|
24 | 24 | import sys
|
25 | 25 | import traceback
|
26 | 26 | import typing
|
27 |
| -from typing import Any, List, Literal, Optional, Union |
| 27 | +from typing import Any |
| 28 | +from typing import List |
| 29 | +from typing import Literal |
| 30 | +from typing import Optional |
| 31 | +from typing import Union |
28 | 32 |
|
29 | 33 | import click
|
30 | 34 | from fastapi import FastAPI
|
|
52 | 56 | from ..agents.base_agent import BaseAgent
|
53 | 57 | from ..agents.live_request_queue import LiveRequest
|
54 | 58 | from ..agents.live_request_queue import LiveRequestQueue
|
55 |
| -from ..agents.llm_agent import Agent, LlmAgent |
| 59 | +from ..agents.llm_agent import Agent |
| 60 | +from ..agents.llm_agent import LlmAgent |
56 | 61 | from ..agents.run_config import StreamingMode
|
57 | 62 | from ..artifacts import InMemoryArtifactService
|
58 | 63 | from ..events.event import Event
|
@@ -467,30 +472,24 @@ async def run_eval(
|
467 | 472 | "Eval ids to run list is empty. We will all evals in the eval set."
|
468 | 473 | )
|
469 | 474 | root_agent = await _get_root_agent_async(app_name)
|
470 |
| - eval_results = list( |
471 |
| - await run_evals( |
| 475 | + return [ |
| 476 | + RunEvalResult( |
| 477 | + app_name=app_name, |
| 478 | + eval_set_id=eval_set_id, |
| 479 | + eval_id=eval_result.eval_id, |
| 480 | + final_eval_status=eval_result.final_eval_status, |
| 481 | + eval_metric_results=eval_result.eval_metric_results, |
| 482 | + session_id=eval_result.session_id, |
| 483 | + ) |
| 484 | + async for eval_result in run_evals( |
472 | 485 | eval_set_to_evals,
|
473 | 486 | root_agent,
|
474 | 487 | getattr(root_agent, "reset_data", None),
|
475 | 488 | req.eval_metrics,
|
476 | 489 | session_service=session_service,
|
477 | 490 | artifact_service=artifact_service,
|
478 | 491 | )
|
479 |
| - ) |
480 |
| - |
481 |
| - run_eval_results = [] |
482 |
| - for eval_result in eval_results: |
483 |
| - run_eval_results.append( |
484 |
| - RunEvalResult( |
485 |
| - app_name=app_name, |
486 |
| - eval_set_id=eval_set_id, |
487 |
| - eval_id=eval_result.eval_id, |
488 |
| - final_eval_status=eval_result.final_eval_status, |
489 |
| - eval_metric_results=eval_result.eval_metric_results, |
490 |
| - session_id=eval_result.session_id, |
491 |
| - ) |
492 |
| - ) |
493 |
| - return run_eval_results |
| 492 | + ] |
494 | 493 |
|
495 | 494 | @app.delete("/apps/{app_name}/users/{user_id}/sessions/{session_id}")
|
496 | 495 | def delete_session(app_name: str, user_id: str, session_id: str):
|
|
0 commit comments