Standard Apps

Gemma-powered applications using standard LLM calls.

Standard Architecture

flowchart TD
    A[User Input] --> B[App Logic]
    B --> C[Gemma API Call]
    C --> D[Process Response]
    D --> E[Display Output]

    style C fill:#fff3e0
    style E fill:#c8e6c9

Demo Scripts

Script	Purpose	Speed
llm_quick_demo_base.py	Basic Qwen demo	Fast
llm_mini_demo_5cases.py	Quick classification	Fast
llm_stress_test_class.py	Load testing	Medium
llm_hierarchical_demo.py	Multi-level routing	Medium

Code Pattern

import requests

def call_gemma(prompt, model="gemma-4b-e4b"):
    url = "http://localhost:1234/v1/chat/completions"
    payload = {
        "model": model,
        "messages": [{"role": "user", "content": prompt}],
        "temperature": 0.7
    }
    response = requests.post(url, json=payload)
    return response.json()["choices"][0]["message"]["content"]

Use Cases

graph TD
    A[Gemma Applications] --> B[Classification]
    A --> C[Summarization]
    A --> D[Entity Extraction]
    A --> E[Routing]

    B --> B1[Ticket routing]
    C --> C1[Email summaries]
    D --> D1[Name/Date extraction]
    E --> E1[Auto-assign]

    style B fill:#e3f2fd
    style C fill:#e3f2fd
    style D fill:#e3f2fd
    style E fill:#e3f2fd

Performance Characteristics

Aspect	Value
Latency	1-3 seconds
Quality	High (4-bit optimized)
Context	8K tokens
Accuracy	90%+ on classification

Comparison: Qwen vs Gemma

Task	Qwen 1.5B	Gemma 4B E4B
Simple classification	✅ Fast	✅ Good
Complex reasoning	❌ Limited	✅ Excellent
Security analysis	❌ Poor	✅ Great
Long context	❌ 2K	✅ 8K

Best Practices

Tip	Description
Batch requests	Process multiple at once
Cache responses	Avoid repeated calls
Use streaming	Better UX for long outputs
Monitor latency	Track performance

Next Steps

• SLM Apps - Smaller, faster models
• MLOps - Production deployment