feat: Add configurable system prompt to Condition Agent (#4587)

* feat: Add configurable system prompt to Condition Agent

* Update system prompt to HTML for UI readability

* fix: Remove invalid default routing and sync hardcoded role-based examples

* Update ConditionAgent.ts

* Update ConditionAgent.ts

---------

Co-authored-by: Henry Heng <henryheng@flowiseai.com>

packages/components/nodes/agentflow/ConditionAgent/ConditionAgent.ts

@@ -27,7 +27,7 @@ class ConditionAgent_Agentflow implements INode {
     constructor() {
         this.label = 'Condition Agent'
         this.name = 'conditionAgentAgentflow'
-        this.version = 1.0
+        this.version = 1.1
         this.type = 'ConditionAgent'
         this.category = 'Agent Flows'
         this.description = `Utilize an agent to split flows based on dynamic conditions`
@@ -80,6 +80,26 @@ class ConditionAgent_Agentflow implements INode {
                         scenario: ''
                     }
                 ]
+            },
+            {
+                label: 'Override System Prompt',
+                name: 'conditionAgentOverrideSystemPrompt',
+                type: 'boolean',
+                description: 'Override initial system prompt for Condition Agent',
+                optional: true
+            },
+            {
+                label: 'Node System Prompt',
+                name: 'conditionAgentSystemPrompt',
+                type: 'string',
+                rows: 4,
+                optional: true,
+                acceptVariable: true,
+                default: CONDITION_AGENT_SYSTEM_PROMPT,
+                description: 'Expert use only. Modifying this can significantly alter agent behavior. Leave default if unsure',
+                show: {
+                    conditionAgentOverrideSystemPrompt: true
+                }
             }
             /*{
                 label: 'Enable Memory',
@@ -242,6 +262,12 @@ class ConditionAgent_Agentflow implements INode {
             const conditionAgentInput = nodeData.inputs?.conditionAgentInput as string
             let input = conditionAgentInput || question
             const conditionAgentInstructions = nodeData.inputs?.conditionAgentInstructions as string
+            const conditionAgentSystemPrompt = nodeData.inputs?.conditionAgentSystemPrompt as string
+            const conditionAgentOverrideSystemPrompt = nodeData.inputs?.conditionAgentOverrideSystemPrompt as boolean
+            let systemPrompt = CONDITION_AGENT_SYSTEM_PROMPT
+            if (conditionAgentSystemPrompt && conditionAgentOverrideSystemPrompt) {
+                systemPrompt = conditionAgentSystemPrompt
+            }
 
             // Extract memory and configuration options
             const enableMemory = nodeData.inputs?.conditionAgentEnableMemory as boolean
@@ -277,31 +303,15 @@ class ConditionAgent_Agentflow implements INode {
             const messages: BaseMessageLike[] = [
                 {
                     role: 'system',
-                    content: CONDITION_AGENT_SYSTEM_PROMPT
+                    content: systemPrompt
                 },
                 {
                     role: 'user',
-                    content: `{"input": "Hello", "scenarios": ["user is asking about AI", "default"], "instruction": "Your task is to check and see if user is asking topic about AI"}`
+                    content: `{"input": "Hello", "scenarios": ["user is asking about AI", "user is not asking about AI"], "instruction": "Your task is to check if the user is asking about AI."}`
                 },
                 {
                     role: 'assistant',
-                    content: `\`\`\`json\n{"output": "default"}\n\`\`\``
+                    content: `\`\`\`json\n{"output": "user is not asking about AI"}\n\`\`\``
-                },
-                {
-                    role: 'user',
-                    content: `{"input": "What is AIGC?", "scenarios": ["user is asking about AI", "default"], "instruction": "Your task is to check and see if user is asking topic about AI"}`
-                },
-                {
-                    role: 'assistant',
-                    content: `\`\`\`json\n{"output": "user is asking about AI"}\n\`\`\``
-                },
-                {
-                    role: 'user',
-                    content: `{"input": "Can you explain deep learning?", "scenarios": ["user is interested in AI topics", "default"], "instruction": "Determine if the user is interested in learning about AI"}`
-                },
-                {
-                    role: 'assistant',
-                    content: `\`\`\`json\n{"output": "user is interested in AI topics"}\n\`\`\``
                 }
             ]
             // Use to store messages with image file references as we do not want to store the base64 data into database
@@ -374,15 +384,19 @@ class ConditionAgent_Agentflow implements INode {
                 )
             }
 
-            let calledOutputName = 'default'
+            let calledOutputName: string
             try {
                 const parsedResponse = this.parseJsonMarkdown(response.content as string)
-                if (!parsedResponse.output) {
+                if (!parsedResponse.output || typeof parsedResponse.output !== 'string') {
-                    throw new Error('Missing "output" key in response')
+                    throw new Error('LLM response is missing the "output" key or it is not a string.')
                 }
                 calledOutputName = parsedResponse.output
             } catch (error) {
-                console.warn(`Failed to parse LLM response: ${error}. Using default output.`)
+                throw new Error(
+                    `Failed to parse a valid scenario from the LLM's response. Please check if the model is capable of following JSON output instructions. Raw LLM Response: "${
+                        response.content as string
+                    }"`
+                )
             }
 
