LLM

Attribute	Value
Node name	LLM
Version	v1.0
Supported scenarios	Chatflow Agent ✓, Workflow Agent ✓
Test run support	✓

Quick Reference

Configuration Item	Required	Description
Model selection	✓	Select from the models configured in service settings
System Prompt	✓	Sets the AI's role and behavioral guidelines
User Prompt	✓	The specific task instruction
Memory	-	Conversation history memory
Structured output	-	Output in JSON Schema format

Feature Description

The LLM node is the core node of the LongbridgeAI Agent Platform, providing powerful AI capabilities:

• Text processing: natural language understanding, reasoning, and generation
• Memory management: maintains conversation context and history
• Structured output: supports structured data output in JSON format

Node Display Content

1. Basic Information Display

• Icon: node-specific icon
• Name: node name
• Model in use:
  • Model provider icon
  • Model name
• Note: explanatory text added by the user

2. Action Buttons

• More button: click to display the context menu content
• Run button: click to enter the test run interface

Node Edit Panel

1. Model Configuration

• Model selection: select from the models configured in service settings; model parameters can be configured
• Parameter configuration: some input parameters can be adjusted; the adjustable parameters are subject to what the configuration panel actually displays
• Model management: there is no model management in this release; Longbridge initializes it internally and provides test model accounts

a. Selecting a Single Model

The model dropdown list is grouped by vendor; you can select a specific model under a vendor (as shown below, qwen-flash under the vendor alibaba), or search for a model name directly via the search box at the top:

b. SOTA Label

Some model names carry a SOTA label, which means the model is more expensive but also better—more capable and consuming more of your usage quota. Choose as needed:

c. Model Combos

Switch to the Model Combos tab to select one of the platform's built-in model combos (such as the Google series, Qwen series, etc.):

• A combo uses the models in order: when the first model is unavailable, it automatically switches to the second, and so on
• Suitable for scenarios with high stability requirements, avoiding flow interruptions caused by a single model failure
• Each combo is labeled with its applicable scenarios (multimodal, Chinese, enterprise, Agent, etc.), so you can pick as needed

2. Context Configuration

• Variable selection: click the variable value to bring up the Select Variable Value feature
• Error check: if the context variable is not filled into the prompt, the error "Please fill the context variable into the prompt" is shown

3. System Prompt

• Input box: same input features as the Answer template
• Limitation: Jinja templates are not supported
• Deletion: cannot be deleted

4. User Prompt / Assistant Prompt

• Input box: same input features as the Answer template
• Deletion: prompts can be deleted
• Add message: after clicking the add message button, new User/Assistant prompt editors appear in turn

• Assistant Prompt: the Assistant Prompt is the AI assistant's output, and it can also serve as context input—as shown above, the Assistant message carries the Start/context variable (with the "context" label), feeding the historical context back to the model
• Order requirement: the last prompt must be a User Prompt (usually carrying Start/sys.query, i.e., the user's input for this turn)

5. Memory

Memory is divided into two independent switches: long-term memory and short-term memory:

a. Long-Term Memory

• When enabled, the LLM automatically extracts content/preferences you explicitly ask it to remember from the conversation (such as your preferred investment style)
• In subsequent conversations, the saved long-term memory is provided to the LLM as context, making answers more personalized

b. Short-Term Memory

• When enabled, short-term memory provides the current session's content to the LLM as context based on the memory window turn count you set, so answers stay coherent with the preceding conversation
• Short-term memory only applies to the current session; it no longer takes effect after switching sessions
• The memory window supports dragging the slider or entering the number of turns directly (set to 10 in the image above)

6. Output Control Switches

a. Output Legality and Compliance Check

• When enabled, the system performs a compliance check on the content as it is output
• If the output content is found to be illegal or non-compliant with financial regulations, the system will immediately stop the output and clear the content already generated

b. Citations / Stock Citations / Widget (Beta)

• These three features will be covered in documentation for future releases

c. Line-by-Line Output

• When enabled, streaming output is delivered line by line (a full line is buffered before display)
• When disabled, output uses a character-by-character typewriter effect

7. Output Variables

• Default output: displays the output variable name "text", data type "String", and description "generated content"
• think: String type, the model's reasoning process. Some models support think output, which can be used with the "Reasoning Process Output" node output (Chatflow only)
• Structured output: the structured output switch
• Configuration page: after turning on the structured output switch, the structured output content "structured_output" of type object is displayed
  • Displays the configured fields, including parameter name, parameter type, and whether the field is required
  • Click the configure button to enter the configuration page

8. Structured Output Configuration

• Specify output variables: output variables can be specified—as shown above, the field aa (string type, required, described as bb) is defined under structured_output (object type)
• Default template: the configuration page displays a default template
• Editor: the configuration page only provides a JSON editor window, a clear-configuration action, and a confirm-save button
• Reset: clicking clear configuration restores the default template

⚠️ Note: structured output is not supported by all models; some older models do not support it. If output is abnormal after enabling it, first confirm whether the selected model supports structured output.

9. Next Step Configuration

• Adding a next node is supported

Execution Logic

The execution logic of the LLM node is implemented by the backend to ensure the large language model is called and processed correctly.

Context Menu Actions

The context menu contains the following options:

1. Change node: change the node type
2. Copy: copy the node
3. Duplicate: duplicate the node content
4. Delete: delete the node
5. Help link: jump to the help documentation
6. Run node: test-run the current node

Use Cases

1. Text Generation

• Content creation
• Article writing
• Creative copywriting

2. Q&A Systems

• Knowledge Q&A
• Technical support
• Customer service conversations

3. Text Processing

• Text summarization
• Translation services
• Text classification

4. Code Generation

• Code writing
• Code explanation
• Code optimization

Best Practices

1. Prompt Design

• Clear and specific: use clear, specific instructions
• Complete context: provide sufficient context information
• Guide with examples: use examples to guide model behavior

2. Memory Management

• Sensible settings: set the memory window based on task needs
• Content optimization: optimize the quality of memory content
• Performance balance: balance memory effectiveness against performance

3. Structured Output

• Schema design: design a sensible output structure
• Type definitions: define field types explicitly
• Required fields: set required fields appropriately

Configuration Steps

1. Basic Configuration

1. Select an appropriate large language model
2. Configure the System Prompt
3. Set the User Prompt

2. Advanced Configuration

1. Configure context variables
2. Set up memory
3. Configure structured output (optional)

3. Testing and Validation

1. Test using the test run feature
2. Check the output results
3. Adjust configuration parameters

Considerations

1. Model selection: choose an appropriate model based on the characteristics of the task
2. Prompt quality: high-quality prompts significantly improve results
3. Context management: manage context length sensibly
4. Cost control: watch token usage and cost
5. Error handling: consider how to handle exceptional cases

FAQ

Q: How do I choose an appropriate model?

A: Choose based on task complexity, response speed requirements, and cost budget.

Q: What is the difference between the System Prompt and the User Prompt?

A: The System Prompt sets the AI's role and behavioral guidelines; the User Prompt is the specific task instruction.

Q: How do I configure structured output?

A: Turn on the structured output switch, click the configure button, and define the output structure using JSON Schema.

Q: How do I use the memory feature?

A: Turn on the memory switch and set the memory template and memory window size; the AI will remember the conversation history.

Q: How do I improve prompt effectiveness?

A: Use clear instructions, provide examples, and set an appropriate role and context.