生成 DuckDB 的 SQL (使用 Ollama、其他向量数据库)

本 Notebook 介绍了使用 vanna Python 包生成 SQL 的过程，该过程使用 AI（RAG + LLMs），包括连接数据库和训练。如果您尚未准备好在自己的数据库上进行训练，仍然可以使用示例 SQLite 数据库进行尝试。

使用 Colab 运行在 GitHub 上打开

您想使用哪个 LLM？

OpenAI
使用您自己的 API 密钥使用 OpenAI
Azure OpenAI
如果您在 Azure 上部署了 OpenAI 模型
Anthropic
使用您的 Anthropic API 密钥使用 Anthropic 的 Claude
[Selected] Ollama
在本地免费使用 Ollama。需要额外设置。
Google Gemini
使用您的 Gemini 或 Vertex API 密钥使用 Google Gemini
通过 Mistral API 使用 Mistral
如果您有 Mistral API 密钥
其他 LLM
如果您有不同的 LLM 模型

您想在哪里存储“训练”数据？

ChromaDB
在本地免费使用 ChromaDB 的开源向量数据库。无需额外设置 -- 所有数据库文件都将在本地创建和存储。
Qdrant
使用 Qdrant 的开源向量数据库
Marqo
在本地免费使用 Marqo。需要额外设置。或者使用他们托管的选项。
[Selected] 其他向量数据库
使用任何其他向量数据库。需要额外设置。

设置

%pip install 'vanna[ollama,duckdb]'

from vanna.ollama import Ollama
from vanna.base import VannaBase

class MyCustomVectorDB(VannaBase):
  def add_ddl(self, ddl: str, **kwargs) -> str:
     # Implement here

  def add_documentation(self, doc: str, **kwargs) -> str:
     # Implement here

  def add_question_sql(self, question: str, sql: str, **kwargs) -> str:
     # Implement here

  def get_related_ddl(self, question: str, **kwargs) -> list:
     # Implement here

  def get_related_documentation(self, question: str, **kwargs) -> list:
     # Implement here

  def get_similar_question_sql(self, question: str, **kwargs) -> list:
     # Implement here

  def get_training_data(self, **kwargs) -> pd.DataFrame:
     # Implement here

  def remove_training_data(id: str, **kwargs) -> bool:
     # Implement here


class MyVanna(MyCustomVectorDB, Ollama):
    def __init__(self, config=None):
        MyCustomVectorDB.__init__(self, config=config)
        Ollama.__init__(self, config=config)

vn = MyVanna(config={'model': 'mistral'})

您想查询哪个数据库？

vn.connect_to_duckdb(url='motherduck:')

训练

您只需要训练一次。除非想添加更多训练数据，否则请勿再次训练。

# The information schema query may need some tweaking depending on your database. This is a good starting point.
df_information_schema = vn.run_sql("SELECT * FROM INFORMATION_SCHEMA.COLUMNS")

# This will break up the information schema into bite-sized chunks that can be referenced by the LLM
plan = vn.get_training_plan_generic(df_information_schema)
plan

# If you like the plan, then uncomment this and run it to train
# vn.train(plan=plan)

# The following are methods for adding training data. Make sure you modify the examples to match your database.

# DDL statements are powerful because they specify table names, colume names, types, and potentially relationships
vn.train(ddl="""
    CREATE TABLE IF NOT EXISTS my-table (
        id INT PRIMARY KEY,
        name VARCHAR(100),
        age INT
    )
""")

# Sometimes you may want to add documentation about your business terminology or definitions.
vn.train(documentation="Our business defines OTIF score as the percentage of orders that are delivered on time and in full")

# You can also add SQL queries to your training data. This is useful if you have some queries already laying around. You can just copy and paste those from your editor to begin generating new SQL.
vn.train(sql="SELECT * FROM my-table WHERE name = 'John Doe'")

# At any time you can inspect what training data the package is able to reference
training_data = vn.get_training_data()
training_data

# You can remove training data if there's obsolete/incorrect information. 
vn.remove_training_data(id='1-ddl')

```## Asking the AI
Whenever you ask a new question, it will find the 10 most relevant pieces of training data and use it as part of the LLM prompt to generate the SQL.
```python
vn.ask(question=...)

启动用户界面

vanna-flask

from vanna.flask import VannaFlaskApp
app = VannaFlaskApp(vn)
app.run()