Move workflow evaluations to standalone runner
This commit is contained in:
48
evals/README.md
Normal file
48
evals/README.md
Normal file
@@ -0,0 +1,48 @@
|
||||
# Workflow evaluations
|
||||
|
||||
The evaluation runner calls a live FastAPI service backed by FastGPT. For now,
|
||||
each conversation step checks only the SSE `stage_code.nextStageCode` value.
|
||||
|
||||
## Case format
|
||||
|
||||
Add YAML files anywhere under `evals/cases/`:
|
||||
|
||||
```yaml
|
||||
name: browser_addon_example
|
||||
clientMode: browser_addon
|
||||
|
||||
steps:
|
||||
- input: hi
|
||||
expect:
|
||||
stageCode: browser_addon.1001
|
||||
|
||||
- input: 继续
|
||||
expect:
|
||||
stageCode: browser_addon.1002
|
||||
```
|
||||
|
||||
Steps in one file share the same generated `sessionId`. Different cases always
|
||||
use different session IDs.
|
||||
|
||||
## Run
|
||||
|
||||
Start the API, then execute:
|
||||
|
||||
```bash
|
||||
python evals/run.py
|
||||
```
|
||||
|
||||
Use another API address:
|
||||
|
||||
```bash
|
||||
python evals/run.py --base-url http://192.168.1.10:8000
|
||||
```
|
||||
|
||||
Run selected cases:
|
||||
|
||||
```bash
|
||||
python evals/run.py --case browser_addon
|
||||
```
|
||||
|
||||
The process exits with code `1` when an expectation fails and code `2` when
|
||||
case configuration is invalid.
|
||||
7
evals/cases/browser_addon/initial.yaml
Normal file
7
evals/cases/browser_addon/initial.yaml
Normal file
@@ -0,0 +1,7 @@
|
||||
name: browser_addon_initial_stage
|
||||
clientMode: browser_addon
|
||||
|
||||
steps:
|
||||
- input: hi
|
||||
expect:
|
||||
stageCode: browser_addon.1001
|
||||
7
evals/cases/direct/initial.yaml
Normal file
7
evals/cases/direct/initial.yaml
Normal file
@@ -0,0 +1,7 @@
|
||||
name: direct_initial_stage
|
||||
clientMode: direct
|
||||
|
||||
steps:
|
||||
- input: hi
|
||||
expect:
|
||||
stageCode: "1001"
|
||||
248
evals/run.py
Normal file
248
evals/run.py
Normal file
@@ -0,0 +1,248 @@
|
||||
#!/usr/bin/env python3
|
||||
"""Run live FastGPT workflow evaluations against the FastAPI service."""
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import os
|
||||
import re
|
||||
import sys
|
||||
from dataclasses import dataclass
|
||||
from datetime import datetime, timezone
|
||||
from pathlib import Path
|
||||
from uuid import uuid4
|
||||
|
||||
import httpx
|
||||
import yaml
|
||||
|
||||
|
||||
EVALS_DIR = Path(__file__).resolve().parent
|
||||
DEFAULT_CASES_DIR = EVALS_DIR / "cases"
|
||||
|
||||
|
||||
@dataclass
|
||||
class ChatResult:
|
||||
stage_codes: list[str]
|
||||
errors: list[dict]
|
||||
|
||||
|
||||
def timestamp() -> str:
|
||||
return datetime.now(timezone.utc).isoformat()
|
||||
|
||||
|
||||
def load_cases(cases_dir: Path, name_filters: list[str]) -> list[dict]:
|
||||
cases = []
|
||||
for path in sorted(cases_dir.rglob("*.yaml")):
|
||||
case = yaml.safe_load(path.read_text(encoding="utf-8"))
|
||||
if not isinstance(case, dict):
|
||||
raise ValueError(f"{path}: case must be a YAML object")
|
||||
|
||||
name = case.get("name")
|
||||
client_mode = case.get("clientMode")
|
||||
steps = case.get("steps")
|
||||
if not isinstance(name, str) or not name:
|
||||
raise ValueError(f"{path}: name is required")
|
||||
if client_mode not in {"direct", "browser_addon"}:
|
||||
raise ValueError(f"{path}: unsupported clientMode={client_mode!r}")
|
||||
if not isinstance(steps, list) or not steps:
|
||||
raise ValueError(f"{path}: at least one step is required")
|
||||
|
||||
for index, step in enumerate(steps, start=1):
|
||||
expected_code = (
|
||||
step.get("expect", {}).get("stageCode")
|
||||
if isinstance(step, dict)
|
||||
else None
|
||||
)
|
||||
if not isinstance(step, dict) or not isinstance(step.get("input"), str):
|
||||
raise ValueError(f"{path}: step {index} input must be a string")
|
||||
if not isinstance(expected_code, str) or not expected_code:
|
||||
