@fugood/bricks-project 2.24.0-beta.4 → 2.24.0-beta.41

package/types/generators/LlmMlx.ts

@@ -0,0 +1,227 @@
+/* Auto generated by build script
+ *
+ * On-device LLM inference using Apple MLX framework on iOS/tvOS
+ *
+ *  ## Features
+ *  - Powered by MLX (Apple's ML framework optimized for Apple Silicon)
+ *  - Download models directly from HuggingFace Hub
+ *  - Streaming token generation
+ *  - Supports LLM and VLM (Vision Language Models)
+ *  - Requires iOS 17+ or tvOS 17+
+ */
+import type { SwitchCondInnerStateCurrentCanvas, SwitchCondData, SwitchDef } from '../switch'
+import type { Data, DataLink } from '../data'
+import type {
+  Brick,
+  Generator,
+  EventAction,
+  ActionWithDataParams,
+  ActionWithParams,
+  Action,
+  EventProperty,
+} from '../common'
+import type { TemplateEventPropsMap } from '../../utils/event-props'
+
+/* Load model */
+export type GeneratorMlxLLMActionLoadModel = ActionWithParams & {
+  __actionName: 'GENERATOR_MLX_LLM_LOAD_MODEL'
+  params?: Array<
+    | {
+        input: 'modelId'
+        value?: string | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'vlm'
+        value?: boolean | DataLink | EventProperty
+        mapping?: string
+      }
+  >
+}
+
+/* Run text completion */
+export type GeneratorMlxLLMActionCompletion = ActionWithParams & {
+  __actionName: 'GENERATOR_MLX_LLM_COMPLETION'
+  params?: Array<
+    | {
+        input: 'messages'
+        value?: Array<any> | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'tools'
+        value?: Array<any> | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'maxTokens'
+        value?: number | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'temperature'
+        value?: number | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'topP'
+        value?: number | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'repetitionPenalty'
+        value?: number | DataLink | EventProperty
+        mapping?: string
+      }
+  >
+}
+
+/* Stop text completion */
+export type GeneratorMlxLLMActionStopCompletion = Action & {
+  __actionName: 'GENERATOR_MLX_LLM_STOP_COMPLETION'
+}
+
+/* Release model context */
+export type GeneratorMlxLLMActionReleaseContext = Action & {
+  __actionName: 'GENERATOR_MLX_LLM_RELEASE_CONTEXT'
+}
+
+interface GeneratorMlxLLMDef {
+  /*
+Default property:
+{
+  "init": false,
+  "modelId": "mlx-community/Qwen3-4B-4bit",
+  "vlm": false,
+  "completionMessages": [
+    {
+      "role": "system",
+      "content": "You are a helpful assistant."
+    },
+    {
+      "role": "user",
+      "content": "Hello"
+    }
+  ],
+  "completionMaxTokens": 1024,
+  "completionTemperature": 0.6,
+  "completionTopP": 1,
+  "completionRepetitionContextSize": 20
+}
+  */
+  property?: {
+    /* Initialize model on generator init */
+    init?: boolean | DataLink
+    /* HuggingFace model ID or local path to model directory
+ e.g. "mlx-community/Qwen3-4B-4bit" */
+    modelId?: string | DataLink
+    /* Enable Vision Language Model (VLM) mode */
+    vlm?: boolean | DataLink
+    /* Chat messages (if first message has role 'system', it will be used as system prompt) */
+    completionMessages?:
+      | Array<
+          | DataLink
+          | {
+              role?: string | DataLink
+              content?: string | DataLink
+            }
+        >
+      | DataLink
+    /* Maximum tokens to generate */
+    completionMaxTokens?: number | DataLink
+    /* Temperature (0.0 to 2.0) */
+    completionTemperature?: number | DataLink
+    /* Top P sampling */
+    completionTopP?: number | DataLink
+    /* Repetition penalty factor */
+    completionRepetitionPenalty?: number | DataLink
+    /* Number of tokens to consider for repetition penalty */
+    completionRepetitionContextSize?: number | DataLink
+    /* Maximum KV cache size (context window). Uses rotating cache to limit memory.
+ Smaller values (e.g. 512) use less memory but lower quality.  
+ Larger values (e.g. 4096) use more memory but better quality.  
+ Leave empty for unlimited (model default). */
+    contextSize?: number | DataLink
+    /* Enable thinking mode (model-dependent, e.g. Qwen3).
+ When enabled, passes enable_thinking=true to the chat template.  
+ The model may output reasoning in &lt;think&gt; tags which will be extracted as reasoning_content. */
+    completionEnableThinking?: boolean | DataLink
+    /* Additional keyword arguments for chat template (object) */
+    completionChatTemplateKwargs?: {} | DataLink
+    /* Tools for chat mode using OpenAI-compatible function calling format
+ Format: Array of objects with {type, function: {name, description, parameters}} structure  
+ See: https://platform.openai.com/docs/guides/function-calling */
+    completionTools?: Array<{} | DataLink> | DataLink
+    /* Buttress connection settings for remote inference */
+    buttressConnectionSettings?:
+      | DataLink
+      | {
+          enabled?: boolean | DataLink
+          url?: string | DataLink
+          fallbackType?: 'use-local' | 'no-op' | DataLink
+          strategy?: 'prefer-local' | 'prefer-buttress' | 'prefer-best' | DataLink
+        }
+  }
+  events?: {
+    /* Event triggered when context state changes */
+    onContextStateChange?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['MlxLlm']['onContextStateChange']>
+    >
+    /* Error event */
+    onError?: Array<EventAction<string & keyof TemplateEventPropsMap['MlxLlm']['onError']>>
+    /* Completion streaming event (emitted for each token) */
+    onCompletion?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['MlxLlm']['onCompletion']>
+    >
+    /* Completion finished event */
+    onCompletionFinished?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['MlxLlm']['onCompletionFinished']>
+    >
+  }
+  outlets?: {
+    /* Context state */
+    contextState?: () => Data<string>
+    /* Model load progress (0-1) */
+    loadProgress?: () => Data<number>
+    /* Whether the model is evaluating */
+    isEvaluating?: () => Data<boolean>
+    /* Completion result */
+    completionResult?: () => Data<string>
+    /* Last token */
+    completionLastToken?: () => Data<string>
+  }
+}
+
+/* On-device LLM inference using Apple MLX framework on iOS/tvOS
+
+ ## Features  
+ - Powered by MLX (Apple's ML framework optimized for Apple Silicon)  
+ - Download models directly from HuggingFace Hub  
+ - Streaming token generation  
+ - Supports LLM and VLM (Vision Language Models)  
+ - Requires iOS 17+ or tvOS 17+ */
+export type GeneratorMlxLLM = Generator &
+  GeneratorMlxLLMDef & {
+    templateKey: 'GENERATOR_MLX_LLM'
+    switches?: Array<
+      SwitchDef &
+        GeneratorMlxLLMDef & {
+          conds?: Array<{
+            method: '==' | '!=' | '>' | '<' | '>=' | '<='
+            cond:
+              | SwitchCondInnerStateCurrentCanvas
+              | SwitchCondData
+              | {
+                  __typename: 'SwitchCondInnerStateOutlet'
+                  outlet:
+                    | 'contextState'
+                    | 'loadProgress'
+                    | 'isEvaluating'
+                    | 'completionResult'
+                    | 'completionLastToken'
+                  value: any
+                }
+          }>
+        }
+    >
+  }

