Vanna 2.0：How to provide DDL, business knowledge, and SQL examples to Vanna 2.0 for correct SQL generation ?

[davychen2007-spec]

Hi Vanna team,

I am trying to use Vanna 2.0 with my MySQL database. Here is a simplified version of my setup:

from vanna import ToolRegistry, User, Agent
from vanna.integrations.mysql import MySQLRunner
from vanna.integrations.openai import OpenAILlmService
from vanna.tools import RunSqlTool, VisualizeDataTool
from vanna.servers.fastapi import VannaFastAPIServer

llm = OpenAILlmService(model="qwen-max-latest", base_url="https://dashscope.aliyuncs.com/compatible-mode/v1")

mysql_runner = MySQLRunner(host="...", port=3306, user="...", password="...", database="...")

db_tool = RunSqlTool(sql_runner=mysql_runner)
tools = ToolRegistry()
tools.register_local_tool(db_tool, access_groups=['admin', 'user'])
tools.register_local_tool(SaveQuestionToolArgsTool(), access_groups=['admin'])
tools.register_local_tool(SearchSavedCorrectToolUsesTool(), access_groups=['admin', 'user'])
tools.register_local_tool(SaveTextMemoryTool(), access_groups=['admin', 'user'])
tools.register_local_tool(VisualizeDataTool(), access_groups=['admin', 'user'])

agent = Agent(llm_service=llm, tool_registry=tools)
server = VannaFastAPIServer(agent=agent)
server.run()

When I ask questions in the frontend, I expect Vanna to generate correct SQL and display the results. However, it seems that Vanna does not understand my database schema, and the generated SQL is incorrect.

In Vanna 0.x, I could do:

vn.train(ddl=...)
vn.train(documentation=...)
vn.train(question=..., sql=...)

to teach Vanna about the database schema, business knowledge, and SQL examples.

In Vanna 2.0, I could not find similar functionality. I assumed Vanna would automatically retrieve the database DDL from the sql runner, but that does not seem to be the case.

My question:
What is the recommended way in Vanna 2.0 to provide:

1. Database DDL
2. Business knowledge / documentation
3. SQL examples

so that the agent can generate correct SQL queries?

Any guidance or examples would be greatly appreciated.

Thanks!

[jeffsburton]

Here is what I'm doing (I'm a novice at vanna btw):
I first annotate my schema model (I'm using sqlalchemy) with info={"docs": "..."} on tables and columns. Then I get a dump of all that information like this:

def get_schema_documentation():
docs = []
# Iterate over all tables in the metadata
for table_name, table in Base.metadata.tables.items():
# Check for table-level docs
if hasattr(table, 'info') and table.info.get('docs'):
docs.append(f"Table {table_name}: {table.info['docs']}")

    # Check for column-level docss
    for column in table.columns:
        if hasattr(column, 'info') and column.info.get('docs'):
            docs.append(f"Column {table_name}.{column.name}: {column.info['docs']}")

return "\n".join(docs)

I have a LlmContextEnhancer that uses that includes that information with the prompt:

    async def enhance_system_prompt(
            self,
            system_prompt: str,
            user_message: str,
            user: User
    ) -> str:

        # Add documentation to system prompt
        schema_section = "\n\n## Relevant Schema Information\n\n"

        docs = get_schema_documentation()

        print(docs)

        return system_prompt + schema_section + docs

It works pretty well. I see that their enhancer code has a primitive way of only including "relevant" schema information, but it seems simplistic and error prone, so I just give the prompt everything. My schema only has 60 tables.

[Abdohashem67]

@jeffsburton Would this be robust on handling the large schemas

[jeffsburton]

I don't know. My schema has 60 tables, probably 500 columns. Jeff Burton ***@***.*** ***@***.***>* *715-222-5655*

…

On Sat, Nov 29, 2025, 8:24 AM عبدالرحمن احمد هاشم علي < ***@***.***> wrote: @jeffsburton <https://github.com/jeffsburton> Would this be robust on handling the large schemas — Reply to this email directly, view it on GitHub <#1027 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BDUDCN4TVUYFKPGCD6KYN7T37GUCFAVCNFSM6AAAAACND4NSR6VHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTKMJRGE4TEOI> . You are receiving this because you were mentioned.Message ID: ***@***.***>