วิธีใช้ Haystack 2.0 เชื่อมต่อ LLM ด้วย Python ใน 4 ขั้นตอน

# System prompt
system_prompt = """
You are an expert curator of mental models across science, philosophy, and applied reasoning.
Your task is to explain mental models clearly and accurately using a fixed schema.
If the origin of a model is unclear or debated, state that explicitly.
Do not invent historical sources. Be concise and concrete.
"""
# User prompt
user_prompt = "Explain the following mental model: {{model_query}}"

# Import required packages
from Haystack.components.builders import ChatPromptBuilder
from Haystack.dataclasses import ChatMessage
# Create a prompt builder instance
prompt_builder = ChatPromptBuilder(
    template=[
        ChatMessage.from_system(system_prompt.strip()),
        ChatMessage.from_user(user_prompt)
    ],
    required_variables=["model_query"]
)

# Import required packages
import os
# Retrieve the API key
GEMINI_API_KEY = os.getenv("GEMINI_API_KEY")

# Import required packages
from pydantic import BaseModel, Field
from typing import List, Literal
# Define the output schema
class MentalModel(BaseModel):
    # Mental model name
    model_name: str = Field(
        description="The commonly accepted name of the mental model"
    )
    # Origin or source
    origin: str = Field(
        description="Where the model comes from (a person, book, field, or cultural origin)"
    )
    # Brief description
    description: str = Field(
        description="A brief explanation of what the mental model is and why it matters"
    )
    # Example
    example: str = Field(
        description="A concrete real-world example illustrating the mental model"
    )
    # Tags
    tags: List[str] = Field(
        description="Short tags such as decision-making, systems thinking, learning, and philosophy"
    )

# Import required packages
from Haystack.utils import Secret
from Haystack_integrations.components.generators.google_genai import GoogleGenAIChatGenerator
# Create an LLM instance
llm_with_structured_output = GoogleGenAIChatGenerator(
    model="gemini-2.5-flash",
    api_key=Secret.from_token(GEMINI_API_KEY),
    generation_kwargs={
        "temperature": 0.5,
        "response_format": MentalModel
    }
)

# Import the required package
from Haystack import Pipeline
# Create a pipeline instance
chain = Pipeline()

# Add the prompt builder and LLM instances
chain.add_component(
    "prompt_builder", # Name
    prompt_builder # Instance
)
chain.add_component(
    "llm", # Name
    llm_with_structured_output # Instance
)

# Connect the components
chain.connect(
    "prompt_builder.prompt", # Sender: prompt from prompt builder
    "llm.messages" # Receiver: message from LLM instance
)

# Run the pipeline with a mental-model query
raw_result = chain.run(
    data={
        "prompt_builder": {"model_query": "Compound Interest"}
    }
)

# Extract the structured-output text from the raw result
reply_json = raw_result["llm"]["replies"][0].text
# Validate the structured output
result = MentalModel.model_validate_json(reply_json)
# Print the validated structured output
print(result.model_dump())