raise ValueError(
|
||||
f"{path}: step {index} expect.stageCode is required"
|
||||
)
|
||||
|
||||
if name_filters and not any(value in name for value in name_filters):
|
||||
continue
|
||||
|
||||
case["_path"] = str(path.relative_to(EVALS_DIR))
|
||||
cases.append(case)
|
||||
|
||||
if not cases:
|
||||
raise ValueError(f"No evaluation cases found under {cases_dir}")
|
||||
return cases
|
||||
|
||||
|
||||
def parse_sse(response: httpx.Response) -> ChatResult:
|
||||
stage_codes: list[str] = []
|
||||
errors: list[dict] = []
|
||||
event_name = ""
|
||||
data_lines: list[str] = []
|
||||
|
||||
def consume_event() -> None:
|
||||
nonlocal event_name, data_lines
|
||||
if data_lines:
|
||||
payload = json.loads("\n".join(data_lines))
|
||||
if event_name == "stage_code":
|
||||
stage_codes.append(payload["nextStageCode"])
|
||||
elif event_name == "error":
|
||||
errors.append(payload)
|
||||
event_name = ""
|
||||
data_lines = []
|
||||
|
||||
for line in response.iter_lines():
|
||||
if not line:
|
||||
consume_event()
|
||||
elif line.startswith("event:"):
|
||||
event_name = line.removeprefix("event:").strip()
|
||||
elif line.startswith("data:"):
|
||||
data_lines.append(line.removeprefix("data:").strip())
|
||||
|
||||
consume_event()
|
||||
return ChatResult(stage_codes=stage_codes, errors=errors)
|
||||
|
||||
|
||||
def run_chat_step(
|
||||
client: httpx.Client,
|
||||
*,
|
||||
session_id: str,
|
||||
client_mode: str,
|
||||
text: str,
|
||||
) -> ChatResult:
|
||||
with client.stream(
|
||||
"POST",
|
||||
"/chat",
|
||||
params={"stream": "true"},
|
||||
json={
|
||||
"sessionId": session_id,
|
||||
"timeStamp": timestamp(),
|
||||
"text": text,
|
||||
"clientMode": client_mode,
|
||||
"needFormUpdate": False,
|
||||
"useTextChunk": False,
|
||||
},
|
||||
) as response:
|
||||
response.raise_for_status()
|
||||
return parse_sse(response)
|
||||
|
||||
|
||||
def cleanup_session(client: httpx.Client, session_id: str) -> str | None:
|
||||
try:
|
||||
response = client.request(
|
||||
"DELETE",
|
||||
"/delete_session",
|
||||
json={"sessionId": session_id, "timeStamp": timestamp()},
|
||||
)
|
||||
response.raise_for_status()
|
||||
payload = response.json()
|
||||
if payload.get("code") != "200":
|
||||
return f"cleanup response={payload}"
|
||||
except (httpx.HTTPError, ValueError) as exc:
|
||||
return f"cleanup error={exc}"
|
||||
return None
|
||||
|
||||
|
||||
def session_id_for(case_name: str) -> str:
|
||||
slug = re.sub(r"[^a-zA-Z0-9_-]+", "-", case_name).strip("-")
|
||||
return f"eval-{slug[:24]}-{uuid4().hex[:12]}"
|
||||
|
||||
|
||||
def run_case(client: httpx.Client, case: dict) -> tuple[bool, list[str]]:
|
||||
session_id = session_id_for(case["name"])
|
||||
messages = []
|
||||
passed = True
|
||||
|
||||
try:
|
||||
for step_number, step in enumerate(case["steps"], start=1):
|
||||
expected = step["expect"]["stageCode"]
|
||||
try:
|
||||
result = run_chat_step(
|
||||
client,
|
||||
session_id=session_id,
|
||||
client_mode=case["clientMode"],
|
||||
text=step["input"],
|
||||
)
|
||||
except (httpx.HTTPError, json.JSONDecodeError, KeyError) as exc:
|
||||
messages.append(f"step {step_number}: request error: {exc}")
|
||||
passed = False
|
||||
break
|
||||
|
||||
if result.stage_codes != [expected]:
|
||||
messages.append(
|
||||
f"step {step_number}: expected={[expected]} "
|
||||
f"actual={result.stage_codes} sse_errors={result.errors}"
|
||||
)
|
||||
passed = False
|
||||
break
|
||||
|
||||
messages.append(f"step {step_number}: stageCode={expected}")
|
||||
finally:
|
||||
cleanup_warning = cleanup_session(client, session_id)
|
||||
if cleanup_warning:
|
||||
messages.append(f"warning: {cleanup_warning}")
|
||||
|
||||
return passed, messages
|
||||
|
||||
|
||||
def parse_args() -> argparse.Namespace:
|
||||
parser = argparse.ArgumentParser(
|
||||
description="Run live workflow evaluations and assert stage codes."
