🍋 Zest (zst1)

The Tangiest 1-Line AI Inference Wrapper for Python.

Why write 20 lines of boilerplate when you can just squeeze a model into one? zest provides a unified interface for local Hugging Face pipelines and cloud-based OpenRouter LLMs.

🚀 Quick Start

import zest

# Local Hugging Face (Sentiment Analysis)
print(zest.run("hf/distilbert-base-uncased-finetuned-sst-2-english", "I love this!"))

# Cloud OpenRouter (GPT-4o)
# Requires os.environ['OPENROUTER_API_KEY']
print(zest.run("or/openai/gpt-4o", "Explain quantum physics in one sentence."))

🛠 Available Routes

1. Hugging Face (hf/ prefix)

Uses the transformers library to run models locally on your CPU/GPU.

• Usage: zest.run("hf/your-model-id", "input")
• Restricted Models: If you need to access gated models (like Llama), set your token:

  import os
  os.environ['HUGGINGFACE_TOKEN'] = 'your_token'
  # OR pass it directly:
  zest.run("hf/meta-llama/Llama-2-7b", "Hi", hf_token="your_token")
  

2. OpenRouter (or/ prefix)

Access 200+ LLMs via the OpenRouter API.

• Usage: zest.run("or/google/gemini-pro-1.5", "Hello world")
• Requirement: Set os.environ['OPENROUTER_API_KEY'].

🛡️ Error Handling (The 'Damn' Level)

zest doesn't just crash. It catches:

• Prefix Errors: Forgot hf/ or or/? It'll tell you.
• Auth Errors: Missing API keys for cloud or restricted models are clearly reported.
• Network Errors: Timeouts and HTTP errors from OpenRouter are intercepted.
• Model Errors: Invalid Hugging Face model IDs return specific config errors.

🧪 Snippets

Multi-turn Logic (Simplified):

model = "or/anthropic/claude-3-opus"
prompt = "Tell me a joke."
result = zest.run(model, prompt)
print(f"AI: {result}")

Batch Processing (Local):

inputs = ["Bad product", "Best ever! "]
results = [zest.run("hf/distilbert-base-uncased-finetuned-sst-2-english", x) for x in inputs]

Stay Tangy. 🍋