Keep-alive: client pre-warms the model to hide idle cold-start

Adds GET /api/warmup that lazy-loads the model (and resets the idle clock)
without scoring — fast if already loaded. The Blazor panel fires it
fire-and-forget when it mounts (right after the user's first analysis), so the
model loads (~1.5s) during the seconds before they click Measure. Reconciles the
idle-unload (frees ~1.2GB when the site is quiet) with a warm experience while
the site is in active use — self-regulating, no always-on pinger.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: peopleworks

src/SignsOfAI.Perplexity.Api/Engine/IPerplexityEngine.cs

@@ -25,4 +25,8 @@ public interface IPerplexityEngine
 
     /// <summary>Runs one forward pass and returns the perplexity of <paramref name="text"/>.</summary>
     Task<PerplexityRaw> ScoreAsync(string text, CancellationToken ct = default);
+
+    /// <summary>Ensures the model is loaded into RAM (and resets the idle clock) without scoring —
+    /// so a subsequent request is warm. Cheap if already loaded.</summary>
+    Task WarmupAsync(CancellationToken ct = default);
 }

src/SignsOfAI.Perplexity.Api/Engine/OnnxPerplexityEngine.cs

@@ -72,6 +72,14 @@ public async Task<PerplexityRaw> ScoreAsync(string text, CancellationToken ct =
         finally { Release(); }
     }
 
+    /// <summary>Loads the model (if needed) and refreshes the idle clock, without running inference.</summary>
+    public async Task WarmupAsync(CancellationToken ct = default)
+    {
+        if (!_filesReady) return;
+        _ = await AcquireAsync(ct);
+        Release();
+    }
+
     // ── Load / unload lifecycle ──────────────────────────────────────────────
     private async Task<(InferenceSession, Tokenizer)> AcquireAsync(CancellationToken ct)
     {

src/SignsOfAI.Perplexity.Api/Program.cs

@@ -44,6 +44,16 @@
 app.MapGet("/healthz", (IPerplexityEngine engine) =>
     engine.IsReady ? Results.Ok("ready") : Results.StatusCode(503));
 
+// Cheap pre-warm: clients call this when the user is about to measure so the model is already
+// in RAM (hides the ~1.5s cold reload after an idle-unload). Returns fast if already loaded.
+app.MapGet("/api/warmup", async (IPerplexityEngine engine, CancellationToken ct) =>
+{
+    if (!engine.IsReady) return Results.Json(new { modelLoaded = false, ready = false });
+    var sw = System.Diagnostics.Stopwatch.StartNew();
+    await engine.WarmupAsync(ct);
+    return Results.Json(new { modelLoaded = engine.IsLoaded, ready = true, elapsedMs = sw.ElapsedMilliseconds });
+});
+
 app.MapPost("/api/perplexity", async (
     PerplexityRequest req, IPerplexityEngine engine, PerplexityScorer scorer, CancellationToken ct) =>
 {

src/SignsOfAI.Web/Components/PerplexityPanel.razor

@@ -86,6 +86,8 @@
     {
         _settings = await Perplexity.LoadSettingsAsync(Storage);
         _endpointEdit = _settings.Endpoint;
+        // Pre-warm the server while the user reads their results, so "Measure" is warm.
+        if (_settings.IsConfigured) _ = Perplexity.WarmupAsync(_settings.Endpoint);
     }
 
     private async Task SaveEndpoint()

src/SignsOfAI.Web/Services/PerplexityClient.cs

@@ -60,6 +60,13 @@ public async Task<PerplexitySettings> LoadSettingsAsync(BrowserStorage storage)
     public async Task SaveSettingsAsync(BrowserStorage storage, PerplexitySettings settings) =>
         await storage.SetAsync(StorageKey, JsonSerializer.Serialize(settings, PerplexityJsonContext.Default.PerplexitySettings));
 
+    /// <summary>Fire-and-forget pre-warm so the server model is loaded before the user clicks Measure.</summary>
+    public async Task WarmupAsync(string endpoint, CancellationToken ct = default)
+    {
+        try { using var _ = await http.GetAsync(endpoint.TrimEnd('/') + "/api/warmup", ct); }
+        catch { /* best-effort; a cold measure just costs ~1.5s */ }
+    }
+
     /// <summary>Measures perplexity for <paramref name="text"/>. Throws with a friendly message on failure.</summary>
     public async Task<PerplexityResult> MeasureAsync(string endpoint, string text, string lang, CancellationToken ct = default)
     {