Use MCP reource

[Hakstar]

Now, MCP can provide prompts, tools, and resources for LLMs. Most agent frameworks, like Autogen and LangChain, support MCP tools, and prompts can be set in system_message. However, how to use resources effectively has become an important challenge.

For example, when setting up a MySQL server with MCP (which only has one tool: executing SQL), the process involves generating an SQL query, executing it with the tool, and then letting the LLM handle the raw data. But in reality, users only need the raw data itself. I'm not sure if using resources can solve this issue or if there's a better way for users to interact with the LLM.

[ekzhu]

Resource could be used to implement memory / RAG, which is controlled by code to add context to model.

See memory: https://microsoft.github.io/autogen/stable/user-guide/agentchat-user-guide/memory.html

cc @victordibia @EItanya

[EItanya]

Currently it's read only, so we couldn't implement the write side of it, but I 1000% agree

[Hakstar]

ahhh....

I am using an MCP server (https://github.com/designcomputer/mysql_mcp_server/tree/main), and I want to build a MySQL agent that can query data from the database. I created a simple demo based on the example, as shown below.

prompt = """
You are a professional MySQL database agent capable of executing SQL statements using tools to fulfill user requests.

## Query Step Guide
1. First, understand the database structure: Use `SHOW TABLES` to view all available tables.
2. Understand table structure: Use `SHOW COLUMNS FROM <TABLE_NAME>` to check table field information.
3. Understand table relationships: Use `SHOW CREATE TABLE <TABLE_NAME>` to view the table creation statement, including foreign key relationships.
4. Construct and execute appropriate queries.

## Common Issue Handling
- When encountering the error `Table <TABLE_NAME> doesn't exist`:
  * Use `SHOW TABLES` to check all available tables.
  * Verify if the table name is misspelled.

- When unsure about the table structure:
  * Use `DESC <TABLE_NAME>` to check column information.
  * Use `SHOW COLUMNS FROM <TABLE_NAME>` to get detailed column information.

- When needing to understand table relationships:
  * Use `SHOW CREATE TABLE <TABLE_NAME>` to check foreign key relationships.

## Execution Process
1. Analyze user requirements to determine the necessary data and tables.
2. If unsure about the table structure, first query the required table structure.
3. Construct appropriate SQL statements, which may involve fuzzy queries.
4. Execute the query and return the tool invocation results.
5. If an error occurs, analyze the issue and modify the query.

## Notes
- Upon task completion, return the tool invocation results using a markdown table and end with `TERMINATE`.
- If the task cannot proceed, inform the user of the issue and end with `TERMINATE`.
"""

However, after execution, the LLM processes the raw data returned from the tool call.

Result:  
It looks like you have provided a tool response containing information about some video devices. I can help parse this data and extract key information. Below is the parsed information:  

1. **Non- Device**:  
   - ID: 10 System  
   - Name: Virtual Video  

2. **Devices List**:  
   - Device IDs range from `************` to `***************`.  
   - Names include: ***, *** Reservoir 1 (this actually shares the same ID, which might be an error), *** Reservoir 3, ***, *** Reservoir, *** Mountain Reservoir, *** , *** Village.  

This is a case of invoking a tool, but perhaps querying MySQL data as a Resource within the MCP server would be a better approach? However, the current agent implementation examples contain very little usage of Resources.

I'm not sure if using SQL queries as an agent tool is the best approach. Workflow may be a better choice, but Workflow cannot fix error during mysql executing. If you have a better implementation model, I would greatly appreciate your valuable suggestions.

[EItanya]

So I actually think using a tool for a DB is the correct thing to do. Resources are static things that can be accessed. So maybe a tool returns a resource ref because it saved a large image file, or something like that. Or maybe it's a dataset that's useful for RAG. Basically dynamic data should probably use a tool, especially when there are parameters.

[Hakstar]

So I actually think using a tool for a DB is the correct thing to do. Resources are static things that can be accessed. So maybe a tool returns a resource ref because it saved a large image file, or something like that. Or maybe it's a dataset that's useful for RAG. Basically dynamic data should probably use a tool, especially when there are parameters.

Thank you very much for your answer! I am going to develop an Agent to complete a database retrieval task, using the aforementioned prompt and a MySQL MCP server as its tools to create an AssistantAgent. If I use agent.run to execute it, the session will terminate after returning the tool execution of SHOW TABLES, and I have to keep telling the agent continue in order to complete the whole task. On the other hand, if I use RoundRobinGroupChat to build a single-agent team, the agent will reprocess the results of the tool call after obtaining the SQL query results, and it will not return the raw SQL query execution results. So far, I haven't found a suitable solution. Could you provide some possible ways to address this issue? Additionally, in this task, the schema from database is needed to assist the LLM in generating SQL query statements. Should the schema be considered as a Resource? Thank you again for your patient explanation!

[ekzhu]

Have you tried customizing your termination condition? Effectively, the round robin team with a termination condition is a while loop.

Also, if you are using a single agent team the second to the last result is going to be the output of the tool.