Shinkai Prompt Testing

Requirements

• Shinkai API running on http://localhost:9950 (or overwritten by env SHINKAI_API_URL)
• Ollama running on http://localhost:11434 (or overwritten by env OLLAMA_API_URL)
• Deno 2.x

Run Prompts & Execute Results

• deno task start test=benchmark-\*

Run specific tests:

• deno task start test=download-url-and-sql - Runs only the download-url-and-sql test
• deno task start test=download-url-and-sql test=download-url-and-summary - Runs the download-url-and-sql and download-url-and-summary tests
• Wildcards can be used: deno task start test=benchmark-\*

Notes on Prompt Results:

• Results will be stored in results/{language}/{model-name}/{test-code}/
• prompt- stores prompts
• raw- store raw responses
• src- store parsed response (valid Typescript or JSON)
• shinkai_local_tools.py stores the dynamic tools file used in the test
• shinkai-local-tools.ts stores the dynamic tools file used in the test
• execute-output stores the output of the executed code

Using models.txt

The models.txt file allows you to specify which models to use for testing. Each line in the file should specify a model in the format:

ollama:qwen2.5-coder:32b

• Engine Prefix: The prefix (ollama in the example) indicates which engine the model belongs to.
• Model Name: The rest of the line specifies the model name, which can include colons.

If models.txt is not present, the system will default to using models obtained from the getInstalledModels() function.