package/types/generators/LlmOnnx.ts

@@ -1,4 +1,8 @@
-/* Auto generated by build script */
+/* Auto generated by build script
+ *
+ * Local LLM inference based on [transformers.js](https://huggingface.co/docs/transformers.js)
+ *  You can use any converted model on HuggingFace.
+ */
 import type { SwitchCondInnerStateCurrentCanvas, SwitchCondData, SwitchDef } from '../switch'
 import type { Data, DataLink } from '../data'
 import type {
@@ -10,6 +14,7 @@ import type {
   Action,
   EventProperty,
 } from '../common'
+import type { TemplateEventPropsMap } from '../../utils/event-props'
 
 /* Load the model */
 export type GeneratorOnnxLLMActionLoadModel = Action & {
@@ -110,7 +115,7 @@ Default property:
     audios?: Array<string | DataLink> | DataLink
     /* Tool call parser */
     toolCallParser?: 'llama3_json' | 'mistral' | 'hermes' | 'internlm' | 'phi4' | DataLink
-    /* Tools for chat mode using OpenAI-compatible function calling format  
+    /* Tools for chat mode using OpenAI-compatible function calling format
  Format: Array of objects with {type, function: {name, description, parameters}} structure  
  See: https://platform.openai.com/docs/guides/function-calling */
     tools?: Array<{} | DataLink> | DataLink
@@ -132,12 +137,12 @@ Default property:
     numBeams?: number | DataLink
     /* Do sampling */
     doSample?: boolean | DataLink
-    /* Executor candidates, descending order of priority  
+    /* Executor candidates, descending order of priority
   Default will be xnnpack, wasm, cpu */
     executors?:
       | Array<'qnn' | 'dml' | 'nnapi' | 'xnnpack' | 'coreml' | 'cpu' | 'wasm' | 'webgpu' | DataLink>
       | DataLink
-    /* Execution mode  
+    /* Execution mode
   Usually when the model has many branches, setting this option to `parallel` will give you better performance. */
     executionMode?: 'sequential' | 'parallel' | DataLink
     /* QNN backend */
@@ -149,32 +154,47 @@ Default property:
   }
   events?: {
     /* Event triggered when state change */
-    onContextStateChange?: Array<EventAction>
+    onContextStateChange?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['OnnxLlm']['onContextStateChange']>
+    >
     /* Event triggered on get function call request */
-    onFunctionCall?: Array<EventAction>
+    onFunctionCall?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['OnnxLlm']['onFunctionCall']>
+    >
     /* Event triggered on completion finished */
-    onCompletionFinished?: Array<EventAction>
+    onCompletionFinished?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['OnnxLlm']['onCompletionFinished']>
+    >
     /* Event triggered when error occurs */
-    onError?: Array<EventAction>
+    onError?: Array<EventAction<string & keyof TemplateEventPropsMap['OnnxLlm']['onError']>>
   }
   outlets?: {
     /* Context state */
-    contextState?: () => Data
+    contextState?: () => Data<string>
     /* Generated output */
-    generated?: () => Data
+    generated?: () => Data<string>
     /* Full result of generation */
-    fullResult?: () => Data
+    fullResult?: () => Data<string>
     /* Last function call */
-    lastFunctionCall?: () => Data
+    lastFunctionCall?: () => Data<{
+      id?: string
+      type?: string
+      function?: {
+        name?: string
+        arguments?: string
+        [key: string]: any
+      }
+      [key: string]: any
+    }>
   }
 }
 