|
||||
)
|
||||
parser.add_argument(
|
||||
"--base-url",
|
||||
default=os.getenv("EVAL_BASE_URL", "http://127.0.0.1:8000"),
|
||||
help="FastAPI base URL (default: EVAL_BASE_URL or http://127.0.0.1:8000)",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--cases-dir",
|
||||
type=Path,
|
||||
default=DEFAULT_CASES_DIR,
|
||||
help="Directory containing YAML cases",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--case",
|
||||
action="append",
|
||||
default=[],
|
||||
help="Run cases whose names contain this value; repeatable",
|
||||
)
|
||||
parser.add_argument("--timeout", type=float, default=90.0)
|
||||
parser.add_argument(
|
||||
"--insecure",
|
||||
action="store_true",
|
||||
help="Disable TLS certificate verification",
|
||||
)
|
||||
return parser.parse_args()
|
||||
|
||||
|
||||
def main() -> int:
|
||||
args = parse_args()
|
||||
try:
|
||||
cases = load_cases(args.cases_dir, args.case)
|
||||
except (OSError, ValueError, yaml.YAMLError) as exc:
|
||||
print(f"ERROR: {exc}", file=sys.stderr)
|
||||
return 2
|
||||
|
||||
passed_count = 0
|
||||
failed_count = 0
|
||||
with httpx.Client(
|
||||
base_url=args.base_url.rstrip("/"),
|
||||
timeout=args.timeout,
|
||||
verify=not args.insecure,
|
||||
) as client:
|
||||
for case in cases:
|
||||
passed, messages = run_case(client, case)
|
||||
status = "PASS" if passed else "FAIL"
|
||||
print(f"{status} {case['name']} ({case['_path']})")
|
||||
for message in messages:
|
||||
print(f" {message}")
|
||||
if passed:
|
||||
passed_count += 1
|
||||
else:
|
||||
failed_count += 1
|
||||
|
||||
print(
|
||||
f"\nSummary: total={len(cases)} "
|
||||
f"passed={passed_count} failed={failed_count}"
|
||||
)
|
||||
return 1 if failed_count else 0
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(main())
|
||||
@@ -2,4 +2,3 @@
|
||||
pythonpath = .
|
||||
markers =
|
||||
integration: marks tests that need access to external services
|
||||
golden_live: runs golden conversation cases against a live API and FastGPT
|
||||
|
||||
@@ -20,3 +20,4 @@ pandas
|
||||
requests
|
||||
sqlalchemy
|
||||
pymysql
|
||||
PyYAML>=6.0
|
||||
|
||||
@@ -1,32 +0,0 @@
|
||||
# Stage-code golden tests
|
||||
|
||||
These tests call the running FastAPI service and its configured FastGPT
|
||||
workflow. Each conversation step asserts only the `nextStageCode` carried by
|
||||
the SSE `stage_code` event.
|
||||
|
||||
Edit `cases.json` to add or update conversation cases:
|
||||
|
||||
```json
|
||||
{
|
||||
"name": "example",
|
||||
"clientMode": "browser_addon",
|
||||
"steps": [
|
||||
{
|
||||
"text": "继续",
|
||||
"expect": {
|
||||
"stageCode": "browser_addon.1002"
|
||||
}
|
||||
}
|
||||
]
|
||||
}
|
||||
```
|
||||
|
||||
Start the API, then run:
|
||||
|
||||
```bash
|
||||
RUN_GOLDEN=1 \
|
||||
GOLDEN_BASE_URL=http://127.0.0.1:8000 \
|
||||
pytest -m golden_live test/golden -v
|
||||
```
|
||||
|
||||
Without `RUN_GOLDEN=1`, the cases are skipped and do not make network calls.
