feat: add support for temperature, top_p to be configurable (#2200)

Co-authored-by: twix03 <sathwikbalaa@gmail.com>
Co-authored-by: RainbowBird <git@luoling.moe>

packages/core-agent/src/runtime/chat-orchestrator-runtime.ts

@@ -62,6 +62,10 @@ export interface ChatOrchestratorSendOptions {
   toolReferences?: ChatToolReference[]
   /** Original transport input metadata used by bridge/devtools observers. */
   input?: ChatStreamEventContext['input']
+  /** Temperature for the LLM request. */
+  temperature?: number
+  /** Top_p for the LLM request. */
+  topP?: number
 }
 
 interface QueuedSend {
@@ -766,6 +770,8 @@ export function createChatOrchestratorRuntime(deps: ChatOrchestratorRuntimeDeps)
           roundId: correlation.roundId,
         },
         tools: options.tools,
+        temperature: options.temperature,
+        topP: options.topP,
         waitForTools: true,
         onMessages: (messages) => {
           const currentTurnMessages = messages.slice(providerInputMessageCount)

packages/core-agent/src/runtime/llm-service.ts

@@ -198,6 +198,8 @@ export async function streamFrom({
         messages: sanitized,
         headers: options?.headers,
         streamOptions: { includeUsage: true },
+        temperature: options?.temperature,
+        topP: options?.topP,
         stopWhen: stepCountAtLeast(10),
         tools,
         toolChoice: options?.toolChoice,

packages/core-agent/src/types/llm.ts

@@ -34,6 +34,17 @@ export interface StreamOptions {
     conversationId: string
     roundId: string
   }
+  /**
+   * The temperature parameter controls the randomness of the model's output.
+   * A lower value results in more deterministic and focused output, while a
+   * higher value increases creativity and diversity.
+   */
+  temperature?: number
+  /**
+   * The top_p parameter controls nucleus sampling. A value of 0.1 means only
+   * the tokens comprising the top 10% probability mass are considered.
+   */
+  topP?: number
   toolsCompatibility?: Map<string, boolean>
   supportsTools?: boolean
   waitForTools?: boolean

packages/i18n/src/locales/en/settings.yaml

@@ -685,6 +685,10 @@ pages:
             search_results: Found {count} of {total} models
             show_less: Show less
             show_more: Show more
+            temperature_label: Temperature
+            temperature_description: Control the randomness of the model's output. A lower value (e.g. 0.2) results in more deterministic, precise, and analytical output, while a higher value (e.g. 1.0) increases creativity and diversity.
+            top_p_label: Top P
+            top_p_description: Control the diversity of the model's output via nucleus sampling. A lower value (e.g. 0.2) limits the output to only high-probability tokens, while a higher value (e.g. 1.0) includes a broader set of possibilities.
             subtitle: Select a default model from the provider
             title: Model
           model-options:

packages/i18n/src/locales/zh-Hans/settings.yaml

@@ -660,6 +660,10 @@ pages:
             search_results: 找到 {count} / {total} 个模型
             show_less: 显示更多
             show_more: 收起
+            temperature_label: 温度
+            temperature_description: 控制模型输出的随机性。较低的温度（例如 0.2）会使输出更精确和确定；较高的温度（例如 1.0）会增加创造力和多样性。
+            top_p_label: Top P
+            top_p_description: 控制模型输出的多样性（核采样）。较低的值（例如 0.2）会将输出限制在概率较高的词汇；较高的值（例如 1.0）则会包含更广泛的可能性。
             subtitle: 选择一个默认模型
             title: 模型
           model-options:

packages/provider-inference/src/providers/cloud/amazon-bedrock/index.ts

@@ -108,6 +108,7 @@ function createBedrockConverseProvider(config: {
           inferenceConfig: {
             maxTokens: body.max_tokens || 4096,
             ...(body.temperature !== undefined && { temperature: body.temperature }),
+            ...(body.top_p !== undefined && { topP: body.top_p }),
           },
         }
         if (system)

packages/stage-pages/src/pages/settings/modules/consciousness.vue

@@ -6,7 +6,7 @@ import { useConsciousnessStore } from '@proj-airi/stage-ui/stores/modules/consci
 import { useConsciousnessSettingsStore } from '@proj-airi/stage-ui/stores/modules/consciousness-settings'
 import { useProviderConfigStore } from '@proj-airi/stage-ui/stores/providers/config'
 import { useProviderStore } from '@proj-airi/stage-ui/stores/providers/provider'
-import { FieldCheckbox } from '@proj-airi/ui'
+import { FieldCheckbox, FieldRange } from '@proj-airi/ui'
 import { storeToRefs } from 'pinia'
 import { watch } from 'vue'
 import { useI18n } from 'vue-i18n'
@@ -29,6 +29,8 @@ const {
   providerModels,
   isLoadingActiveProviderModels,
   activeProviderModelError,
+  activeTemperature,
+  activeTopP,
 } = storeToRefs(consciousnessStore)
 
 const { t } = useI18n()
@@ -304,6 +306,29 @@ async function updateReasoning(value: boolean) {
     </section>
   </div>
 
