WorkflowsCommon Architectures

Fallback Models

Fallback Models

When building production LLM applications, it’s often beneficial to implement a fallback strategy that can dynamically select different models based on various conditions. This approach allows you to:

  1. Optimize for cost: Start with a cheaper, faster model and only use more expensive models when necessary
  2. Handle context limitations: Automatically switch to models with larger context windows when needed
  3. Implement error recovery: Gracefully handle model-specific errors by trying alternative models
  4. Balance performance and cost: Use the most appropriate model for each specific task

This tutorial demonstrates how to build a workflow that dynamically selects models and implements fallback logic when errors occur.

Implementation Overview

We’ll create a workflow that:

  1. Uses a Templating Node to dynamically select which model to use
  2. Connects the Templating Node output to a Prompt Node’s model selection
  3. Implements error handling with a Try adornment
  4. Creates a feedback loop to try a different model if the first attempt fails
1

Add a Prompt Node and a Templating Node

Start by adding a Prompt Node and a Templating Node to your workflow sandbox. The Templating Node will determine which model we should use in the Prompt Node.

Adding a Prompt Node and a Templating Node to the workflow
2

Configure Dynamic Model Selection

Open the “Model” tab in the Prompt Node, then click the 3-dot icon next to the currently selected model. You’ll see an option for “Expressions” - this allows you to dynamically select a model. Click it to enable dynamic model selection.

Enabling dynamic model selection via Expressions
3

Connect the Templating Node to the Model Selection

Connect the Templating Node’s output to the Prompt Node’s model field. This will allow the Templating Node to dynamically determine which model the Prompt Node should use.

Connecting the Templating Node to the Prompt Node's model field
4

Add a Try Adornment to the Prompt Node

Click anywhere on the Prompt Node to open the side panel. Navigate to the Adornments section and add a “Try” adornment. This will help us expose a new output from the Prompt Node that will let us detect errors and handle them appropriately.

Adding a Try adornment to the Prompt Node
5

Configure Conditional Routing

Navigate to the Ports tab and add a new port. Create an expression for the “If” case that checks if there was no error. If there was an error, we’ll take the “Else” path and route back to the Templating Node to select a different model.

Configuring conditional routing based on error detection
6

Connect the Nodes

Your canvas should now look something like this, with the Prompt Node having two output paths - one for successful execution and one for error handling:

Workflow with conditional paths set up
7

Complete the Connections

Finish setting up your workflow by:

  • Connecting the Templating Node output to the Prompt Node input
  • Connecting the Entrypoint to the input of the Templating Node
  • Renaming the Templating Node to “Model Selection”
Completed workflow connections
8

Implement the Model Selection Logic

Open the Templating Node and rename the input variable to times_executed. Connect it to the Execution Count of the Prompt Node by selecting Execution Count: Prompt.

Then, paste the following Jinja template into the Template area:

{%- if times_executed == 0 -%}
gpt-3.5-turbo
{%- else -%}
gpt-4o
{%- endif -%}

This logic says:

  • If our Prompt Node hasn’t run yet (execution count is 0), use gpt-3.5-turbo (a model with a smaller context window)
  • If we have executed it once before (meaning there was likely an error), use gpt-4o which has a larger context window
Configuring the model selection logic in the Templating Node
9

Test the Workflow

To test the fallback functionality:

  1. Go back to your Prompt Node and delete any unused input variables
  2. Try pasting in a large amount of text (>16k tokens) that would exceed the context window of gpt-3.5-turbo
  3. Make sure your Final Output Node has its output value connected to the Prompt Node’s output
  4. Run the workflow
10

Verify the Fallback Behavior

Use the Workflow Scrubber at the bottom of the page to validate that the workflow did in fact re-run once, switching from gpt-3.5-turbo to gpt-4o after encountering an error due to the context size limitation.

Workflow execution showing the fallback behavior - first attempt
Workflow execution showing the fallback behavior - second attempt

Additional Use Cases

This fallback model architecture can be extended for various scenarios:

Cost Optimization

Start with cheaper models for simple tasks, and only escalate to more expensive models when necessary.

Error Recovery

Handle various types of model-specific errors (rate limits, timeouts, etc.) by trying alternative models.

Progressive Enhancement

Begin with faster models for quick responses, then use more capable models only when the initial response doesn’t meet quality thresholds.

Context Window Management

Automatically switch to models with larger context windows when content exceeds the limits of smaller models.

Customization Options

You can customize this architecture in several ways:

  1. Multiple Fallback Levels: Add more conditions to try several models in sequence
  2. Error-Specific Handling: Check the specific error type and choose different fallback models accordingly
  3. Quality-Based Routing: Add a Guardrail Node to evaluate the quality of responses and trigger fallbacks for low-quality outputs
  4. Timeout Management: Implement timeouts to fall back to faster models when response time is critical

By implementing dynamic model selection with fallback logic, you can build more resilient, cost-effective, and performant LLM applications that gracefully handle errors and optimize for different requirements.