adk-typescript/dist/cli/utils/evals.js

TypeScript port of Google's Agent Development Kit (ADK)

"use strict";
Object.defineProperty(exports, "__esModule", { value: true });
exports.convertSessionToEvalFormat = convertSessionToEvalFormat;
/**
 * Converts a session into an evaluation format
 *
 * @param session The session to convert
 * @returns Array of evaluation cases extracted from the session
 */
function convertSessionToEvalFormat(session) {
    const evalCases = [];
    const events = session.events || [];
    for (let i = 0; i < events.length; i++) {
        const event = events[i];
        // Skip if not a user event or has no content
        if (event.author !== 'user' || !event.content || !event.content.parts || event.content.parts.length === 0) {
            continue;
        }
        // Extract user query
        const query = event.content.parts[0].text || '';
        // Find the corresponding tool usage or response for the query
        const expectedToolUse = [];
        const intermediateAgentResponses = [];
        // Check subsequent events to extract tool uses or responses for this turn
        for (let j = i + 1; j < events.length; j++) {
            const subsequentEvent = events[j];
            const eventAuthor = subsequentEvent.author || 'agent';
            // If we find a user event, this is the start of a new turn
            if (eventAuthor === 'user') {
                break;
            }
            if (!subsequentEvent.content || !subsequentEvent.content.parts) {
                continue;
            }
            for (const part of subsequentEvent.content.parts) {
                if (part.functionCall) {
                    const toolName = part.functionCall.name || '';
                    const toolInput = part.functionCall.args || {};
                    expectedToolUse.push({
                        tool_name: toolName,
                        tool_input: toolInput,
                    });
                }
                else if (part.text) {
                    // Also keep track of all the natural language responses that
                    // agent (or sub agents) generated
                    intermediateAgentResponses.push({
                        author: eventAuthor,
                        text: part.text
                    });
                }
            }
        }
        // We assume the last natural language intermediate response is the final response
        // from the agent/model. We treat that as a reference.
        evalCases.push({
            query,
            expected_tool_use: expectedToolUse,
            expected_intermediate_agent_responses: intermediateAgentResponses.slice(0, -1),
            reference: intermediateAgentResponses.length > 0 ?
                intermediateAgentResponses[intermediateAgentResponses.length - 1].text : '',
        });
    }
    return evalCases;
}