             // Clean up empty inputs

packages/components/nodes/agentflow/prompt.ts

@@ -39,37 +39,45 @@ export const DEFAULT_HUMAN_INPUT_DESCRIPTION_HTML = `<p>Summarize the conversati
 </ul>
 `
 
-export const CONDITION_AGENT_SYSTEM_PROMPT = `You are part of a multi-agent system designed to make agent coordination and execution easy. Your task is to analyze the given input and select one matching scenario from a provided set of scenarios. If none of the scenarios match the input, you should return "default."
+export const CONDITION_AGENT_SYSTEM_PROMPT = `
+<p>You are part of a multi-agent system designed to make agent coordination and execution easy. Your task is to analyze the given input and select one matching scenario from a provided set of scenarios.</p>
 
-- **Input**: A string representing the user's query or message.
+    <ul>
-- **Scenarios**: A list of predefined scenarios that relate to the input.
+        <li><strong>Input</strong>: A string representing the user's query, message or data.</li>
-- **Instruction**: Determine if the input fits any of the scenarios.
+        <li><strong>Scenarios</strong>: A list of predefined scenarios that relate to the input.</li>
+        <li><strong>Instruction</strong>: Determine which of the provided scenarios is the best fit for the input.</li>
+    </ul>
 
-## Steps
+    <h2>Steps</h2>
+    <ol>
+        <li><strong>Read the input string</strong> and the list of scenarios.</li>
+        <li><strong>Analyze the content of the input</strong> to identify its main topic or intention.</li>
+        <li><strong>Compare the input with each scenario</strong>: Evaluate how well the input's topic or intention aligns with each of the provided scenarios and select the one that is the best fit.</li>
+        <li><strong>Output the result</strong>: Return the selected scenario in the specified JSON format.</li>
+    </ol>
 
-1. **Read the input string** and the list of scenarios.
+    <h2>Output Format</h2>
-2. **Analyze the content of the input** to identify its main topic or intention.
+    <p>Output should be a JSON object that names the selected scenario, like this: <code>{"output": "<selected_scenario_name>"}</code>. No explanation is needed.</p>
-3. **Compare the input with each scenario**:
-   - If a scenario matches the main topic of the input, select that scenario.
-   - If no scenarios match, prepare to output "\`\`\`json\n{"output": "default"}\`\`\`"
-4. **Output the result**: If a match is found, return the corresponding scenario in JSON; otherwise, return "\`\`\`json\n{"output": "default"}\`\`\`"
 
-## Output Format
+    <h2>Examples</h2>
+    <ol>
+       <li>
+            <p><strong>Input</strong>: <code>{"input": "Hello", "scenarios": ["user is asking about AI", "user is not asking about AI"], "instruction": "Your task is to check if the user is asking about AI."}</code></p>
+            <p><strong>Output</strong>: <code>{"output": "user is not asking about AI"}</code></p>
+        </li>
+        <li>
+            <p><strong>Input</strong>: <code>{"input": "What is AIGC?", "scenarios": ["user is asking about AI", "user is asking about the weather"], "instruction": "Your task is to check and see if the user is asking a topic about AI."}</code></p>
+            <p><strong>Output</strong>: <code>{"output": "user is asking about AI"}</code></p>
+        </li>
+        <li>
+            <p><strong>Input</strong>: <code>{"input": "Can you explain deep learning?", "scenarios": ["user is interested in AI topics", "user wants to order food"], "instruction": "Determine if the user is interested in learning about AI."}</code></p>
+            <p><strong>Output</strong>: <code>{"output": "user is interested in AI topics"}</code></p>
+        </li>
+    </ol>
 
-Output should be a JSON object that either names the matching scenario or returns "\`\`\`json\n{"output": "default"}\`\`\`" if no scenarios match. No explanation is needed.
+    <h2>Note</h2>
-
+    <ul>
-## Examples
+        <li>Ensure that the input scenarios align well with potential user queries for accurate matching.</li>
-
+        <li>DO NOT include anything other than the JSON in your response.</li>
-1. **Input**: {"input": "Hello", "scenarios": ["user is asking about AI", "default"], "instruction": "Your task is to check and see if user is asking topic about AI"}  
+    </ul>
-   **Output**: "\`\`\`json\n{"output": "default"}\`\`\`"
-
-2. **Input**: {"input": "What is AIGC?", "scenarios": ["user is asking about AI", "default"], "instruction": "Your task is to check and see if user is asking topic about AI"}  
-   **Output**: "\`\`\`json\n{"output": "user is asking about AI"}\`\`\`"
-
-3. **Input**: {"input": "Can you explain deep learning?", "scenarios": ["user is interested in AI topics", "default"], "instruction": "Determine if the user is interested in learning about AI"}  
-   **Output**: "\`\`\`json\n{"output": "user is interested in AI topics"}\`\`\`"
-
-## Note
-- Ensure that the input scenarios align well with potential user queries for accurate matching
-- DO NOT include anything other than the JSON in your response.
 `