Как создать ИИ-агента с помощью DSPy и MCP-сервера ClickHouse

В этом руководстве вы узнаете, как создать ИИ-агента с помощью DSPy, который может работать с Песочницей ClickHouse через MCP-сервер ClickHouse.

Предварительные требования

В системе должен быть установлен Python.
В системе должен быть установлен pip.
Вам понадобится API-ключ Anthropic или API-ключ другого провайдера LLM

Вы можете выполнить следующие шаги либо в Python REPL, либо с помощью скрипта.

Пример ноутбукаЭтот пример доступен в виде ноутбука в репозитории examples.

Установка библиотек

Выполните следующие команды с помощью pip, чтобы установить необходимые библиотеки:

pip install -q --upgrade pip
pip install -q dspy
pip install -q mcp

Настройка учетных данных

Затем вам потребуется указать свой API-ключ Anthropic:

import os
os.environ["ANTHROPIC_API_KEY"] = getpass.getpass("Enter Anthropic API Key:")

Использование другого провайдера LLMЕсли у вас нет API-ключа Anthropic и вы хотите использовать другого провайдера LLM, инструкции по настройке учетных данных можно найти в документации DSPy

Далее укажите учетные данные, необходимые для подключения к Песочнице ClickHouse SQL:

env = {
    "CLICKHOUSE_HOST": "sql-clickhouse.clickhouse.com",
    "CLICKHOUSE_PORT": "8443",
    "CLICKHOUSE_USER": "demo",
    "CLICKHOUSE_PASSWORD": "",
    "CLICKHOUSE_SECURE": "true"
}

Инициализируйте MCP-сервер

Теперь настройте MCP-сервер ClickHouse так, чтобы он был подключён к Песочнице ClickHouse.

from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client
import dspy

server_params = StdioServerParameters(
    command="uv",
    args=[
        'run',
        '--with', 'mcp-clickhouse',
        '--python', '3.13',
        'mcp-clickhouse'
    ],
    env=env
)

Инициализируйте LLM

Затем инициализируйте LLM следующей строкой:

dspy.configure(lm=dspy.LM("anthropic/claude-sonnet-4-20250514"))

Запуск агента

Наконец, инициализируйте и запустите агент:

class DataAnalyst(dspy.Signature):
    """You are a data analyst. You'll be asked questions and you need to try to answer them using the tools you have access to. """

    user_request: str = dspy.InputField()
    process_result: str = dspy.OutputField(
        desc=(
            "Answer to the query"
        )
    )

from utils import print_dspy_result

async with stdio_client(server_params) as (read, write):
    async with ClientSession(read, write) as session:
        await session.initialize()
        tools = await session.list_tools()

        dspy_tools = []
        for tool in tools.tools:
            dspy_tools.append(dspy.Tool.from_mcp_tool(session, tool))

        react = dspy.ReAct(DataAnalyst, tools=dspy_tools)
        result = await react.acall(user_request="What's the most popular Amazon product category")
        print_dspy_result(result)

Response

================================================================================
🤖 DSPy ReAct Result
================================================================================

📍 STEP 1
----------------------------------------
🧠 THINKING: I need to find information about Amazon product categories and determine which one is most popular. First, I should explore what databases are available to see if there's any Amazon-related data.

🔧 TOOL: list_databases

📊 RESULT:
   amazon
bluesky
country
covid
default
dns
environmental
food
forex
geo
git
github
hackernews
imdb
logs
metrica
mgbench
mta
noaa
nyc_taxi
nypd
ontime
opensky
otel
otel_v2
pypi
random
reddit
rubygems
sta...

📍 STEP 2
----------------------------------------
🧠 THINKING: Great! I can see there's an "amazon" database available. This is exactly what I need to find information about Amazon product categories. Let me explore the tables in the amazon database to see what data is available.

🔧 TOOL: list_tables
   Args: {'database': 'amazon'}

📊 RESULT:
   {
  "database": "amazon",
  "name": "amazon_reviews",
  "comment": "",
  "columns": [
    {
      "name": "review_date",
      "type": "Date",
      "default_type": "",
      "default_expression": "",...

📍 STEP 3
----------------------------------------
🧠 THINKING: Perfect! I found the amazon_reviews table which contains a product_category column. This table has over 150 million rows of Amazon review data, which should give me a good representation of product popularity. To find the most popular Amazon product category, I need to count the number of reviews per category and order them by count in descending order. The category with the most reviews would be considered the most popular.

🔧 TOOL: run_select_query
   Args: {'query': 'SELECT product_category, COUNT(*) as review_count FROM amazon.amazon_reviews GROUP BY product_category ORDER BY review_count DESC LIMIT 10'}

📊 RESULT:
   1. Books: 19,530,930 reviews
   2. Digital_Ebook_Purchase: 17,622,107 reviews
   3. Wireless: 9,001,567 reviews
   4. PC: 6,908,245 reviews
   5. Home: 6,221,321 reviews
   ... and 5 more

📍 STEP 4
----------------------------------------
🧠 THINKING: Perfect! I have found the answer to the user's question. Based on the Amazon reviews data, I can clearly see that "Books" is the most popular Amazon product category with 19,530,930 reviews, followed by "Digital_Ebook_Purchase" with 17,622,107 reviews. The data shows a clear ranking of product categories by popularity based on the number of reviews each category has received. I now have all the information needed to answer the user's question.

🔧 TOOL: finish

📊 RESULT:
   Completed.

🎯 REASONING
----------------------------------------
To answer the question about the most popular Amazon product category, I accessed the Amazon database and found a table called "amazon_reviews" with over 150 million review records. This table contains a "product_category" column that allows me to analyze product popularity based on review volume. I ran a query to count the number of reviews per product category and ordered them by count in descending order. The assumption here is that categories with more reviews indicate higher popularity, as more reviews typically correlate with more purchases and customer engagement.

✅ FINAL RESULT
----------------------------------------
Based on Amazon review data containing over 150 million records, the most popular Amazon product category is **Books** with 19,530,930 reviews. 

Here are the top 10 most popular Amazon product categories by review count:

1. **Books** - 19,530,930 reviews
2. **Digital_Ebook_Purchase** - 17,622,107 reviews  
3. **Wireless** - 9,001,567 reviews
4. **PC** - 6,908,245 reviews
5. **Home** - 6,221,321 reviews
6. **Apparel** - 5,906,085 reviews
7. **Health & Personal Care** - 5,331,239 reviews
8. **Beauty** - 5,115,462 reviews
9. **Video DVD** - 5,069,014 reviews
10. **Mobile_Apps** - 5,033,164 reviews

It's interesting to note that Books and Digital Ebook Purchase (which are related categories) together account for over 37 million reviews, showing the strong popularity of reading materials on Amazon's platform.
================================================================================

Руководства по ClickHouse

Руководства по сценариям использования

Как создать ИИ-агента с помощью DSPy и MCP-сервера ClickHouse

Предварительные требования

​Предварительные требования

Предварительные требования