Skip to main content

Access Model Playground

Overview​

The Model Playground is an interactive interface in Armada Bridge that lets you test deployed models, run predictions, and validate model behavior before integrating them into production applications.

Use the playground to:

  • Send prompts and inspect model responses in real time
  • Verify that a deployed model behaves as expected
  • Iterate on prompts and parameters without writing code
  • Share the playground URL with your team for quick testing (within tenant isolation)

Prerequisites​

Before you can use the Model Playground, ensure that:

  • A model has been deployed (as a Tenant Admin or by a user with deployment permissions)
  • You have a tenant user account (created by a Tenant Admin)
  • You are logged in as that tenant user

Accessing the Model Playground​

Step 1: Open AI Studio and Models​

  1. Log in to Armada Bridge as a tenant user.

  2. In the left sidebar, click AI Studio.

    AI Studio

  3. Under Models, locate the model you want to test.

  4. Click Open in Playground for that model.

    Model Playground

Step 2: Open the Playground​

Click Open in Playground. The playground opens in your browser.

Browser security warning (when the endpoint has no authorized certificate)

A security warning is shown only when the model endpoint does not have an authorized certificate (for example, when it uses a self-signed or untrusted certificate). If you see this warning, you can proceed as follows:

  1. Copy the model endpoint URL from the Playground page.
  2. Open a new browser tab and paste that URL.
  3. When the browser shows a security warning, click Advanced (or equivalent in your browser).
  4. Choose Proceed to [site] (unsafe) or Accept the risk and continue to load the endpoint page once.
  5. You may see a "404 page not found" message which is expected.
  6. Return to the Playground tab and refresh the page if needed. You can then run queries in the prompting interface.

This one-time step establishes trust for the endpoint in your browser. You do not need to repeat it for the same endpoint.

Playground CURL

Step 3: Access the Playground​

Return to the Playground tab. You can now run queries in the prompting interface, as shown below.

Playground Prompt

Tenant user isolation

Tenant user isolation is enforced in Armada Bridge. One tenant user cannot see or use another tenant user's Model Playground. Each tenant user (for example, tenantuser1 and tenantuser2) must complete the steps above with their own login to access their own model playground.

Using the Playground​

In the prompting interface you can:

  • Enter text prompts and submit them to the deployed model
  • View the model’s response in the same screen
  • Adjust parameters (if supported by the model) and re-run tests

Use these results to validate behavior before calling the model from applications or APIs.

Playground Interface​

The playground provides:

  • Input Editor - Define model inputs
  • Prediction Viewer - View model outputs
  • History - Track prediction calls
  • Export - Save test results

Playground Interface

Test Models​

Single Prediction​

Make a single prediction:

  1. Enter input data in the input editor
  2. Click Predict or Submit
  3. View model output
{
  "inputs": {
    "feature1": 1.0,
    "feature2": 2.0,
    "feature3": 3.0
  }
}

Single Prediction

Output:

{
  "prediction": 0.95,
  "confidence": 0.98
}

Batch Predictions​

Test multiple inputs:

  1. Prepare CSV or JSON file with multiple inputs
  2. Upload file to playground
  3. Submit batch request
  4. Download results
feature1,feature2,feature3
1.0,2.0,3.0
2.0,3.0,4.0
3.0,4.0,5.0

Batch Predictions

Input Formats​

JSON Format​

{
  "instances": [
    [1.0, 2.0, 3.0],
    [2.0, 3.0, 4.0]
  ]
}

CSV Format​

col1,col2,col3
1.0,2.0,3.0
2.0,3.0,4.0

Text Format​

Raw text input for text models

Validate Models​

Check Predictions​

Verify model behaves correctly:

  • Test with known inputs
  • Verify expected outputs
  • Check confidence scores
  • Test edge cases

Validate Predictions

Test Different Inputs​

Explore model behavior:

{
  "test_cases": [
    {"normal_input": [1.0, 2.0, 3.0]},
    {"edge_case": [0.0, 0.0, 0.0]},
    {"extreme": [100.0, 100.0, 100.0]}
  ]
}

Analyze Predictions​

Review results:

  • Prediction values
  • Confidence scores
  • Inference time
  • Resource usage

Export Results​

Download Predictions​

Save test results:

  1. Run predictions
  2. Click Download Results
  3. Choose format (JSON, CSV, Excel)

Export Results

Share Results​

Share findings with team:

  1. Export results
  2. Include in reports
  3. Attach to documentation
  4. Reference in model validation

Performance Testing​

Latency Testing​

Measure model response time:

Request: 1.5 seconds
Average: 1.2 seconds (from 10 requests)

Latency Testing

Throughput Testing​

Test throughput capacity:

Requests per second: 100 RPS
Concurrent connections: 10

Resource Usage​

Monitor resource consumption:

  • CPU usage
  • Memory usage
  • GPU usage
  • Network bandwidth

Resource Monitoring

Debugging​

View Model Logs​

Check model serving logs:

[2024-01-09 10:15:32] Model loaded successfully
[2024-01-09 10:15:45] Prediction request: {input: [1,2,3]}
[2024-01-09 10:15:46] Inference complete: 0.95

Error Handling​

Debug prediction errors:

  1. Check input format
  2. Verify input data types
  3. Check model requirements
  4. Review error messages

Error Debugging

A/B Testing​

Compare Models​

Test multiple model versions:

  1. Deploy Model A and Model B
  2. Send same inputs to both
  3. Compare outputs
  4. Analyze differences

A/B Testing

Model Comparison​

{
  "model_a_output": 0.95,
  "model_b_output": 0.92,
  "difference": 0.03
}

Best Practices​

Before Production​

  • Test with diverse inputs
  • Validate edge cases
  • Check performance metrics
  • Review resource usage
  • Document test results

Continuous Validation​

  • Regular model testing
  • Monitor prediction distribution
  • Track confidence scores
  • Detect data drift
  • Compare against baselines

Next Steps​

Last updated on