Apply workspace temperature to agents and omit the parameter when unset

frontend/src/locales/en/common.js

@@ -251,7 +251,8 @@ const TRANSLATIONS = {
     temperature: {
       title: "LLM Temperature",
       "desc-end":
-        "The higher the number the more creative. For some models this can lead to incoherent responses when set too high.",
+        "The higher the number the more creative. For some models this can lead to incoherent responses when set too high. Leave blank to use your model provider's default.",
+      placeholder: "Provider default",
     },
   },
   "vector-workspace": {

frontend/src/pages/WorkspaceSettings/ChatSettings/ChatTemperatureSettings/index.jsx

@@ -1,22 +1,9 @@
 import { useTranslation } from "react-i18next";
-function recommendedSettings(provider = null) {
-  switch (provider) {
-    case "mistral":
-      return { temp: 0 };
-    default:
-      return { temp: 0.7 };
-  }
-}
 
-export default function ChatTemperatureSettings({
-  settings,
-  workspace,
-  setHasChanges,
-}) {
-  const defaults = recommendedSettings(settings?.LLMProvider);
+export default function ChatTemperatureSettings({ workspace, setHasChanges }) {
   const { t } = useTranslation();
   return (
-    <div>
+    <div className="flex flex-col gap-y-[8px]">
       <div className="flex flex-col gap-y-[8px]">
         <label htmlFor="name" className="block input-label">
           {t("chat.temperature.title")}
@@ -31,10 +18,9 @@ export default function ChatTemperatureSettings({
         min={0.0}
         step={0.1}
         onWheel={(e) => e.target.blur()}
-        defaultValue={workspace?.openAiTemp ?? defaults.temp}
+        defaultValue={workspace?.openAiTemp ?? ""}
         className="border-none bg-theme-settings-input-bg text-white placeholder:text-theme-settings-input-placeholder text-sm rounded-lg focus:outline-primary-button active:outline-primary-button outline-none block w-full p-2.5"
-        placeholder="0.7"
-        required={true}
+        placeholder={t("chat.temperature.placeholder")}
         autoComplete="off"
         onChange={() => setHasChanges(true)}
       />

frontend/src/pages/WorkspaceSettings/ChatSettings/index.jsx

@@ -85,7 +85,6 @@ export default function ChatSettings({ workspace }) {
           setHasChanges={setHasChanges}
         />
         <ChatTemperatureSettings
-          settings={settings}
           workspace={workspace}
           setHasChanges={setHasChanges}
         />

frontend/src/utils/types.js

@@ -1,7 +1,7 @@
 export function castToType(key, value) {
   const definitions = {
     openAiTemp: {
-      cast: (value) => Number(value),
+      cast: (value) => (value === "" ? null : Number(value)),
     },
     openAiHistory: {
       cast: (value) => Number(value),

server/__tests__/utils/chats/openaiCompatible.test.js

@@ -59,7 +59,6 @@ describe('OpenAICompatibleChat', () => {
         metrics: {},
       }),
       handleStream: jest.fn().mockResolvedValue('Mock streamed response'),
-      defaultTemp: 0.7,
     };
     getLLMProvider.mockReturnValue(mockLLMConnector);
     resolveProviderConnector.mockResolvedValue({

server/endpoints/agentWebsocket.js

@@ -34,6 +34,10 @@ function agentWebsocket(app) {
         return;
       }
 
+      AgentHandler.registerSessionSocket(
+        agentHandler.invocation.workspace_id,
+        socket
+      );
       socket.on("message", relayToSocket);
       socket.on("close", () => {
         // Abort the running agent loop (stop button, tab close, disconnect) so

server/models/workspace.js

@@ -270,7 +270,17 @@ const Workspace = {
       validatedUpdates.router_id = null;
     }
 