-/* Local LLM inference based on [transformers.js](https://huggingface.co/docs/transformers.js)  
+/* Local LLM inference based on [transformers.js](https://huggingface.co/docs/transformers.js)
  You can use any converted model on HuggingFace. */
 export type GeneratorOnnxLLM = Generator &
   GeneratorOnnxLLMDef & {
     templateKey: 'GENERATOR_ONNX_LLM'
-    switches: Array<
+    switches?: Array<
       SwitchDef &
         GeneratorOnnxLLMDef & {
           conds?: Array<{

package/types/generators/LlmOpenAiCompat.ts

@@ -1,4 +1,17 @@
-/* Auto generated by build script */
+/* Auto generated by build script
+ *
+ * LLM inference using OpenAI-compatible API endpoints
+ *
+ *  ## Features
+ *  - Compatible with OpenAI API format
+ *  - Supports function calling
+ *  - Streaming responses
+ *  - Custom API endpoints, like
+ *    - OpenAI API: https://platform.openai.com/docs/guides/text?api-mode=chat
+ *    - Anthropic API: https://docs.anthropic.com/en/api/openai-sdk
+ *    - Gemini API: https://ai.google.dev/gemini-api/docs/openai
+ *    - llama.cpp server: https://github.com/ggml-org/llama.cpp/tree/master/tools/server
+ */
 import type { SwitchCondInnerStateCurrentCanvas, SwitchCondData, SwitchDef } from '../switch'
 import type { Data, DataLink } from '../data'
 import type {
@@ -10,6 +23,7 @@ import type {
   Action,
   EventProperty,
 } from '../common'
+import type { TemplateEventPropsMap } from '../../utils/event-props'
 
 /* Run text completion */
 export type GeneratorOpenAILLMActionCompletion = ActionWithParams & {
@@ -122,7 +136,7 @@ Default property:
             }
         >
       | DataLink
-    /* Tools for chat mode following OpenAI function calling format  
+    /* Tools for chat mode following OpenAI function calling format
  Format: Array of objects with {type, function: {name, description, parameters}} structure  
  See: https://platform.openai.com/docs/guides/function-calling */
     completionTools?: Array<{} | DataLink> | DataLink
@@ -157,25 +171,47 @@ Default property:
   }
   events?: {
     /* Error event */
-    onError?: Array<EventAction>
+    onError?: Array<EventAction<string & keyof TemplateEventPropsMap['OpenaiLlm']['onError']>>
     /* Completion event */
-    onCompletion?: Array<EventAction>
+    onCompletion?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['OpenaiLlm']['onCompletion']>
+    >
     /* Completion finished event */
-    onCompletionFinished?: Array<EventAction>
+    onCompletionFinished?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['OpenaiLlm']['onCompletionFinished']>
+    >
     /* Completion function call event */
-    onCompletionFunctionCall?: Array<EventAction>
+    onCompletionFunctionCall?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['OpenaiLlm']['onCompletionFunctionCall']>
+    >
   }
   outlets?: {
     /* Evaluating outlet */
-    isEvaluating?: () => Data
+    isEvaluating?: () => Data<boolean>
     /* Completion result outlet */
-    completionResult?: () => Data
+    completionResult?: () => Data<string>
     /* Completion details outlet */
-    completionDetails?: () => Data
+    completionDetails?: () => Data<{
+      text?: string
+      content?: string
+      finish_reason?: string
+      usage?: { [key: string]: any }
+      tool_calls?: Array<{
+        id?: string
+        type?: string
+        function?: {
+          name?: string
+          arguments?: string
+          [key: string]: any
+        }
+        [key: string]: any
+      }>
+      [key: string]: any
+    }>
   }
 }
 
