Dynamic LLM Switching and Self-Correction for AI Agents - n8n Workflow

• AI Engineers building self-correcting systems.


• Developers needing reliable fallback mechanisms in their n8n workflow.


• Users looking for advanced examples of the Langchain Code n8n node and dynamic connections.


• Organizations running critical automation where LLM downtime is unacceptable.

This n8n workflow addresses the challenge of building robust AI agents by implementing a model cascading mechanism. Traditional workflows often hardcode a single Large Language Model (LLM). If that LLM fails, rate-limits, or provides a poor quality response, the entire n8n automation stalls. This innovative n8n templates solution utilizes the Switch Model n8n node (a Langchain Code node) to select from a pool of available OpenAI LLMs based on an index.

If the initial attempt fails or is flagged as inadequate by a separate validation step, the system increases the index and loops back, automatically trying the next LLM in the sequence. This ensures that your AI agent maximizes reliability and response quality within the defined n8n process, demonstrating a powerful capability of the n8n platform.

1. Trigger: The n8n workflow starts with the When chat message received n8n trigger, capturing the user's input (e.g., a customer complaint).


2. Index Initialization: The Set LLM index n8n node initializes the selection index, determining which LLM to try first (defaulting to index 0).


3. Dynamic Selection: The Switch Model Langchain Code n8n node dynamically reads the list of connected LLMs (e.g., OpenAI 4o-mini, OpenAI 4o, etc.) and selects one based on the current llm_index. This is a sophisticated application of a custom n8n node.


4. Response Generation: The Generate response n8n node uses the selected LLM along with a fixed prompt to generate a customer support reply. This Langchain Chain LLM n8n node is configured to continue on error, allowing the n8n workflow to handle failures gracefully.


5. Validation: The Validate response n8n node (Sentiment Analysis) assesses the quality of the generated text against predetermined criteria (polite tone, clear resolution) using a separate, fixed LLM (the OpenAI Chat Model).


6. Fallback/Iteration Path: If the generation fails, the Check for expected error n8n node determines if the failure was a known condition (like index exhaustion). If the system is iterating due to poor response quality or a model error, the Increase LLM index n8n node increments the index, and the flow uses the No Operation, do nothing n8n node to loop back to the index setting step, initiating a new run with the next available model in this self-healing n8n workflow.

1. Import: Import this dynamic n8n workflow JSON into your n8n instance using the Templates feature or the 'New workflow' > 'Import from JSON' option. This n8n templates setup requires the @n8n/n8n-nodes-langchain community package to be installed.


2. Credentials: You must provide valid OpenAI API Key credentials for all language model nodes. Locate the following n8n node setups:

1. LLM Configuration: Ensure the LLM nodes connected to the Switch Model are configured in the desired order, as the llm_index depends on this arrangement within the n8n node connection structure.


2. Activate: Save and activate the n8n workflow trigger to begin using this agent.

When chat message received (Langchain Chat Trigger n8n node): The starting n8n trigger for the agent, listening for incoming user chat inputs.
Set LLM index (Set n8n node): Initializes the index used for model selection. Key Configuration: llmindex set to ={{ $json.llmindex || 0 }}.
Switch Model (Langchain Code n8n node): The core dynamic routing logic. It selects the specific LLM n8n node based on the incoming llmindex. Key Configuration: Custom JavaScript code fetching and selecting the input LLM by index, ensuring a robust n8n workflow.
OpenAI 4o-mini, OpenAI 4o, OpenAI o1 (Langchain LLM Chat n8n node): The pool of Large Language Models available for the agent, feeding into the dynamic switching logic.
Generate response (Langchain Chain LLM n8n node): Generates the agent's response using the dynamically selected LLM. Key Configuration: onError set to continueErrorOutput for resilience.
Validate response (Langchain Sentiment Analysis n8n node): Evaluates the quality of the generated response using a specific prompt template, crucial for the self-correction mechanism in this n8n workflow.
Increase LLM index (Set n8n node): Used in the self-correction loop to prepare for the next iteration by incrementing the llmindex.
Check for expected error (If n8n node): Handles flow control based on whether a specific error occurred in the preceding generation step, facilitating robust error handling within the n8n workflow.