441 lines
13 KiB
Python
441 lines
13 KiB
Python
|
from __future__ import annotations
|
||
|
|
|
||
|
|
import json
|
||
|
|
import os
|
||
|
|
import time
|
||
|
|
from uuid import uuid4
|
||
|
|
|
||
|
|
import httpx
|
||
|
|
import jwt
|
||
|
|
import pytest
|
||
|
|
|
||
|
|
from backend.tests.quality.evaluators import ModelScorecard, ScoreDetail, ScenarioScore
|
||
|
|
from backend.tests.quality.scenarios import ALL_SCENARIOS
|
||
|
|
|
||
|
|
CANDIDATE_MODELS = ["qwen3.5-flash", "deepseek-chat"]
|
||
|
|
|
||
|
|
MODEL_LLM_IDS = {
|
||
|
|
"qwen3.5-flash": "c625bce4-970e-4a76-bebe-cb8840fed854",
|
||
|
|
"deepseek-chat": "12bc1963-4b67-404b-b952-5948bea0f690",
|
||
|
|
}
|
||
|
|
|
||
|
|
BASE_URL = os.getenv("AGENT_LIVE_BASE_URL", "http://localhost:5775")
|
||
|
|
|
||
|
|
|
||
|
|
def _load_env() -> None:
|
||
|
|
from pathlib import Path
|
||
|
|
|
||
|
|
env_path = Path(__file__).resolve().parents[3] / ".env"
|
||
|
|
if env_path.exists():
|
||
|
|
for line in env_path.read_text().splitlines():
|
||
|
|
line = line.strip()
|
||
|
|
if not line or line.startswith("#") or "=" not in line:
|
||
|
|
continue
|
||
|
|
key, _, value = line.partition("=")
|
||
|
|
key = key.strip()
|
||
|
|
value = value.strip().strip('"').strip("'")
|
||
|
|
if key and key not in os.environ:
|
||
|
|
os.environ[key] = value
|
||
|
|
|
||
|
|
|
||
|
|
_load_env()
|
||
|
|
|
||
|
|
|
||
|
|
def _get_jwt_secret() -> str:
|
||
|
|
secret = (
|
||
|
|
os.getenv("SOCIAL_SUPABASE__JWT_SECRET")
|
||
|
|
or os.getenv("SUPABASE_JWT_SECRET")
|
||
|
|
or os.getenv("JWT_SECRET")
|
||
|
|
)
|
||
|
|
if not secret:
|
||
|
|
raise RuntimeError("JWT_SECRET not found in environment")
|
||
|
|
return secret
|
||
|
|
|
||
|
|
|
||
|
|
def _get_supabase_url() -> str:
|
||
|
|
return (
|
||
|
|
os.getenv("SOCIAL_SUPABASE__PUBLIC_URL")
|
||
|
|
or os.getenv("SOCIAL_SUPABASE__URL")
|
||
|
|
or os.getenv("SUPABASE_URL")
|
||
|
|
or "http://localhost:54321"
|
||
|
|
)
|
||
|
|
|
||
|
|
|
||
|
|
def _get_supabase_key() -> str:
|
||
|
|
from core.config.settings import config
|
||
|
|
|
||
|
|
key = os.getenv("SOCIAL_SUPABASE__SERVICE_ROLE_KEY", "")
|
||
|
|
if key:
|
||
|
|
return key
|
||
|
|
return config.supabase.service_role_key
|
||
|
|
|
||
|
|
|
||
|
|
def _get_test_user_id() -> str:
|
||
|
|
user_id = os.getenv("TEST_USER_ID")
|
||
|
|
if user_id:
|
||
|
|
return user_id
|
||
|
|
raise RuntimeError("TEST_USER_ID not set")
|
||
|
|
|
||
|
|
|
||
|
|
def _create_jwt(user_id: str) -> str:
|
||
|
|
now = int(time.time())
|
||
|
|
payload = {
|
||
|
|
"sub": user_id,
|
||
|
|
"role": "authenticated",
|
||
|
|
"aud": "authenticated",
|
||
|
|
"iss": _get_supabase_url(),
|
||
|
|
"iat": now,
|
||
|
|
"exp": now + 3600,
|
||
|
|
}
|
||
|
|
return jwt.encode(payload, _get_jwt_secret(), algorithm="HS256")