-    return this._update(id, validatedUpdates);
+    const result = await this._update(id, validatedUpdates);
+
+    // A live agent session keeps the settings it was built with, so end any
+    // open sessions for this workspace - the next agent message starts a
+    // fresh session with the updated settings.
+    if (result.workspace) {
+      const { AgentHandler } = require("../utils/agents");
+      AgentHandler.closeWorkspaceSessions(id);
+    }
+
+    return result;
   },
 
   /**

server/utils/AiProviders/anthropic/index.js

@@ -17,13 +17,22 @@ class AnthropicLLM {
    * These models reject `temperature`/`top_p`/`top_k` with a 400 error.
    * @type {string[]}
    */
-  noTemperatureModels = [
+  static noTemperatureModels = [
     "claude-opus-4-7",
     "claude-opus-4-8",
     "claude-sonnet-5",
     // Add other models here if identified
   ];
 
+  /**
+   * Whether the model supports the temperature parameter at all.
+   * @param {string} modelName
+   * @returns {boolean}
+   */
+  static modelSupportsTemperature(modelName = "") {
+    return !this.noTemperatureModels.some((model) => modelName.includes(model));
+  }
+
   constructor(embedder = null, modelPreference = null) {
     if (!process.env.ANTHROPIC_API_KEY)
       throw new Error("No Anthropic API key was set.");
@@ -50,7 +59,6 @@ class AnthropicLLM {
 
     this.maxTokens = null;
     this.embedder = embedder ?? new NativeEmbedder();
-    this.defaultTemp = 0.7;
     this.log(
       `Initialized with ${this.model}. Cache ${this.cacheControl ? `enabled (${this.cacheControl.ttl})` : "disabled"}`
     );
@@ -92,10 +100,9 @@ class AnthropicLLM {
    * @param {number} temperature - The temperature to use.
    * @returns {number|undefined} The temperature value or undefined if not supported.
    */
-  temperatureParam(temperature = this.defaultTemp) {
+  temperatureParam(temperature = this.temperature) {
     if (typeof temperature !== "number") return undefined;
-    if (this.noTemperatureModels.some((model) => this.model.includes(model)))
-      return undefined;
+    if (!AnthropicLLM.modelSupportsTemperature(this.model)) return undefined;
     return parseFloat(temperature);
   }
 
@@ -210,7 +217,10 @@ class AnthropicLLM {
     ];
   }
 
-  async getChatCompletion(messages = null, { temperature = 0.7 }) {
+  async getChatCompletion(
+    messages = null,
+    { temperature = this.temperature } = {}
+  ) {
     await this.assertModelMaxTokens();
     try {
       const systemContent = messages[0].content;
@@ -258,7 +268,10 @@ class AnthropicLLM {
     }
   }
 
-  async streamGetChatCompletion(messages = null, { temperature = 0.7 }) {
+  async streamGetChatCompletion(
+    messages = null,
+    { temperature = this.temperature } = {}
+  ) {
     await this.assertModelMaxTokens();
     const systemContent = messages[0].content;
     const measuredStreamRequest = await LLMPerformanceMonitor.measureStream({

server/utils/AiProviders/apipie/index.js

@@ -41,7 +41,6 @@ class ApiPieLLM {
     };
 
     this.embedder = embedder ?? new NativeEmbedder();
-    this.defaultTemp = 0.7;
 
     if (!fs.existsSync(cacheFolder))
       fs.mkdirSync(cacheFolder, { recursive: true });
@@ -187,7 +186,10 @@ class ApiPieLLM {
     ];
   }
 
-  async getChatCompletion(messages = null, { temperature = 0.7 }) {
+  async getChatCompletion(
+    messages = null,
+    { temperature = this.temperature } = {}
+  ) {
     if (!(await this.isValidChatCompletionModel(this.model)))
       throw new Error(
         `ApiPie chat: ${this.model} is not valid for chat completion!`
@@ -227,7 +229,10 @@ class ApiPieLLM {
     };
   }
 
