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>

apps/stage-pocket/src/pages/index.vue

@@ -57,7 +57,7 @@ const hearingPipeline = useHearingSpeechInputPipeline()
 const { removeStreamingTranscriptionConsumer, transcribeForRecording, transcribeForMediaStream, stopStreamingTranscription } = hearingPipeline
 const { supportsStreamInput } = storeToRefs(hearingPipeline)
 const consciousnessStore = useConsciousnessStore()
-const { activeProvider: activeChatProvider, activeModel: activeChatModel } = storeToRefs(consciousnessStore)
+const { activeProvider: activeChatProvider, activeModel: activeChatModel, activeTemperature, activeTopP } = storeToRefs(consciousnessStore)
 const chatStore = useChatStore()
 
 /** Identifies this page in the shared streaming transcription session. */
@@ -91,7 +91,12 @@ async function sendVoiceInputTextToChat(text: string | undefined) {
 
     const provider = await consciousnessStore.getChatProviderInstance(providerId)
 
-    await chatStore.ingest(text, { model, chatProvider: provider })
+    await chatStore.ingest(text, {
+      model,
+      chatProvider: provider,
+      temperature: activeTemperature.value,
+      topP: activeTopP.value,
+    })
   }
   catch (error) {
     console.error('Failed to send chat from voice:', error)

apps/stage-web/src/pages/index.vue

@@ -59,7 +59,7 @@ const hearingPipeline = useHearingSpeechInputPipeline()
 const { removeStreamingTranscriptionConsumer, stopStreamingTranscription, transcribeForMediaStream, transcribeForRecording } = hearingPipeline
 const { supportsStreamInput } = storeToRefs(hearingPipeline)
 const consciousnessStore = useConsciousnessStore()
-const { activeProvider: activeChatProvider, activeModel: activeChatModel } = storeToRefs(consciousnessStore)
+const { activeProvider: activeChatProvider, activeModel: activeChatModel, activeTemperature, activeTopP } = storeToRefs(consciousnessStore)
 const chatStore = useChatStore()
 
 /** Identifies this page in the shared streaming transcription session. */
@@ -93,7 +93,12 @@ async function sendVoiceInputTextToChat(text: string | undefined) {
 
     const provider = await consciousnessStore.getChatProviderInstance(providerId)
 
-    await chatStore.ingest(text, { model, chatProvider: provider })
+    await chatStore.ingest(text, {
+      model,
+      chatProvider: provider,
+      temperature: activeTemperature.value,
+      topP: activeTopP.value,
+    })
   }
   catch (error) {
     console.error('Failed to send chat from voice:', error)

apps/stage-web/src/pages/settings/characters/components/CharacterDialog.vue

@@ -40,6 +40,7 @@ const form = reactive({
   // Capability: LLM
   llmModel: '',
   llmTemperature: 0.7,
+  llmTopP: 1.0,
 
   // Capability: TTS
   ttsVoiceId: '',
@@ -61,6 +62,7 @@ watch(() => props.character, (char) => {
 
     form.llmModel = llm?.config.llm?.model || ''
     form.llmTemperature = llm?.config.llm?.temperature || 0.7
+    form.llmTopP = llm?.config.llm?.topP || 1.0
 
     form.ttsVoiceId = tts?.config.tts?.voiceId || ''
     form.ttsSpeed = tts?.config.tts?.speed || 1.0
@@ -74,6 +76,7 @@ watch(() => props.character, (char) => {
     form.description = ''
     form.llmModel = 'gpt-4o-mini'
     form.llmTemperature = 0.7
+    form.llmTopP = 1.0
     form.ttsVoiceId = ''
     form.ttsSpeed = 1.0
   }
@@ -108,6 +111,7 @@ async function handleSubmit() {
           llm: {
             model: form.llmModel,
             temperature: form.llmTemperature,
+            topP: form.llmTopP,
           },
         },
       },
@@ -153,15 +157,9 @@ async function handleSubmit() {
         characterId: form.characterId,
         version: form.version,
         coverUrl: form.coverUrl,
+        capabilities: payload.capabilities,
       })
       trackCharacterUpdated({ character_id: props.character.id })
-      // Capabilities/I18n update not supported in simple UpdateCharacterSchema yet?
-      // Checking types/character.ts: UpdateCharacterSchema only has version, coverUrl, characterId.
-      // So deep update is not supported by the simple endpoint yet?
-      // The plan said "update(id, payload)".
-      // The backend `update` endpoint only updates the `character` table fields.
-      // To update relations, we'd need specific endpoints or a smarter update endpoint.
-      // I will only update basic info for now.
     }
     else {
       await characterStore.create(payload)
@@ -270,6 +268,18 @@ const isOpen = computed({
                       class="w-full border border-neutral-200 rounded-lg bg-white px-3 py-2 text-sm outline-none dark:border-neutral-700 focus:border-primary-500 dark:bg-neutral-800 focus:ring-2 focus:ring-primary-500/20"
                     >
                   </div>
+                  <!-- Use number input for top_p properly -->
+                  <div class="flex flex-col gap-1.5">
+                    <label class="text-sm text-neutral-700 font-medium dark:text-neutral-300">Top P</label>
+                    <input
+                      v-model.number="form.llmTopP"
+                      type="number"
+                      step="0.1"
+                      min="0"
+                      max="1"
+                      class="w-full border border-neutral-200 rounded-lg bg-white px-3 py-2 text-sm outline-none dark:border-neutral-700 focus:border-primary-500 dark:bg-neutral-800 focus:ring-2 focus:ring-primary-500/20"
+                    >
+                  </div>
                 </div>
 
