Helix v1

Helix v1 is Rustellar's flagship conversational AI model, designed for general-purpose dialogue and natural language understanding.

Overview

Helix v1 is built on the proven Transformer Decoder-only architecture, making it exceptionally capable at understanding context and generating coherent, natural responses across a wide range of conversational scenarios.

Key Features

General Conversation Excellence

Helix v1 excels at:

• Natural dialogue and chat applications
• Question answering systems
• Customer support automation
• Personal assistants
• Educational tutoring

Technical Specifications

Specification	Value
Architecture	Transformer Decoder-only
Context Window	8,192 tokens
Maximum Output	4,096 tokens
Languages	95+ languages
Training Cutoff	January 2025

Use Cases

Customer Support Chatbot

import requests

# カスタマーサポート用の設定
response = requests.post(
    "https://api.rustellar.com/v1/chat/completions",
    headers={
        "Authorization": "Bearer YOUR_API_KEY",
        "Content-Type": "application/json"
    },
    json={
        "model": "helix-v1",
        "messages": [
            {
                "role": "system",
                "content": "You are a helpful customer support agent. "
                          "Be polite, professional, and provide clear solutions."
            },
            {
                "role": "user",
                "content": "I can't log in to my account. What should I do?"
            }
        ],
        "temperature": 0.7,  # バランスの取れた応答
        "max_tokens": 500
    }
)

print(response.json()['choices'][0]['message']['content'])

Educational Q&A

# 教育用Q&Aシステムの例
response = requests.post(
    "https://api.rustellar.com/v1/chat/completions",
    headers={
        "Authorization": "Bearer YOUR_API_KEY",
        "Content-Type": "application/json"
    },
    json={
        "model": "helix-v1",
        "messages": [
            {
                "role": "system",
                "content": "You are an educational tutor. "
                          "Explain concepts clearly with examples."
            },
            {
                "role": "user",
                "content": "Explain photosynthesis in simple terms."
            }
        ],
        "temperature": 0.5,  # より一貫した説明のため低めに設定
        "max_tokens": 800
    }
)

print(response.json()['choices'][0]['message']['content'])

Code Assistant

# プログラミングアシスタントの例
response = requests.post(
    "https://api.rustellar.com/v1/chat/completions",
    headers={
        "Authorization": "Bearer YOUR_API_KEY",
        "Content-Type": "application/json"
    },
    json={
        "model": "helix-v1",
        "messages": [
            {
                "role": "system",
                "content": "You are an expert programming assistant. "
                          "Provide clear code examples with explanations."
            },
            {
                "role": "user",
                "content": "How do I read a CSV file in Python?"
            }
        ],
        "temperature": 0.3,  # 正確なコード生成のため低めに設定
        "max_tokens": 1000
    }
)

print(response.json()['choices'][0]['message']['content'])

Performance Characteristics

Response Speed

Helix v1 is optimized for fast response times, typically generating completions in:

• Short responses (< 100 tokens): 0.5-1 second
• Medium responses (100-500 tokens): 1-3 seconds
• Long responses (500+ tokens): 3-8 seconds

Accuracy

Helix v1 demonstrates high accuracy in:

• ✅ Factual question answering
• ✅ Task completion and instructions following
• ✅ Multi-turn conversation coherence
• ✅ Context understanding and retention

Best Practices

Temperature Settings

Choose temperature based on your use case:

# 事実に基づく応答（低温度）
# 推奨: 0.1 - 0.5
{
    "temperature": 0.3,
    "use_cases": ["Q&A", "Code generation", "Technical support"]
}

# バランスの取れた応答（中温度）
# 推奨: 0.6 - 0.8
{
    "temperature": 0.7,
    "use_cases": ["Customer service", "General chat", "Tutoring"]
}

# 創造的な応答（高温度）
# 推奨: 0.9 - 1.2
{
    "temperature": 1.0,
    "use_cases": ["Creative writing", "Brainstorming", "Ideation"]
}

System Message Guidelines

Craft effective system messages:

# ❌ 悪い例: 曖昧で具体性がない
"You are helpful."

# ✅ 良い例: 明確で具体的
"You are a professional customer service agent for TechCorp. "
"Always be polite, empathetic, and solution-focused. "
"If you don't know the answer, offer to escalate to a human agent."

Context Management

For long conversations, manage context effectively:

def manage_conversation_context(messages, max_tokens=6000):
    """
    会話コンテキストを管理する関数
    最大トークン数を超えないように古いメッセージを削除
    """
    # トークン数を概算（実際にはトークナイザーを使用）
    estimated_tokens = sum(len(m['content']) // 4 for m in messages)

    if estimated_tokens > max_tokens:
        # システムメッセージは保持
        system_msgs = [m for m in messages if m['role'] == 'system']
        # 最新のユーザー/アシスタントメッセージのみ保持
        recent_msgs = messages[-20:]  # 最新20メッセージ

        return system_msgs + [m for m in recent_msgs if m['role'] != 'system']

    return messages

# 使用例
conversation_history = [
    {"role": "system", "content": "You are a helpful assistant."},
    # ... 多数のメッセージ ...
]

# コンテキストをトリミング
trimmed_messages = manage_conversation_context(conversation_history)

# APIリクエスト
response = requests.post(
    "https://api.rustellar.com/v1/chat/completions",
    headers={
        "Authorization": "Bearer YOUR_API_KEY",
        "Content-Type": "application/json"
    },
    json={
        "model": "helix-v1",
        "messages": trimmed_messages
    }
)

Pricing

Helix v1 uses token-based pricing:

Token Type	Price per 1M tokens
Input	$0.50
Output	$1.50

See our Pricing page for more details.

Comparison with Iroha

Feature	Helix v1	Iroha
Best for	General conversation	Story generation
Architecture	Transformer	Mamba
Context Window	8,192 tokens	16,384 tokens
Response Speed	Faster	Moderate
Creativity	Moderate	High

Limitations

While Helix v1 is highly capable, be aware of these limitations:

1. Knowledge Cutoff: Training data up to January 2025
2. Context Window: Limited to 8,192 tokens
3. Factual Accuracy: May occasionally produce incorrect information
4. Calculation: Not optimized for complex mathematical calculations
5. Real-time Data: No access to current events or live data

Support

For technical support or questions about Helix v1:

• Email: support@rustellar.com
• Documentation: platform.rustellar.com/docs

Related Resources

• Quick Start Guide
• Input Formats
• Output Formats
• Rate Limits