-  async streamGetChatCompletion(messages = null, { temperature = 0.7 }) {
+  async streamGetChatCompletion(
+    messages = null,
+    { temperature = this.temperature } = {}
+  ) {
     if (!(await this.isValidChatCompletionModel(this.model)))
       throw new Error(
         `ApiPie chat: ${this.model} is not valid for chat completion!`

server/utils/AiProviders/azureOpenAi/index.js

@@ -39,7 +39,6 @@ class AzureOpenAiLLM {
     };
 
     this.embedder = embedder ?? new NativeEmbedder();
-    this.defaultTemp = 0.7;
     this.#log(
       `Initialized. Model "${this.model}" @ ${this.promptWindowLimit()} tokens.\nAPI-Version: ${this.apiVersion}.\nModel Type: ${this.isOTypeModel ? "reasoning" : "default"}`
     );
@@ -65,6 +64,15 @@ class AzureOpenAiLLM {
     }
   }
 
+  /**
+   * Whether the deployment supports the temperature parameter. Azure does not
+   * expose model metadata, so this relies on the user-declared AZURE_OPENAI_MODEL_TYPE.
+   * @returns {boolean}
+   */
+  static modelSupportsTemperature() {
+    return process.env.AZURE_OPENAI_MODEL_TYPE !== "reasoning";
+  }
+
   #log(text, ...args) {
     console.log(`\x1b[32m[AzureOpenAi]\x1b[0m ${text}`, ...args);
   }
@@ -150,7 +158,10 @@ class AzureOpenAiLLM {
     ];
   }
 
-  async getChatCompletion(messages = [], { temperature = 0.7 }) {
+  async getChatCompletion(
+    messages = [],
+    { temperature = this.temperature } = {}
+  ) {
     if (!this.model)
       throw new Error(
         "No AZURE_OPENAI_MODEL_PREF ENV defined. This must the name of a deployment on your Azure account for an LLM chat model like GPT-3.5."
@@ -185,7 +196,10 @@ class AzureOpenAiLLM {
     };
   }
 
-  async streamGetChatCompletion(messages = [], { temperature = 0.7 }) {
+  async streamGetChatCompletion(
+    messages = [],
+    { temperature = this.temperature } = {}
+  ) {
     if (!this.model)
       throw new Error(
         "No AZURE_OPENAI_MODEL_PREF ENV defined. This must the name of a deployment on your Azure account for an LLM chat model like GPT-3.5."

server/utils/AiProviders/bedrock/index.js

@@ -52,12 +52,21 @@ class AWSBedrockLLM {
     "us.deepseek.r1-v1:0",
   ];
 
-  noTemperatureModels = [
+  static noTemperatureModels = [
     "anthropic.claude-opus-4-7",
     "anthropic.claude-opus-4-8",
     "anthropic.claude-sonnet-5",
   ];
 
+  /**
+   * Whether the model supports the temperature parameter at all.
+   * @param {string} modelName
+   * @returns {boolean}
+   */
+  static modelSupportsTemperature(modelName = "") {
+    return !this.noTemperatureModels.some((model) => modelName.includes(model));
+  }
+
   constructor(embedder = null, modelPreference = null) {
     if (!process.env.AWS_BEDROCK_LLM_API_KEY)
       throw new Error("AWS_BEDROCK_LLM_API_KEY is required for AWS Bedrock.");
@@ -91,7 +100,6 @@ class AWSBedrockLLM {
     }
 
     this.embedder = embedder ?? new NativeEmbedder();
-    this.defaultTemp = 0.7;
     this.#log(
       `Initialized with model: ${this.model}. Region: ${this.region}. Context Window: ${contextWindowLimit}.`
     );
@@ -105,10 +113,9 @@ class AWSBedrockLLM {
     return Number(process.env.AWS_BEDROCK_LLM_MAX_TOKENS) || 4096;
   }
 
