Ship Assistent 0.8.1: split modules and shared+personal vector memory.

Personal RAG never leaks into the shared store; retrieve merges shared plus the persona chain, with personal overwrite on kind+key.

Co-authored-by: Cursor <cursoragent@cursor.com>

AssistentVram.cs

@@ -0,0 +1,118 @@
+using System;
+using System.Net.Http;
+using System.Text;
+using System.Threading.Tasks;
+using Newtonsoft.Json.Linq;
+using SwarmUI.Accounts;
+using SwarmUI.Utils;
+
+namespace Mrleo1nid.SwarmAssistent;
+
+/// <summary>VRAM handover between Ollama and the image backend: park the chat model before
+/// Generate, warm it again once the user is back in the chat. Embed / memory models are never
+/// parked — they are tiny and reloading them stalls every retrieve.</summary>
+public partial class SwarmAssistentExtension
+{
+    const string WarmKeepAlive = "15m";
+
+    /// <summary>Unloads the chat model from VRAM (<c>keep_alive: 0</c>) so Krea 2 gets the whole GPU.</summary>
+    public async Task<JObject> AssistentParkLlm(Session session, string baseUrl, string model)
+    {
+        string root = NormalizeBaseUrl(baseUrl);
+        string name = (model ?? "").Trim();
+        if (string.IsNullOrWhiteSpace(name))
+        {
+            return new JObject { ["error"] = "model is required" };
+        }
+        if (LooksLikeEmbedModel(name))
+        {
+            return new JObject { ["success"] = true, ["parked"] = false, ["skipped"] = "memory model — never parked" };
+        }
+        JObject generate = new()
+        {
+            ["model"] = name,
+            ["prompt"] = "",
+            ["stream"] = false,
+            ["keep_alive"] = 0,
+        };
+        (bool ok, string body) = await PostOllamaJson(root, "/api/generate", generate);
+        if (!ok)
+        {
+            // Older Ollama builds only unload through /api/chat.
+            JObject chat = new()
+            {
+                ["model"] = name,
+                ["messages"] = new JArray(),
+                ["stream"] = false,
+                ["keep_alive"] = 0,
+            };
+            (ok, body) = await PostOllamaJson(root, "/api/chat", chat);
+        }
+        if (!ok)
+        {
+            Logs.Debug($"AssistentParkLlm {name}: {Clip(body, 200)}");
+            return new JObject { ["success"] = true, ["parked"] = false, ["note"] = Clip(body, 200) };
+        }
+        return new JObject { ["success"] = true, ["parked"] = true, ["model"] = name, ["base_url"] = root };
+    }
+
+    /// <summary>Single-token chat so the model is resident again by the time the user types.</summary>
+    public async Task<JObject> AssistentWarmLlm(Session session, string baseUrl, string model, string persona = null)
+    {
+        string root = NormalizeBaseUrl(baseUrl);
+        string name = (model ?? "").Trim();
+        if (string.IsNullOrWhiteSpace(name))
+        {
+            return new JObject { ["error"] = "model is required" };
+        }
+        if (LooksLikeEmbedModel(name))
+        {
+            return new JObject { ["success"] = true, ["warmed"] = false, ["skipped"] = "memory model" };
+        }
+        int numCtx = CfgInt("num_ctx", DefaultNumCtxFallback);
+        JObject payload = new()
+        {
+            ["model"] = name,
+            ["stream"] = false,
+            ["messages"] = new JArray
+            {
+                new JObject { ["role"] = "user", ["content"] = "ok" },
+            },
+            ["options"] = new JObject
+            {
+                ["num_ctx"] = numCtx,
+                ["num_predict"] = 1,
+            },
+            ["keep_alive"] = WarmKeepAlive,
+        };
+        (bool ok, string body) = await PostOllamaJson(root, "/api/chat", payload);
+        if (!ok)
+        {
+            Logs.Debug($"AssistentWarmLlm {name}: {Clip(body, 200)}");
+            return new JObject { ["success"] = true, ["warmed"] = false, ["note"] = Clip(body, 200) };
+        }
+        return new JObject
+        {
+            ["success"] = true,
+            ["warmed"] = true,
+            ["model"] = name,
+            ["num_ctx"] = numCtx,
+            ["keep_alive"] = WarmKeepAlive,
+        };
+    }
+
+    static async Task<(bool ok, string body)> PostOllamaJson(string root, string route, JObject payload)
+    {
+        try
+        {
+            using StringContent content = new(payload.ToString(Newtonsoft.Json.Formatting.None), Encoding.UTF8, "application/json");
+            using HttpResponseMessage resp = await HttpClient.PostAsync($"{root}{route}", content);
+            string body = await resp.Content.ReadAsStringAsync();
+            return (resp.IsSuccessStatusCode, resp.IsSuccessStatusCode ? body : $"HTTP {(int)resp.StatusCode}: {body}");
+        }
+        catch (Exception ex)
+        {
+            return (false, ex.Message);
+        }
+    }
+}