+  <div v-if="activeProvider" :class="['bg-neutral-50 dark:bg-[rgba(0,0,0,0.3)]', 'rounded-xl', 'p-4', 'flex flex-col gap-4', 'mt-4']">
+    <div :class="['flex flex-col gap-4']">
+      <FieldRange
+        v-model="activeTemperature"
+        :label="t('settings.pages.modules.consciousness.sections.section.provider-model-selection.temperature_label')"
+        :description="t('settings.pages.modules.consciousness.sections.section.provider-model-selection.temperature_description')"
+        :min="0"
+        :max="2"
+        :step="0.1"
+        :format-value="value => value.toFixed(1)"
+      />
+      <FieldRange
+        v-model="activeTopP"
+        :label="t('settings.pages.modules.consciousness.sections.section.provider-model-selection.top_p_label')"
+        :description="t('settings.pages.modules.consciousness.sections.section.provider-model-selection.top_p_description')"
+        :min="0"
+        :max="1"
+        :step="0.1"
+        :format-value="value => value.toFixed(1)"
+      />
+    </div>
+  </div>
+
   <div
     v-motion
     text="neutral-200/50 dark:neutral-600/20" pointer-events-none

packages/stage-ui/src/stores/chat.ts

@@ -58,6 +58,10 @@ export interface ChatSendPayload {
   text: string
   /** Request-specific tools selected by their model-facing names. */
   tools?: ChatToolReference[]
+  /** Request-specific temperature override. */
+  temperature?: number
+  /** Request-specific top_p override. */
+  topP?: number
 }
 
 /** The durable messages appended while one chat request executes. */
@@ -401,6 +405,8 @@ export const useChatStore = defineStore('chat', () => {
       attachments: payload.attachments,
       input: payload.input,
       toolReferences: payload.tools,
+      temperature: payload.temperature ?? consciousnessStore.activeTemperature,
+      topP: payload.topP ?? consciousnessStore.activeTopP,
       // Resolve this function after the request reaches the per-session queue.
       // The history then contains tool names from every earlier queued turn.
       tools: async () => {

packages/stage-ui/src/stores/mods/api/context-bridge.ts

@@ -56,7 +56,7 @@ export const useContextBridgeStore = defineStore('mods:api:context-bridge', () =
   const contextObservability = useContextObservabilityStore()
   const characterOrchestratorStore = useCharacterOrchestratorStore()
   const consciousnessStore = useConsciousnessStore()
-  const { activeProvider, activeModel } = storeToRefs(consciousnessStore)
+  const { activeProvider, activeModel, activeTemperature, activeTopP } = storeToRefs(consciousnessStore)
   const streamingControl = useLlmStreamingControlStore()
 
   type SparkNotifyBridgeMessage
@@ -724,6 +724,8 @@ export const useContextBridgeStore = defineStore('mods:api:context-bridge', () =
               await chatOrchestrator.ingest(messageText, {
                 model: activeModel.value,
                 chatProvider,
+                temperature: activeTemperature.value,
+                topP: activeTopP.value,
                 input: {
                   type: 'input:text',
                   data: {

packages/stage-ui/src/stores/modules/consciousness.ts

@@ -41,6 +41,18 @@ export const useConsciousnessStore = defineStore('consciousness', () => {
     return providersStore.modelLoadError[activeProvider.value] || null
   })
 
+  const activeTemperature = useLocalStorageManualReset<number>(
+    'settings/consciousness/active-temperature',
+    0.7,
+    persistenceOptions,
+  )
+
+  const activeTopP = useLocalStorageManualReset<number>(
+    'settings/consciousness/active-top-p',
+    1.0,
+    persistenceOptions,
+  )
+
   const filteredModels = computed(() => {
     if (!modelSearchQuery.value.trim()) {
       return providerModels.value
@@ -109,6 +121,8 @@ export const useConsciousnessStore = defineStore('consciousness', () => {
   function resetState() {
     activeProvider.reset()
     resetModelSelection()
+    activeTemperature.reset()
+    activeTopP.reset()
   }
 
   return {
@@ -116,6 +130,8 @@ export const useConsciousnessStore = defineStore('consciousness', () => {
     configured,
     activeProvider,
     activeModel,
+    activeTemperature,
+    activeTopP,
     customModelName: activeCustomModelName,
     expandedDescriptions,
     modelSearchQuery,

packages/stage-ui/src/types/character.ts

@@ -21,6 +21,7 @@ export const CharacterCapabilityConfigSchema = object({
   apiBaseUrl: string(),
   llm: optional(object({
     temperature: number(),
+    topP: optional(number()),
     model: string(),
   })),
   tts: optional(object({
@@ -171,6 +172,10 @@ export const UpdateCharacterSchema = object({
   version: optional(string()),
   coverUrl: optional(string()),
   characterId: optional(string()),
+  capabilities: optional(array(object({
+    type: CharacterCapabilityTypeSchema,
+    config: CharacterCapabilityConfigSchema,
+  }))),
 })
 
 // --- Type Exports ---