-  temperatureParam(temperature = this.defaultTemp) {
+  temperatureParam(temperature = this.temperature) {
     if (typeof temperature !== "number") return undefined;
-    if (this.noTemperatureModels.some((model) => this.model.includes(model)))
-      return undefined;
+    if (!AWSBedrockLLM.modelSupportsTemperature(this.model)) return undefined;
     return parseFloat(temperature);
   }
 
@@ -211,7 +218,10 @@ class AWSBedrockLLM {
 
   // --- Chat completions ---
 
-  async getChatCompletion(messages = null, { temperature }) {
+  async getChatCompletion(
+    messages = null,
+    { temperature = this.temperature } = {}
+  ) {
     if (!messages?.length)
       throw new Error(
         "AWSBedrock::getChatCompletion requires a non-empty messages array."
@@ -246,7 +256,10 @@ class AWSBedrockLLM {
     };
   }
 
-  async streamGetChatCompletion(messages = null, { temperature }) {
+  async streamGetChatCompletion(
+    messages = null,
+    { temperature = this.temperature } = {}
+  ) {
     if (!Array.isArray(messages) || messages.length === 0) {
       throw new Error(
         "AWSBedrock::streamGetChatCompletion requires a non-empty messages array."

server/utils/AiProviders/cerebras/index.js

@@ -26,7 +26,6 @@ class CerebrasLLM {
     this.limits = null;
 
     this.embedder = embedder ?? new NativeEmbedder();
-    this.defaultTemp = 0;
 
     CerebrasLLM.cacheContextWindows(true);
     this.#log(`Initialized with model: ${this.model}`);
@@ -197,7 +196,12 @@ class CerebrasLLM {
     return textResponse;
   }
 
-  async getChatCompletion(messages = null, { temperature = 0.7 }) {
+  async getChatCompletion(
+    messages = null,
+    // These models degrade quickly at higher temperatures, so an unset
+    // workspace temperature falls back to 0 instead of the provider default.
+    { temperature = this.temperature ?? 0 } = {}
+  ) {
     const result = await LLMPerformanceMonitor.measureAsyncFunction(
       this.openai.chat.completions
         .create({
@@ -234,7 +238,12 @@ class CerebrasLLM {
     };
   }
 
-  async streamGetChatCompletion(messages = null, { temperature = 0.7 }) {
+  async streamGetChatCompletion(
+    messages = null,
+    // These models degrade quickly at higher temperatures, so an unset
+    // workspace temperature falls back to 0 instead of the provider default.
+    { temperature = this.temperature ?? 0 } = {}
+  ) {
     const measuredStreamRequest = await LLMPerformanceMonitor.measureStream({
       func: this.openai.chat.completions.create({
         model: this.model,

server/utils/AiProviders/cohere/index.js

@@ -28,7 +28,6 @@ class CohereLLM {
     };
 
     this.embedder = embedder ?? new NativeEmbedder();
-    this.defaultTemp = 0.7;
     this.#log(
       `Initialized with model ${this.model}. ctx: ${this.promptWindowLimit()}`
     );
@@ -79,7 +78,10 @@ class CohereLLM {
     return [prompt, ...chatHistory, { role: "user", content: userPrompt }];
   }
 
-  async getChatCompletion(messages = null, { temperature = 0.7 }) {
+  async getChatCompletion(
+    messages = null,
+    { temperature = this.temperature } = {}
+  ) {
     const result = await LLMPerformanceMonitor.measureAsyncFunction(
       this.openai.chat.completions
         .create({
@@ -115,7 +117,10 @@ class CohereLLM {
     };
   }
 
-  async streamGetChatCompletion(messages = null, { temperature = 0.7 }) {
+  async streamGetChatCompletion(
+    messages = null,
+    { temperature = this.temperature } = {}
+  ) {
     const measuredStreamRequest = await LLMPerformanceMonitor.measureStream({
       func: this.openai.chat.completions.create({
         model: this.model,