LLM OpenAI

Überblick

Die Komponente LLM OpenAI bietet Integration mit OpenAI und OpenAI-kompatiblen APIs (Azure OpenAI, Ollama usw.).

Konfiguration

Setup

answers.json

{
  "component-llm-openai": {
    "api_key": "sk-xxx...",
    "api_base": "https://api.openai.com/v1",
    "default_model": "gpt-4"
  }
}

Für Azure OpenAI

{
  "component-llm-openai": {
    "api_key": "your-azure-key",
    "api_base": "https://your-resource.openai.azure.com",
    "api_type": "azure",
    "api_version": "2024-02-15-preview",
    "deployment_name": "gpt-4"
  }
}

Verwendung in Flows

Einfache Completion

- id: ask_llm
  type: llm
  config:
    model: "gpt-4"
    prompt: "What is the capital of France?"
  next: use_response

Mit System-Prompt

- id: classify
  type: llm
  config:
    model: "gpt-4"
    system_prompt: |
      You are a customer service intent classifier.
      Classify the user's message into one of: greeting, question, complaint, other.
      Respond with JSON: {"intent": "...", "confidence": 0.0-1.0}
    prompt: "{{message}}"
    output_format: json
  next: route_intent

Chat-Completion

- id: chat
  type: llm
  config:
    model: "gpt-4"
    messages:
      - role: system
        content: "You are a helpful assistant."
      - role: user
        content: "{{user_message}}"
    temperature: 0.7
    max_tokens: 500

Mit Function Calling

- id: extract_entities
  type: llm
  config:
    model: "gpt-4"
    prompt: "Extract order information from: {{message}}"
    functions:
      - name: extract_order
        description: "Extract order details"
        parameters:
          type: object
          properties:
            order_id:
              type: string
              description: "The order ID"
            product:
              type: string
              description: "Product name"
          required: ["order_id"]
    function_call: "auto"

Parameter

Parameter	Typ	Standard	Beschreibung
model	string	”gpt-4”	Modellname
prompt	string	-	Benutzer-Prompt
system_prompt	string	-	Systemnachricht
messages	array	-	Vollständiges Message-Array
temperature	float	1.0	Kreativität (0-2)
max_tokens	int	-	Maximale Antwort-Tokens
output_format	string	”text”	"text" oder "json"
functions	array	-	Funktionsdefinitionen
function_call	string	”auto”	Function-Call-Modus

Antwortverarbeitung

- id: ask_llm
  type: llm
  config:
    model: "gpt-4"
    prompt: "{{question}}"
  output: llm_response
  next: use_response

- id: use_response
  type: reply
  config:
    message: "{{llm_response}}"

Fehlerbehandlung

- id: safe_llm_call
  type: llm
  config:
    model: "gpt-4"
    prompt: "{{message}}"
    timeout: 30000
    retry_count: 3
  on_error: handle_error
  next: success

- id: handle_error
  type: reply
  config:
    message: "I'm having trouble processing that. Please try again."

Nächste Schritte

• Templates-Komponente
• fast2flow