Multimodal Node

​

Typical Usage

​

Example Configuration

1. Connection: The node can be connected to a ‘Start’ node or any of the other nodes for input or output.
2. Configuration: The configuration panel allows for detailed setup of the model’s behavior.

​

Configuration Details

​

Core Settings

• Title: Give your node a descriptive name (e.g., “Analyze Receipt Image”).
• Description: Add a brief summary of what this node does for documentation purposes.

​

LLM Configuration

• LLM: Select the specific Large Language Model to power this node (e.g., GIDR LLM 2). The capabilities listed below are dynamically shown based on whether the selected LLM supports them.
• Reasoning level: Controls the depth of the model’s analysis. Options include Low (faster), Medium (balanced), and High (thorough). This option is only displayed when the selected LLM supports reasoning capabilities.
• Web Search: Retrieves information from the web by executing search queries and returning relevant, up-to-date results. This option is only displayed when the selected LLM supports tool-based web search.
• Get input by reference: Processes inputs via URL reference rather than direct value, enabling the model to handle large documents, videos, and audio files. This option is only displayed when the selected LLM supports reference-based input (e.g., Google Gemini).

​

Web Search Prompt Guidelines

You are a Multimodal Processing Node responsible for understanding user queries and producing accurate, grounded responses using the correct data source.

1. Understand the Query
Identify the user's intent and required information.
Decide whether the query needs real-time / latest / externally verifiable data or can be answered from pre-existing inputs and context.

2. Data Source & Tool Rules (STRICT)
If web_search is NOT available:
Do not call or simulate any tool.
Answer using pre-existing data only.
If the answer cannot be determined, explicitly state the limitation.
If web_search IS available:
Use it only when the query requires:
Real-time or time-sensitive information, or Latest updates, news, pricing, releases, or availability, or
Information not guaranteed in provided inputs.
Otherwise, do not use any tool.
🚫 Never call a tool that is not provided
🚫 Never use web_search if pre-existing data is sufficient

3. Multimodal Reasoning
Extract relevant information from all provided modalities.
Cross-check data when multiple inputs exist.
Prioritize explicit inputs over assumptions or inference.

4. Response Rules
Be clear, concise, and structured.
Ground responses only in:
web_search results (if used), or
Pre-existing data (if no tool is used).
Do not hallucinate or fabricate information.
Clearly state limitations when data is unavailable.

5. Tool Discipline
Use only one strategy per query:
Tool-based (web_search) OR
Non-tool (pre-existing data).
Never mix both.

​

Get Input by Reference

​

Prompts & Conversation History

• Prompt: The main instruction for the AI (e.g., “Please provide a detailed analysis of the input”). You can mix static text with dynamic variables.
• No. of previous exchanges: Controls how much conversation history (context) is passed to the model. ‘0’ means no history (stateless).

​

Advanced Options

• Skip if no image: Automatically bypasses this node if the input does not contain image data. This is useful for building workflows that can gracefully handle both text-only and multimodal inputs without error.
• Allow conditional input: Enables logic to conditionally trigger this node based on input criteria.
• Variable Selectors:
  • Input variable selector: Map specific input variables to the node.
  • Prompt variable selector: Inject variables (e.g., user name, date) directly into your prompt.