-/* LLM inference using OpenAI-compatible API endpoints  
+/* LLM inference using OpenAI-compatible API endpoints
 
  ## Features  
  - Compatible with OpenAI API format  
@@ -189,7 +225,7 @@ Default property:
 export type GeneratorOpenAILLM = Generator &
   GeneratorOpenAILLMDef & {
     templateKey: 'GENERATOR_OPENAI_LLM'
-    switches: Array<
+    switches?: Array<
       SwitchDef &
         GeneratorOpenAILLMDef & {
           conds?: Array<{

package/types/generators/LlmQualcommAiEngine.ts

@@ -1,4 +1,7 @@
-/* Auto generated by build script */
+/* Auto generated by build script
+ *
+ * Local LLM inference using Qualcomm AI Engine
+ */
 import type { SwitchCondInnerStateCurrentCanvas, SwitchCondData, SwitchDef } from '../switch'
 import type { Data, DataLink } from '../data'
 import type {
@@ -10,6 +13,7 @@ import type {
   Action,
   EventProperty,
 } from '../common'
+import type { TemplateEventPropsMap } from '../../utils/event-props'
 
 /* Load the model */
 export type GeneratorQnnLlmActionLoadModel = Action & {
@@ -108,7 +112,7 @@ Default property:
       | DataLink
     /* SOC model */
     socModel?: 'X Elite' | 'X Plus' | '8 Elite' | '8 Gen 3' | 'QCS8550' | DataLink
-    /* Custom model base URL  
+    /* Custom model base URL
  The model should be bundled, for details see https://github.com/mybigday/node-qnn-llm?tab=readme-ov-file#bundled-file */
     customModelUrl?: string | DataLink
     /* Custom model MD5 */
@@ -142,7 +146,7 @@ Default property:
     stopWords?: Array<string | DataLink> | DataLink
     /* Tool call parser */
     toolCallParser?: 'llama3_json' | 'mistral' | 'hermes' | 'internlm' | 'phi4' | DataLink
-    /* Tools for chat mode using OpenAI-compatible function calling format  
+    /* Tools for chat mode using OpenAI-compatible function calling format
  Format: Array of objects with {type, function: {name, description, parameters}} structure  
  See: https://platform.openai.com/docs/guides/function-calling */
     tools?: Array<{} | DataLink> | DataLink
@@ -165,25 +169,53 @@ Default property:
   }
   events?: {
     /* Event triggered when context state changes */
-    onContextStateChange?: Array<EventAction>
+    onContextStateChange?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['QnnLlm']['onContextStateChange']>
+    >
     /* Event triggered when generate is done */
-    onGenerate?: Array<EventAction>
+    onGenerate?: Array<EventAction<string & keyof TemplateEventPropsMap['QnnLlm']['onGenerate']>>
     /* Event triggered on get function call request */
-    onFunctionCall?: Array<EventAction>
+    onFunctionCall?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['QnnLlm']['onFunctionCall']>
+    >
     /* Event triggered when error occurs */
-    onError?: Array<EventAction>
+    onError?: Array<EventAction<string & keyof TemplateEventPropsMap['QnnLlm']['onError']>>
   }
   outlets?: {
     /* Context state */
-    contextState?: () => Data
+    contextState?: () => Data<string>
     /* Generation result */
-    result?: () => Data
+    result?: () => Data<string>
     /* Full context (Prompt + Generation Result) */
-    fullContext?: () => Data
+    fullContext?: () => Data<string>
     /* Last function call details */
-    lastFunctionCall?: () => Data
+    lastFunctionCall?: () => Data<{
+      id?: string
+      type?: string
+      function?: {
+        name?: string
+        arguments?: string
+        [key: string]: any
+      }
+      [key: string]: any
+    }>
     /* Completion details */
-    completionDetails?: () => Data
+    completionDetails?: () => Data<{
+      text?: string
+      content?: string
+      reasoning_content?: string
+      tool_calls?: Array<{
+        id?: string
+        type?: string
+        function?: {
+          name?: string
+          arguments?: string
+          [key: string]: any
+        }
+        [key: string]: any
+      }>
+      [key: string]: any
+    }>
   }
 }
 
@@ -191,7 +223,7 @@ Default property:
 export type GeneratorQnnLlm = Generator &
   GeneratorQnnLlmDef & {
     templateKey: 'GENERATOR_QNN_LLM'
-    switches: Array<
+    switches?: Array<
       SwitchDef &
         GeneratorQnnLlmDef & {
           conds?: Array<{