|
||
|
|
|
||
|
|
|
||
|
|
async def _run_via_http(
|
||
|
|
*,
|
||
|
|
user_message: str,
|
||
|
|
token: str,
|
||
|
|
timeout: float = 120.0,
|
||
|
|
) -> dict:
|
||
|
|
thread_id = str(uuid4())
|
||
|
|
run_id = f"q-{uuid4().hex[:12]}"
|
||
|
|
|
||
|
|
async with httpx.AsyncClient(timeout=httpx.Timeout(timeout)) as client:
|
||
|
|
headers = {"Authorization": f"Bearer {token}"}
|
||
|
|
|
||
|
|
run_resp = await client.post(
|
||
|
|
f"{BASE_URL}/api/v1/agent/runs",
|
||
|
|
headers=headers,
|
||
|
|
json={
|
||
|
|
"threadId": thread_id,
|
||
|
|
"runId": run_id,
|
||
|
|
"state": {},
|
||
|
|
"messages": [
|
||
|
|
{"id": "u1", "role": "user", "content": user_message}
|
||
|
|
],
|
||
|
|
"tools": [],
|
||
|
|
"context": [],
|
||
|
|
"forwardedProps": {"runtime_mode": "chat"},
|
||
|
|
},
|
||
|
|
)
|
||
|
|
run_data = run_resp.json()
|
||
|
|
eff_thread = str(run_data.get("threadId", thread_id))
|
||
|
|
eff_run = run_data.get("runId", run_id)
|
||
|
|
events_url = (
|
||
|
|
f"{BASE_URL}/api/v1/agent/runs/{eff_thread}/events"
|
||
|
|
f"?runId={eff_run}"
|
||
|
|
)
|
||
|
|
|
||
|
|
t_start = time.monotonic()
|
||
|
|
|
||
|
|
tool_results: list[dict] = []
|
||
|
|
all_events: list[dict] = []
|
||
|
|
final_answer = ""
|
||
|
|
run_finished = False
|
||
|
|
token_usage: dict = {}
|
||
|
|
|
||
|
|
async with client.stream(
|
||
|
|
"GET", events_url, headers=headers, timeout=timeout
|
||
|
|
) as sse:
|
||
|
|
buffer = ""
|
||
|
|
async for line in sse.aiter_lines():
|
||
|
|
if line.startswith("data:"):
|
||
|
|
data_str = line.split(":", 1)[1].strip()
|
||
|
|
if data_str:
|
||
|
|
buffer = data_str
|
||
|
|
elif line == "" and buffer:
|
||
|
|
try:
|
||
|
|
ev = json.loads(buffer)
|
||
|
|
all_events.append(ev)
|
||
|
|
etype = ev.get("type")
|
||
|
|
|
||
|
|
if etype == "TOOL_CALL_RESULT":
|
||
|
|
tool_results.append(ev)
|
||
|
|
elif etype == "TEXT_MESSAGE_END":
|
||
|
|
final_answer = ev.get("answer", "") or ev.get("text", "")
|
||
|
|
token_usage = {
|
||
|
|
"totalTokens": ev.get("totalTokens", 0),
|
||
|
|
"inputTokens": ev.get("inputTokens", 0),
|
||
|
|
"outputTokens": ev.get("outputTokens", 0),
|
||
|
|
"promptCacheMissTokens": ev.get(
|
||
|
|
"promptCacheMissTokens", 0
|
||
|
|
),
|
||
|
|
"promptCacheHitTokens": ev.get(
|
||
|
|
"promptCacheHitTokens", 0
|
||
|
|
),
|
||
|
|
}
|
||
|
|
elif etype in {"RUN_FINISHED", "RUN_ERROR"}:
|
||
|
|
run_finished = True
|
||
|
|
except json.JSONDecodeError:
|
||
|
|
pass
|
||
|
|
buffer = ""
|
||
|
|
|
||
|
|
t_end = time.monotonic()
|
||
|
|
|
||
|
|
tool_names = [
|
||
|
|
tr.get("tool_name", "") or tr.get("toolName", "")
|
||
|
|
for tr in tool_results
|
||
|
|
]
|
||
|
|
successful_tool_names = [
|
||
|
|
tr.get("tool_name", "") or tr.get("toolName", "")
|
||
|
|
for tr in tool_results
|
||
