Spice.ai Demo App

This is a Spice.ai data and AI app.

Prerequisites

• Spice.ai CLI installed
• OpenAI API key
• Hugging Face API token (optional, for LLaMA model)
• curl and jq for API calls

Learn More

To learn more about Spice.ai, take a look at the following resources:

• Spice.ai - learn about Spice.ai features, data, and API.
• Get started with Spice.ai - try out the API and make basic queries.

Connect with us on Discord - your feedback is appreciated!

---

Demo Steps

Publishing a Spice App in the Cloud

Step 1: Forking and Using the Dataset

1. Fork the repository https://github.com/jeadie/evals into your GitHub org.

Step 2: Creating a New App in the Cloud

1. Log into the Spice.ai Cloud Platform and create a new app called evals. The app will start empty.
2. Connect the app to your repository:
   • Go to the App Settings tab and select Connect Repository.
   • If the repository is not yet linked, follow the prompts to authenticate and link it.

Step 3: Deploying the App

1. Set the app to Public:
   • Navigate to the app's settings and toggle the visibility to public.
2. Redeploy the app:
   • Click Redeploy to load the datasets and configurations from the repository.

Step 4: Verifying and Testing

1. Check the datasets in the Spice.ai Cloud:
   • Verify that the datasets are correctly loaded and accessible.
2. Test public access:
   • Log in with a different account to confirm the app is accessible to external users.

---

Initializing a Local Spice App

1. Initialize a new local Spice app

   mkdir demo
   cd demo
   spice init
   

2. Login to Spice.ai Cloud

   spice login
   

3. Get spicepod from Spicerack Navigate to spicerack.org, search for evals.

Click on /evals, click on Use this app, and copy the spice connect command.

Paste the command into the terminal. Navigate to spicerack.org, search for evals, click on /evals, click on Use this app, and copy the spice connect command. Paste the command into the terminal.

spice connect <username>/evals

The spicepod.yml should be updated to:

version: v1beta1
kind: Spicepod
name: demo

dependencies:
  - Jeadie/evals

5. Add a model to the spicepod

   models:
     - name: gpt-4o
       from: openai:gpt-4o
       params:
         openai_api_key: ${ secrets:SPICE_OPENAI_API_KEY }
   

6. Start spice

   spice run
   

7. Run an eval

   curl -XPOST "http://localhost:8090/v1/evals/taxes"      -H "Content-Type: application/json"      -d '{
       "model": "gpt-4o"
     }' | jq
   

8. Explore incorrect results

   spice sql
   

   SELECT
     input,
     output,
     actual
   FROM eval.results
   WHERE value=0.0 LIMIT 5;
   

---

Optional: Create an Eval to Use a Smaller Model

1. Track the outputs of all AI model calls:

   runtime:
     task_history:
       captured_output: truncated
   

2. Define a new view and evaluation:

   views:
     - name: user_queries
       sql: |
         SELECT
           json_get_json(input, 'messages') AS input,
           json_get_str((captured_output -> 0), 'content') as ideal
         FROM runtime.task_history
         WHERE task='ai_completion'
     - name: latest_eval_runs
       sql: |
         SELECT model, MAX(created_at) as latest_run
            FROM eval.runs
            GROUP BY model
     - name: model_stats
       sql: |
         SELECT
           r.model,
           COUNT(*) as total_queries,
           SUM(CASE WHEN res.value = 1.0 THEN 1 ELSE 0 END) as correct_answers,
           AVG(res.value) as accuracy
         FROM eval.runs r
         JOIN latest_eval_runs lr ON r.model = lr.model AND r.created_at = lr.latest_run
         JOIN eval.results res ON res.run_id = r.id
         GROUP BY r.model
   
   evals:
     - name: mimic-user-queries
       description: |
         Evaluates how well a model can copy the exact answers already returned to a user. Useful for testing if a smaller/cheaper model is sufficient.
       dataset: user_queries
       scorers:
         - match
   

3. Add a smaller model to the spicepod:

   models:
     - name: llama3
       from: huggingface:huggingface.co/meta-llama/Llama-3.2-3B-Instruct
       params:
         hf_token: ${ secrets:SPICE_HUGGINGFACE_API_KEY }
   
     - name: gpt-4o # Keep previous model.
   

4. Verify models are loaded:

   spice models
   

   You should see both models listed:

   NAME    FROM                                                         STATUS
   gpt-4o  openai:gpt-4o                                                ready
   llama3  huggingface:huggingface.co/meta-llama/Llama-3.3-70B-Instruct ready
   

5. Restart the Spice app:

   spice run
   

6. Test the larger model or run another eval:

   spice chat
   

7. Run evaluations on both models:

   # Run eval with GPT-4
   curl -XPOST "http://localhost:8090/v1/evals/mimic-user-queries" \
     -H "Content-Type: application/json" \
     -d '{"model": "gpt-4o"}' | jq
   
   # Run eval with LLaMA
   curl -XPOST "http://localhost:8090/v1/evals/mimic-user-queries" \
     -H "Content-Type: application/json" \
     -d '{"model": "llama3"}' | jq
   

8. Compare model performance:

   spice sql
   

   SELECT
     model,
     total_queries,
     correct_answers,
     ROUND(accuracy * 100, 2) as accuracy_percentage
   FROM model_stats
   ORDER BY accuracy_percentage DESC;
   

   This query will show:

   • Total number of queries processed
   • Number of correct answers
   • Accuracy percentage as a percentage

   You can use these metrics to decide if the smaller model provides acceptable performance for your use case.

---

Full Spicepod Configuration

Include the following spicepod.yml for reference:

version: v1beta1
kind: Spicepod
name: demo

dependencies:
  - Jeadie/evals

runtime:
  task_history:
    captured_output: truncated

views:
  - name: user_queries
    sql: |
      SELECT
        json_get_json(input, 'messages') AS input,
        json_get_str((captured_output -> 0), 'content') as ideal
      FROM runtime.task_history
      WHERE task='ai_completion'

evals:
  - name: mimic-user-queries
    description: |
      Evaluates how well a model can copy the exact answers already returned to a user. Useful for testing if a smaller/cheaper model is sufficient.
    dataset: user_queries
    scorers:
      - match

models:
  - name: gpt-4o
    from: openai:gpt-4o
    params:
      openai_api_key: ${ secrets:SPICE_OPENAI_API_KEY }

  - name: llama3
    from: huggingface:huggingface.co/meta-llama/Llama-3.2-3B-Instruct
    params:
      hf_token: ${ secrets:SPICE_HUGGINGFACE_API_KEY }