Add ADK instrumentation sample#418
Conversation
|
Here is the summary of changes. You are about to add 1 region tag.
This comment is generated by snippet-bot.
|
| @tracer.start_as_current_span("run_sql") | ||
| def run_sql_tool(sql_query: str, tool_context: ToolContext) -> dict[str, Any]: | ||
| """Runs a SQLite query. The SQL query can be DDL or DML. Returns the rows if it's a SELECT query.""" | ||
| current_session_db_path = tool_context.state[SESSION_DB_KEY] |
There was a problem hiding this comment.
Should we check if this is None (model hasn't decided to create the session yet) and return error? I think as is it would throw an exception but I'm not sure how ADK handles tool call exceptions
There was a problem hiding this comment.
We can certainly do that, but, the model should figure out what tool to use and use it accordingly - if the dbpath is null, then the model should invoke the create_database tool to create one before executing the query.
If we start anticipating exceptions then we should also validate the sql_query before executing it.
There was a problem hiding this comment.
Did some experiments -
Added a tool that always raises exception. It halts the agent's response. It is recommended to always return structured responses from tools even in case of exception.
Based on this, I added a return statement to the create_db tool and added a check. LLM should be able to read the responses and recover.
2 participants
Sample adopted from https://github.com/aabmass/opentelemetry-operations-python/tree/adk-sample/samples/adk-sql-agent.
Testing
TODO