_timeout(std::time::Duration::from_secs(3))
             .build()
             .map_err(|e| format!("HTTP client: {}", e))?;
 
+        // Jos OLLAMA_URL on asetettu, käytetään sitä suoraan
+        let ollama_url = if let Ok(url) = std::env::var("OLLAMA_URL") {
+            tracing::info!("Ollama backend (env): {}", url);
+            url
+        } else {
+            // Haistellaan Ollamaa tunnetuista osoitteista
+            let candidates = [
+                "http://localhost:11434",
+                "http://127.0.0.1:11434",
+                "http://ollama:11434",
+                "http://host.docker.internal:11434",
+            ];
+            let mut found = None;
+            for url in &candidates {
+                let probe = reqwest::Client::builder()
+                    .connect_timeout(std::time::Duration::from_secs(2))
+                    .build().unwrap_or(client.clone());
+                if let Ok(resp) = probe.get(format!("{}/api/version", url)).send().await {
+                    if resp.status().is_success() {
+                        tracing::info!("Ollama löytyi osoitteesta: {}", url);
+                        found = Some(url.to_string());
+                        break;
+                    }
+                }
+            }
+            found.unwrap_or_else(|| {
+                tracing::warn!("Ollamaa ei löytynyt — käytetään oletusta http://localhost:11434");
+                "http://localhost:11434".to_string()
+            })
+        };
+
+        tracing::info!("Ollama backend: {} | malli: {}", ollama_url, model);
         Ok(LlmEngine { ollama_url, model: RefCell::new(model), client })
     }
 
@@ -49,7 +79,7 @@ impl LlmEngine {
     }
 
     pub async fn generate(&self, prompt: &str, max_tokens: usize) -> Result<GenerateResult, String> {
-        let system = "You are a coding assistant. Respond with ONLY code. No explanations, no markdown, no comments unless asked.";
+        let system = "You are a coding assistant. Respond with ONLY code. Use proper newlines and indentation. No explanations, no markdown fences, no comments unless asked.";
         let model = self.model.borrow().clone();
 
         let start = Instant::now();

network-poc/native-node/src/main.rs

@@ -287,7 +287,7 @@ async fn main() {
 
     // Ollama-backend
     tracing::info!("Alustetaan Ollama-yhteyttä...");
-    let llm = match inference::LlmEngine::load() {
+    let llm = match inference::LlmEngine::load().await {
         Ok(engine) => {
             // Varmistetaan malli (ollama pull) — odotetaan kunnes valmis
             match engine.ensure_model().await {

network-poc/node/src/lib.rs

@@ -118,6 +118,27 @@ async fn run_ai_tensor_inference(difficulty: usize) -> String {
     format!("PoC {} Matmul ({}x{}) >> {}", backend_name, active_workload_size, active_workload_size, result)
 }
 
+/// JS-exportti: tokenisoi tekstin ja palauttaa JSON-merkkijonon
+/// Tokenizer ladataan IndexedDB:stä (täytyy olla ladattu aiemmin)
+#[wasm_bindgen]
+pub async fn tokenize_js(text: String) -> Result<String, JsValue> {
+    let cached_tok = storage::load_from_idb("tokenizer.json").await.unwrap_or(None);
+    let Some(bytes) = cached_tok else {
+        // Yritetään ladata verkosta
+        let resp = reqwest::get("https://huggingface.co/Qwen/Qwen2.5-Coder-0.5B/resolve/main/tokenizer.json").await
+            .map_err(|e| JsValue::from_str(&format!("Tokenizer-lataus epäonnistui: {}", e)))?;
+        let bytes = resp.bytes().await
+            .map_err(|e| JsValue::from_str(&format!("Tokenizer-lataus epäonnistui: {}", e)))?;
+        let _ = storage::save_to_idb("tokenizer.json", &bytes).await;
+        let tokenizer = tokenizers::Tokenizer::from_bytes(&bytes)
+            .map_err(|e| JsValue::from_str(&format!("Tokenizer-parsinta: {}", e)))?;
+        return Ok(tokenize_text(&tokenizer, &text).to_string());
+    };
+    let tokenizer = tokenizers::Tokenizer::from_bytes(&bytes)
+        .map_err(|e| JsValue::from_str(&format!("Tokenizer-parsinta: {}", e)))?;
+    Ok(tokenize_text(&tokenizer, &text).to_string())
+}
+
 /// Tokenisoi yhden tekstin ja palauttaa metriikat
 fn tokenize_text(tokenizer: &tokenizers::Tokenizer, text: &str) -> serde_json::Value {
     let char_count = text.chars().count();

network-poc/static/GUIDE.md

@@ -36,29 +36,13 @@ sanan osa, kokonainen sana tai välilyönti. Tokenisaatio tehdään
 BPE-algoritmilla (Byte Pair Encoding) joka oppii yleisimmät
 merkkijonot harjoitusdatasta.
 
-### Esimerkki: koodi
-
-```
-"print('Hello')" → [print] [(' ] [Hello] [')]  = 4 tokenia
-"tulosta('Hei')" → [tul] [osta] [(' ] [He] [i] [')]  = 6 tokenia
-```
-
-Koodi tokenisoidaan tehokkaasti koska `print`, `def`, `return` yms.
-ovat kokonaisia tokeneita. Suomenkielinen `tulosta` joudutaan pilkkomaan
-osiin koska se ei esiinny harjoitusdatassa kokonaisena.
-
 ### Esimerkki: suomi vs. englanti
 
-Sama lause kahdella kielellä Qwen2.5-Coder -tokenisaattorilla:
+Alla oikea tokenisointitulos Qwen2.5-Coder-tokenisaattorilla. Jokainen
+värikoodattu lohko on yksi tokeni — huomaa miten suomi vaatii enemmän
+tokeneita saman merkityksen välittämiseen:
 
-| | Teksti | Tokenit | Määrä | Merkkejä/token |
-|---|---|---|---|---|
-| EN | The cat sat on the mat | [The] [ cat] [ sat] [ on] [ the] [ mat] | **6** | 3.7 |
-| FI | Kissa istui matolla | [K] [issa] [ ist] [ui] [ mat] [olla] | **6** | 3.2 |
-| EN | Distributed computing in the browser | [Dist] [ributed] [ computing] [ in] [ the] [ browser] | **6** | 6.0 |
-| FI | Hajautettu laskenta selaimessa | [H] [aj] [au] [tettu] [ las] [kenta] [ sel] [aim] [essa] | **9** | 3.3 |
-| EN | Write a function that sorts a list | [Write] [ a] [ function] [ that] [ sorts] [ a] [ list] | **7** | 5.0 |
-| FI | Kirjoita funktio joka lajittelee listan | [K] [irj] [oita] [ funkt] [io] [ joka] [ laj] [ittel] [ee] [ listan] | **10** | 4.0 |
+![Tokenisointivertailu EN/FI](/images/tokenization-example.png)
 
 **Huomaa miten:**
 - Englannin yleiset sanat (`the`, `in`, `a`, `function`) ovat kokonaisia tokeneita