|
||||
@@ -1,26 +0,0 @@
|
||||
[
|
||||
{
|
||||
"name": "direct_initial_stage",
|
||||
"clientMode": "direct",
|
||||
"steps": [
|
||||
{
|
||||
"text": "hi",
|
||||
"expect": {
|
||||
"stageCode": "1001"
|
||||
}
|
||||
}
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "browser_addon_initial_stage",
|
||||
"clientMode": "browser_addon",
|
||||
"steps": [
|
||||
{
|
||||
"text": "hi",
|
||||
"expect": {
|
||||
"stageCode": "browser_addon.1001"
|
||||
}
|
||||
}
|
||||
]
|
||||
}
|
||||
]
|
||||
@@ -1,112 +0,0 @@
|
||||
import json
|
||||
import os
|
||||
from datetime import datetime, timezone
|
||||
from pathlib import Path
|
||||
from uuid import uuid4
|
||||
|
||||
import httpx
|
||||
import pytest
|
||||
|
||||
|
||||
CASES_PATH = Path(__file__).with_name("cases.json")
|
||||
RUN_GOLDEN = os.getenv("RUN_GOLDEN") == "1"
|
||||
BASE_URL = os.getenv("GOLDEN_BASE_URL", "http://127.0.0.1:8000").rstrip("/")
|
||||
|
||||
|
||||
def load_cases() -> list[dict]:
|
||||
return json.loads(CASES_PATH.read_text(encoding="utf-8"))
|
||||
|
||||
|
||||
def parse_stage_codes(response: httpx.Response) -> list[str]:
|
||||
stage_codes: list[str] = []
|
||||
event_name = ""
|
||||
data_lines: list[str] = []
|
||||
|
||||
def consume_event() -> None:
|
||||
nonlocal event_name, data_lines
|
||||
if event_name == "stage_code" and data_lines:
|
||||
payload = json.loads("\n".join(data_lines))
|
||||
stage_codes.append(payload["nextStageCode"])
|
||||
event_name = ""
|
||||
data_lines = []
|
||||
|
||||
for line in response.iter_lines():
|
||||
if not line:
|
||||
consume_event()
|
||||
elif line.startswith("event:"):
|
||||
event_name = line.removeprefix("event:").strip()
|
||||
elif line.startswith("data:"):
|
||||
data_lines.append(line.removeprefix("data:").strip())
|
||||
|
||||
consume_event()
|
||||
return stage_codes
|
||||
|
||||
|
||||
def run_chat_step(
|
||||
client: httpx.Client,
|
||||
*,
|
||||
session_id: str,
|
||||
client_mode: str,
|
||||
text: str,
|
||||
) -> list[str]:
|
||||
payload = {
|
||||
"sessionId": session_id,
|
||||
"timeStamp": datetime.now(timezone.utc).isoformat(),
|
||||
"text": text,
|
||||
"clientMode": client_mode,
|
||||
"needFormUpdate": False,
|
||||
"useTextChunk": False,
|
||||
}
|
||||
|
||||
with client.stream(
|
||||
"POST",
|
||||
"/chat",
|
||||
params={"stream": "true"},
|
||||
json=payload,
|
||||
) as response:
|
||||
response.raise_for_status()
|
||||
return parse_stage_codes(response)
|
||||
|
||||
|
||||
def cleanup_session(client: httpx.Client, session_id: str) -> None:
|
||||
try:
|
||||
client.request(
|
||||
"DELETE",
|
||||
"/delete_session",
|
||||
json={
|
||||
"sessionId": session_id,
|
||||
"timeStamp": datetime.now(timezone.utc).isoformat(),
|
||||
},
|
||||
)
|
||||
except httpx.HTTPError:
|
||||
# A unique session ID is used for every run, so cleanup failure must not
|
||||
# hide the stage-code assertion that the golden test is responsible for.
|
||||
pass
|
||||
|
||||
|
||||
@pytest.mark.golden_live
|
||||
@pytest.mark.skipif(
|
||||
not RUN_GOLDEN,
|
||||
reason="Set RUN_GOLDEN=1 to run live golden cases",
|
||||
)
|
||||
@pytest.mark.parametrize("case", load_cases(), ids=lambda case: case["name"])
|
||||
def test_expected_stage_codes(case):
|
||||
session_id = f"golden-{case['name']}-{uuid4().hex}"
|
||||
|
||||
with httpx.Client(base_url=BASE_URL, timeout=90.0) as client:
|
||||
try:
|
||||
for step_number, step in enumerate(case["steps"], start=1):
|
||||
actual_codes = run_chat_step(
|
||||
client,
|
||||
session_id=session_id,
|
||||
client_mode=case["clientMode"],
|
||||
text=step["text"],
|
||||
)
|
||||
expected_code = step["expect"]["stageCode"]
|
||||
|
||||
assert actual_codes == [expected_code], (
|
||||
f"case={case['name']} step={step_number} "
|
||||
f"expected={[expected_code]} actual={actual_codes}"
|
||||
)
|
||||
finally:
|
||||
cleanup_session(client, session_id)
|
||||
Reference in New Issue
Block a user