bricks/docs/en/llm.md

LlmMsgAudio

Widget Functionality: Handles voice message streams, supports segmentation by Chinese and non-Chinese punctuation, and plays the model's audio response through an audio player.
Type: Regular widget
Parent Widget: bricks.UpStreaming

Initialization Parameters

Parameter	Type	Description
opts	Object	Configuration options passed to the parent class, handled by UpStreaming

Note

: This widget internally initializes the following properties:

• olddata / data: Used for accumulating text data
• cn_p / other_p: Arrays of common Chinese and non-Chinese punctuation marks
• audio: Audio playback instance created using AudioPlayer({})

Main Events

No explicit events are bound, but the following methods are overridden to participate in data stream processing:

• send(data): Receives incremental text, splits it by language-specific punctuation, and sends playable segments
• go(): Initiates a request, sets the audio source, and plays the voice based on the response content

---

ModelOutput

Widget Functionality: Displays output from large language models, supporting dynamic updates, status indication, user feedback (like/dislike), and integrated TTS audio playback.
Type: Container widget
Parent Widget: bricks.VBox

Initialization Parameters

Parameter	Type	Required	Description
opts.modelname	String	No	Displayed model name
opts.icon	String	No	Model icon URL; defaults to a default LLM icon if not provided
opts.response_mode	String	No	Response mode: stream, sync, or async
opts.estimate_url	String	No	API endpoint for submitting user ratings (e.g., like/dislike)
opts.textvoice	Boolean	No	Whether to enable text-to-speech reading
opts.tts_url	String	No	TTS service API endpoint for generating speech

Main Events

Event Name	Trigger Condition	Callback Function
click (on like icon)	User clicks the "like" icon	estimate_llm(icon, 1)
click (on unlike icon)	User clicks the "dislike" icon	estimate_llm(icon, -1)

Other Behaviors:

• update_data(data): Receives model output and updates displayed content
• finish(): Called when streaming ends; currently logs only

---

LlmModel

Widget Functionality: Encapsulates the logic for calling a single large language model, including input preprocessing, request sending, response stream handling, and result rendering. Supports multiple interaction modes (synchronous, streaming, asynchronous).
Type: Regular widget
Parent Widget: bricks.JsWidget

Initialization Parameters

Parameter	Type	Required	Description
llmio	Object	Yes	The owning LlmIO instance, used for shared configuration
opts.model	String	Yes	Model identifier
opts.modelname	String	Yes	Display name
opts.url	String	Yes	Model API endpoint
opts.icon	String	No	Custom icon path
opts.params	Object	No	Additional request parameters
opts.user_message_format	String	No	Template format for user messages
opts.system_message_format	String	No	Template format for system messages
opts.llm_message_format	Object	No	Structure definition for assistant messages
opts.use_session	Boolean	No	Whether to maintain conversation context
opts.input_from	String	No	Source identifier for input
opts.textvoice	Boolean	No	Whether to enable voice output
opts.tts_url	String	No	TTS API endpoint
opts.response_mode	String	Yes	Request mode: stream, sync, or async

Main Events

Event Name	Trigger Condition	Callback Function
click (on title)	Clicking the model title area	show_setup_panel(event) — Can be extended by subclasses

Key Internal Methods:

• model_inputed(data): Triggers a model request upon receiving input data
• chunk_response(mout, line): Processes each chunk in a streaming response
• is_accept_source(source): Determines whether data from a given source should be accepted

---

LlmIO

Widget Functionality: Serves as the core container for the overall LLM interaction interface. Manages multiple model instances, input dialogs, knowledge base configurations, user inputs, and output displays. Provides a unified entry point for coordinating multiple models.
Type: Container widget
Parent Widget: bricks.VBox

Initialization Parameters

Parameter	Type	Required	Description
opts.user_icon	String	No	User avatar icon URL
opts.list_models_url	String	No	API endpoint for retrieving available model list
opts.input_fields	Array	Yes	Form field definitions (e.g., prompt, temperature, etc.)
opts.models	Array	Yes	Initial array of model configurations
opts.tts_url	String	No	Global TTS service URL
opts.get_kdb_url	String	No	API endpoint for retrieving knowledge base list
opts.estimate_url	String	No	Endpoint for submitting user feedback
opts.enabled_kdb	Boolean	No	Whether to enable knowledge base augmentation

Example models structure:

{
  model: "qwen",
  modelname: "通义千问",
  url: "/api/llm/qwen",
  response_mode: "stream"
}

Main Events

Event Name	Trigger Condition	Callback Function
click (i_w)	Clicking the input button	open_input_widget(event) — Opens input form dialog
click (nm_w)	Clicking the "Add Model" button	open_search_models(event) — Opens model selection panel
click (kdb_w)	Clicking the knowledge base settings button	setup_kdb(event) — Opens KDB configuration form
submit (in form)	User submits the input form	handle_input(event) — Distributes input to all models
record_click (in Cols)	Selecting a model record	add_new_model(event) — Adds a new model instance
submit (in kdb form)	Submitting knowledge base configuration	handle_kdb_setup(event) — Saves and applies configuration

Other Core Behaviors:

• show_input(params): Displays user input in the chat area
• show_added_model(m): Registers and displays a new model instance