                 <div class="space-y-4">

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 ---

server/apps/api/src/routes/characters/schema.ts

@@ -1,5 +1,5 @@
 import { createInsertSchema, createSelectSchema } from 'drizzle-valibot'
-import { array, literal, number, object, optional, pipe, string, transform, union } from 'valibot'
+import { array, literal, number, object, optional, string, union } from 'valibot'
 
 import * as schema from '../../schemas/characters'
 
@@ -17,6 +17,7 @@ export const CharacterCapabilityConfigSchema = object({
   apiBaseUrl: string(),
   llm: optional(object({
     temperature: number(),
+    topP: optional(number()),
     model: string(),
   })),
   tts: optional(object({
@@ -63,11 +64,6 @@ export const InsertCharacterCoverSchema = createInsertSchema(schema.characterCov
 export const CharacterPromptSchema = createSelectSchema(schema.characterPrompts)
 export const InsertCharacterPromptSchema = createInsertSchema(schema.characterPrompts)
 
-const DateSchema = pipe(
-  string(),
-  transform(v => new Date(v)),
-)
-
 export const CreateCharacterSchema = object({
   // TODO: Replace createInsertSchema-derived request bodies with explicit HTTP DTO schemas.
   // The current shape still leaks persistence fields such as ownerId/creatorId into the API boundary.
@@ -102,16 +98,13 @@ export const CreateCharacterSchema = object({
 
 // TODO: Split update request schema from DB insert schema.
 // This route should reject server-managed fields like id/ownerId/creatorId/timestamps instead of allowing them here.
-export const UpdateCharacterSchema = createInsertSchema(schema.character, {
-  id: optional(string()),
+export const UpdateCharacterSchema = object({
   version: optional(string()),
   coverUrl: optional(string()),
-  avatarUrl: optional(string()),
-  creatorRole: optional(string()),
-  priceCredit: optional(string()),
-  creatorId: optional(string()),
-  ownerId: optional(string()),
   characterId: optional(string()),
-  createdAt: optional(DateSchema),
-  updatedAt: optional(DateSchema),
+  capabilities: optional(array(createInsertSchema(schema.characterCapabilities, {
+    characterId: optional(string()),
+    type: CharacterCapabilityTypeSchema,
+    config: CharacterCapabilityConfigSchema,
+  }))),
 })

server/apps/api/src/services/domain/characters.ts

@@ -2,7 +2,7 @@ import type { Database } from '../../libs/db'
 import type { EngagementMetrics } from '../../otel'
 
 import { useLogger } from '@guiiai/logg'
-import { and, eq, isNull, or, sql } from 'drizzle-orm'
+import { and, eq, inArray, isNull, or, sql } from 'drizzle-orm'
 
 import * as schema from '../../schemas/characters'
 import * as userCharacterSchema from '../../schemas/user-character'
@@ -190,16 +190,60 @@ export function createCharacterService(db: Database, metrics?: EngagementMetrics
       return inserted
     },
 
-    async update(id: string, data: Partial<schema.NewCharacter>) {
-      // TODO: Return a stable single-object response shape for HTTP callers.
-      // leaking Drizzle returning() arrays across the service boundary makes route contracts drift.
-      const result = await db.update(schema.character)
-        .set({ ...data, updatedAt: new Date() })
+    async update(id: string, data: {
+      version?: string
+      coverUrl?: string
+      characterId?: string
+      capabilities?: {
+        type: 'llm' | 'tts' | 'vlm' | 'asr'
+        config: any
+      }[]
+    }) {
+      const { capabilities, ...characterData } = data
+
+      const result = await db.transaction(async (tx) => {
+        let updatedChar
+        if (Object.keys(characterData).length > 0) {
+          const [res] = await tx.update(schema.character)
+            .set({ ...characterData, updatedAt: new Date() })
             .where(and(
               eq(schema.character.id, id),
               isNull(schema.character.deletedAt),
             ))
             .returning()
+          updatedChar = res
+        }
+
+        if (capabilities) {
+          const submittedTypes = capabilities.map(c => c.type)
+          if (submittedTypes.length > 0) {
+            await tx.delete(schema.characterCapabilities)
+              .where(and(
+                eq(schema.characterCapabilities.characterId, id),
+                inArray(schema.characterCapabilities.type, submittedTypes),
+              ))
+
+            await tx.insert(schema.characterCapabilities).values(
+              capabilities.map(c => ({ ...c, characterId: id })),
+            )
+          }
+        }
+
+        if (updatedChar) {
+          return updatedChar
+        }
+
+        const fallback = await tx.query.character.findFirst({
+          where: and(
+            eq(schema.character.id, id),
+            isNull(schema.character.deletedAt),
+          ),
+        })
+        if (!fallback)
+          throw new Error('Character not found')
+        return fallback
+      })
+
       logger.withFields({ id }).log('Updated character')
       return result
     },

server/apps/api/src/types/character-capability.ts

@@ -7,6 +7,7 @@ interface CharacterCapabilityBaseConfig {
 export interface CharacterCapabilityConfig extends CharacterCapabilityBaseConfig {
   llm: {
     temperature: number
+    topP?: number
     model: string
   }
   tts: {