Data Reflux System
In recent years, increasing engineering experience has shown that building a data flywheel aligned with business scenarios is the most critical factor in enhancing agent performance and the underlying large model performance. The data reflux system serves as a powerful tool to achieve this.
Here, the openmcp project team is also attempting to build an open-source data reflux system as an auxiliary module of openmcp. For details about the reflux system, please refer to my previous blog: Why Do I Need to Design a Reflux System? openmcp Reflux Uses Quantitative Metrics to Help You Better Understand Your Agent System and MCP.
Enabling Data Reflux
In openmcp, you can enable data reflux through the reflux option:
const agent = new OmAgent();
agent.loadMcpConfig(useMcpConfig('word-mcp'));
const prompt = await agent.getPrompt('word_operations_prompt', {});
const query = prompt + '\n' + userInput;
const result = await agent.ainvoke({
messages: query,
reflux: {
enabled: true,
saveDir: './dataset'
}
});This way, openmcp will save each interaction of the current MCP in the ./dataset folder, with the filename {MCP server name}.omdb.
You can use the latest version of the openmcp plugin to view this file and perform performance analysis and data export.
Validator System
to be continue
LSTM-Kirigaya (锦恢)