|
|
if tr.get("status") in ("success", "partial")
|
||
|
|
]
|
||
|
|
|
||
|
|
return {
|
||
|
|
"final_answer": final_answer,
|
||
|
|
"tool_results": tool_results,
|
||
|
|
"tool_names": tool_names,
|
||
|
|
"successful_tool_names": successful_tool_names,
|
||
|
|
"run_finished": run_finished,
|
||
|
|
"latency_ms": round((t_end - t_start) * 1000),
|
||
|
|
"token_usage": token_usage,
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
def _switch_model(model_code: str) -> None:
|
||
|
|
from supabase import create_client
|
||
|
|
|
||
|
|
sb = create_client(_get_supabase_url(), _get_supabase_key())
|
||
|
|
llm_id = MODEL_LLM_IDS[model_code]
|
||
|
|
for agent_type in ("router", "worker"):
|
||
|
|
(
|
||
|
|
sb.table("system_agents")
|
||
|
|
.update({"llm_id": llm_id})
|
||
|
|
.eq("agent_type", agent_type)
|
||
|
|
.execute()
|
||
|
|
)
|
||
|
|
|
||
|
|
|
||
|
|
def _save_original_models() -> list[dict]:
|
||
|
|
from supabase import create_client
|
||
|
|
|
||
|
|
sb = create_client(_get_supabase_url(), _get_supabase_key())
|
||
|
|
return (
|
||
|
|
sb.table("system_agents")
|
||
|
|
.select("agent_type, llm_id")
|
||
|
|
.execute()
|
||
|
|
.data
|
||
|
|
)
|
||
|
|
|
||
|
|
|
||
|
|
def _restore_models(original_rows: list[dict]) -> None:
|
||
|
|
from supabase import create_client
|
||
|
|
|
||
|
|
sb = create_client(_get_supabase_url(), _get_supabase_key())
|
||
|
|
for row in original_rows:
|
||
|
|
(
|
||
|
|
sb.table("system_agents")
|
||
|
|
.update({"llm_id": row["llm_id"]})
|
||
|
|
.eq("agent_type", row["agent_type"])
|
||
|
|
.execute()
|
||
|
|
)
|
||
|
|
|
||
|
|
|
||
|
|
def _evaluate_answer_quality(
|
||
|
|
*,
|
||
|
|
answer: str,
|
||
|
|
run_finished: bool,
|
||
|
|
expect_tool_use: bool,
|
||
|
|
has_tool_success: bool,
|
||
|
|
tool_names: list[str],
|
||
|
|
) -> float:
|
||
|
|
if not run_finished:
|
||
|
|
return 0.0
|
||
|
|
if not answer or not answer.strip():
|
||
|
|
return 0.0
|
||
|
|
|
||
|
|
score = 0.6
|
||
|
|
|
||
|
|
if expect_tool_use:
|
||
|
|
if has_tool_success:
|
||
|
|
score += 0.2
|
||
|
|
elif tool_names:
|
||
|
|
score += 0.1
|
||
|
|
else:
|
||
|
|
score -= 0.3
|
||
|
|
else:
|
||
|
|
if not tool_names:
|
||
|
|
score += 0.2
|
||
|
|
else:
|
||
|
|
score -= 0.1
|
||
|
|
|
||
|
|
if len(answer) > 10:
|
||
|
|
score += 0.1
|
||
|
|
|
||
|
|
if "无法" in answer or "失败" in answer or "错误" in answer:
|
||
|
|
if expect_tool_use:
|
||
|
|
score -= 0.1
|
||
|
|
|
||
|
|
return max(0.0, min(1.0, score))
|
||
|
|
|
||
|
|
|
||
|
|
def _evaluate_criteria(
|
||
|
|
*,
|
||
|
|
answer: str,
|
||
|
|
run_finished: bool,
|
||
|
|
tool_names: list[str],
|
||
|
|
has_tool_success: bool,
|
||
|
|
tool_results: list[dict],
|
||
|
|
scenario: object,
|
||
|
|
) -> list[ScoreDetail]:
|
||
|
|
details: list[ScoreDetail] = []
|
||
|
|
for criterion in getattr(scenario, "quality_criteria", []):
|
||
|
|
passed = False
|
||
|
|
note = ""
|
||
|
|
|
||
|
|
if "调用" in criterion or "project_cli" in criterion:
|
||
|
|
passed = any("project_cli" in tn for tn in tool_names)
|
||
|
|
note = f"tools: {tool_names}" if not passed else ""
|
||
|
|
elif "mode" in criterion and "day" in criterion:
|
||
|
|
for tr in tool_results:
|
||
|
|
args = tr.get("tool_call_args", {}) or tr.get("toolCallArgs", {})
|
||
|
|
inp = args.get("input", {})
|
||
|
|
if isinstance(inp, dict) and inp.get("mode") == "day":
|
||
|
|
passed = True
|
||
|
|
break
|
||
|
|
elif "具体" in criterion or "时间戳" in criterion:
|
||
|
|
passed = has_tool_success
|
||
|
|
elif "基于工具" in criterion or "返回" in criterion:
|
||
|
|
passed = has_tool_success
|
||
|
|
elif "无日程" in criterion:
|
||
|
|
passed = "无" in answer or "没有" in answer
|
||
|
|
elif "简短" in criterion or "简洁" in criterion:
|
||
|
|
passed = 0 < len(answer) < 200
|
||
|
|
elif "自我介绍" in criterion:
|
||
|
|
passed = "Linksy" in answer or "助手" in answer
|
||
|
|
elif "礼貌" in criterion:
|
||
|
|
passed = len(answer) > 0
|
||
|
|
else:
|
||
|
|
passed = run_finished and len(answer) > 0
|
||
|
|
|
||
|
|
details.append(ScoreDetail(criterion=criterion, passed=passed, note=note))
|
||
|
|
return details
|
||
|
|
|
||
|
|
|
||
|
|
async def _run_model_scenarios(model_code: str, user_id: str) -> ModelScorecard:
|
||
|
|
from services.llm_pricing.service import LlmPricingService
|
||
|
|
|
||
|
|
pricing = LlmPricingService()
|
||
|
|
token = _create_jwt(user_id)
|
||
|
|
scores: list[ScenarioScore] = []
|
||
|
|
|
||
|
|
for scenario in ALL_SCENARIOS:
|
||
|
|
result = await _run_via_http(
|
||
|
|
user_message=scenario.prompt,
|
||
|
|
token=token,
|
||
|
|
)
|
||
|
|
|
||
|
|
answer = result["final_answer"]
|
||
|
|
tool_names = result["tool_names"]
|
||
|
|
has_tool_success = len(result["successful_tool_names"]) > 0
|
||
|
|
tu = result["token_usage"]
|
||
|
|
|
||
|
|
total_tokens = tu.get("totalTokens", 0)
|
||
|
|
input_tokens = tu.get("inputTokens", 0) or tu.get("promptCacheMissTokens", 0)
|
||
|
|
output_tokens = tu.get("outputTokens", 0) or max(total_tokens - input_tokens, 0)
|
||
|
|
|
||
|
|
try:
|
||
|
|
cost_usd = pricing.calculate_cost(
|
||
|
|
model=model_code,
|
||
|
|
prompt_tokens=input_tokens,
|
||
|
|
completion_tokens=output_tokens,
|
||
|
|
cached_prompt_tokens=tu.get("promptCacheHitTokens", 0),
|
||
|
|
)
|
||
|
|
except ValueError:
|
||
|
|
cost_usd = 0.0
|
||
|
|
cost_usd = round(cost_usd, 8)
|
||
|
|
|
||
|
|
tool_called = any("project_cli" in tn for tn in tool_names)
|
||
|
|
tool_succeeded = has_tool_success if scenario.expect_tool_use else True
|
||
|
|
|
||
|
|
answer_quality = _evaluate_answer_quality(
|
||
|
|
answer=answer,
|
||
|
|
run_finished=result["run_finished"],
|
||
|
|
expect_tool_use=scenario.expect_tool_use,
|
||
|
|
has_tool_success=has_tool_success,
|
||
|
|
tool_names=tool_names,
|
||
|
|
)
|
||
|
|
|
||
|
|
details = _evaluate_criteria(
|
||
|
|
answer=answer,
|
||
|
|
run_finished=result["run_finished"],
|
||
|
|
tool_names=tool_names,
|
||
|
|
has_tool_success=has_tool_success,
|
||
|
|
tool_results=result["tool_results"],
|
||
|
|
scenario=scenario,
|
||
|
|
)
|
||
|
|
|
||
|
|
print(
|
||
|
|
f" [{model_code}] {scenario.id:<25} "
|
||
|
|
f"lat={result['latency_ms']}ms "
|
||
|
|
f"tokens={total_tokens} "
|
||
|
|
f"cost=${cost_usd:.6f} "
|
||
|
|
f"tool={'OK' if has_tool_success else 'FAIL'} "
|
||
|
|
f"answer={answer[:60]}"
|
||
|
|
)
|
||
|
|
|
||
|
|
scores.append(
|
||
|
|
ScenarioScore(
|
||
|
|
scenario_id=scenario.id,
|
||
|
|
model_code=model_code,
|
||
|
|
latency_ms=result["latency_ms"],
|
||
|
|
input_tokens=input_tokens,
|
||
|
|
output_tokens=output_tokens,
|
||
|
|
cost_usd=cost_usd,
|
||
|
|
tool_called=tool_called,
|
||
|
|
tool_succeeded=tool_succeeded,
|
||
|
|
answer_quality=answer_quality,
|
||
|
|
details=details,
|
||
|
|
raw_answer=answer[:500],
|
||
|
|
run_finished=result["run_finished"],
|
||
|
|
)
|
||
|
|
)
|
||
|
|
|
||
|
|
return ModelScorecard(model_code=model_code, scenario_scores=scores)
|
||
|
|
|
||
|
|
|
||
|
|
@pytest.fixture(autouse=True)
|
||
|
|
def _check_env():
|
||
|
|
if os.getenv("QUALITY_TEST") != "1":
|
||
|
|
pytest.skip("set QUALITY_TEST=1 to run quality tests")
|
||
|
|
|
||
|
|
|
||
|
|
@pytest.fixture(autouse=True)
|
||
|
|
def _require_test_user_id():
|
||
|
|
_get_test_user_id()
|
||
|
|
|
||
|
|
|
||
|
|
@pytest.mark.asyncio
|
||
|
|
@pytest.mark.quality
|
||
|
|
@pytest.mark.live
|
||
|
|
async def test_model_ab_comparison():
|
||
|
|
user_id = _get_test_user_id()
|
||
|
|
original_rows = _save_original_models()
|
||
|
|
|
||
|
|
scorecards: list[ModelScorecard] = []
|
||
|
|
try:
|
||
|
|
for model_code in CANDIDATE_MODELS:
|
||
|
|
_switch_model(model_code)
|
||
|
|
card = await _run_model_scenarios(model_code, user_id)
|
||
|
|
scorecards.append(card)
|
||
|
|
print(card.summary_table())
|
||
|
|
finally:
|
||
|
|
_restore_models(original_rows)
|
||
|
|
|
||
|
|
print("\n" + "=" * 60)
|
||
|
|
print("COMPARISON")
|
||
|
|
print("=" * 60)
|
||
|
|
for card in scorecards:
|
||
|
|
print(
|
||
|
|
f" {card.model_code:<20} "
|
||
|
|
f"overall={card.avg_overall:.2f} "
|
||
|
|
f"latency={card.avg_latency_ms:.0f}ms "
|
||
|
|
f"cost=${card.avg_cost_usd:.6f} "
|
||
|
|
f"tool_success={card.tool_success_rate:.0%}"
|
||
|
|
)
|
||
|
|
|
||
|
|
if len(scorecards) == 2:
|
||
|
|
a, b = scorecards
|
||
|
|
winner = a.model_code if a.avg_overall >= b.avg_overall else b.model_code
|
||
|
|
print(f"\n Winner: {winner} (by overall score)")
|