File size: 9,655 Bytes
9c3ba60 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 | """Phase 6E — Live-Eval Harness CLI with opt-in /api/generate adapter.
Without ``NEXUS_RUN_LIVE_EVAL=true`` in the environment this script
**refuses** to make any real generation calls and exits non-zero. This is
the guard that keeps live evaluation out of the normal backend test gate.
Live evaluation is a thin orchestration layer:
1. POST to ``/api/generate`` with the prompt as ``topic``.
2. Poll ``/api/slides/{task_id}`` until 200 (done), failure, or timeout.
3. Shape the response into the deck dict expected by
:func:`services.eval_service.evaluate_deck`.
4. Write one JSON record per prompt under
``backend/storage/evals/`` (or ``$NEXUS_EVAL_OUTPUT_DIR``).
Network calls happen only via the injectable ``HttpClient`` protocol so
tests can exercise the orchestration without real HTTP.
Usage::
# Refuse — no env flag:
python -m scripts.run_live_eval --prompt-id biz-001
# Real run — explicit opt-in:
NEXUS_RUN_LIVE_EVAL=true python -m scripts.run_live_eval \\
--prompt-id biz-001 --base-url http://localhost:8000
"""
from __future__ import annotations
import argparse
import json
import os
import sys
import time
from datetime import datetime, timezone
from pathlib import Path
from typing import Any, Callable, Protocol
from services.eval_service import evaluate_deck, get_prompt_spec, load_prompts
DEFAULT_TIMEOUT_SECONDS = 600
DEFAULT_POLL_INTERVAL_SECONDS = 3.0
class HttpResponse(Protocol):
status_code: int
def json(self) -> Any: ...
class HttpClient(Protocol):
def post(self, url: str, *, json: dict[str, Any], timeout: float) -> HttpResponse: ...
def get(self, url: str, *, timeout: float) -> HttpResponse: ...
def close(self) -> None: ...
def _output_dir() -> Path:
base = Path(os.environ.get("NEXUS_EVAL_OUTPUT_DIR") or "/app/storage/evals")
base.mkdir(parents=True, exist_ok=True)
return base
def _refuse_if_not_opted_in() -> None:
flag = (os.environ.get("NEXUS_RUN_LIVE_EVAL") or "").strip().lower()
if flag != "true":
sys.stderr.write(
"[run_live_eval] refusing to run: NEXUS_RUN_LIVE_EVAL is not set to 'true'.\n"
" Set NEXUS_RUN_LIVE_EVAL=true to opt in to live calls.\n"
)
raise SystemExit(2)
def _build_default_client() -> HttpClient: # pragma: no cover - real network
import httpx
return httpx.Client() # type: ignore[return-value]
class LiveGenerationError(RuntimeError):
"""Raised for non-recoverable generation failures (HTTP error, task failed)."""
class LiveGenerationTimeout(RuntimeError):
"""Raised when polling exceeds the configured timeout."""
def run_live_generation(
prompt_text: str,
*,
base_url: str,
theme: str,
search_web: bool,
slide_count: int,
timeout_seconds: float,
poll_interval_seconds: float,
client: HttpClient,
sleep: Callable[[float], None] = time.sleep,
now: Callable[[], float] = time.monotonic,
min_sources: int = 0,
) -> dict[str, Any]:
"""Drive a single end-to-end generate→poll→fetch round-trip.
Returns a deck dict shaped like ``{"slides": [...], "sources": [...]}``
suitable for :func:`services.eval_service.evaluate_deck`.
Pure-orchestration: all HTTP goes through ``client`` so tests can
inject a fake. The default client is created by ``_build_default_client``.
"""
base = base_url.rstrip("/")
payload = {
"topic": prompt_text,
"slide_count": int(slide_count),
"theme": theme,
"search_web": bool(search_web),
}
# Phase 6U: forward expected min_sources so the agent loop's harvest
# step has an explicit target. Default 0 preserves the prior behaviour.
if int(min_sources or 0) > 0:
payload["min_sources"] = int(min_sources)
create_resp = client.post(f"{base}/api/generate", json=payload, timeout=30.0)
if create_resp.status_code >= 400:
raise LiveGenerationError(
f"POST /api/generate failed: HTTP {create_resp.status_code}"
)
body = create_resp.json()
task_id = body.get("task_id")
if not task_id:
raise LiveGenerationError(f"POST /api/generate missing task_id: {body!r}")
deadline = now() + float(timeout_seconds)
while True:
if now() >= deadline:
raise LiveGenerationTimeout(
f"task {task_id} did not complete within {timeout_seconds}s"
)
resp = client.get(f"{base}/api/slides/{task_id}", timeout=30.0)
if resp.status_code == 200:
data = resp.json()
slides = data.get("slides") or []
sources = data.get("sources") or []
return {"slides": slides, "sources": sources, "_task_id": task_id}
if resp.status_code == 409:
# Task still pending; /api/slides 409s until status==done. If
# the task transitions to "failed", /api/slides also 409s with
# a body.detail mentioning failed — surface that explicitly.
try:
detail = resp.json().get("detail", "")
except Exception:
detail = ""
if isinstance(detail, str) and "failed" in detail.lower():
raise LiveGenerationError(f"task {task_id} failed: {detail}")
sleep(float(poll_interval_seconds))
continue
if resp.status_code == 404:
raise LiveGenerationError(f"task {task_id} not found")
raise LiveGenerationError(
f"GET /api/slides/{task_id} failed: HTTP {resp.status_code}"
)
def _write_record(record: dict[str, Any]) -> Path:
out_dir = _output_dir()
ts = datetime.now(timezone.utc).strftime("%Y%m%dT%H%M%SZ")
path = out_dir / f"{record['prompt_id']}-{ts}.json"
path.write_text(json.dumps(record, indent=2, sort_keys=True), encoding="utf-8")
return path
def main(
argv: list[str] | None = None,
*,
client_factory: Callable[[], HttpClient] | None = None,
write_record: Callable[[dict[str, Any]], Path] = _write_record,
) -> int:
parser = argparse.ArgumentParser(description="NEXUS live-eval harness (opt-in).")
parser.add_argument(
"--prompt-id",
required=False,
help="Run a single prompt id from benchmarks/prompts.json. Omit to run all.",
)
parser.add_argument(
"--base-url",
default="http://localhost:8000",
help="Backend base URL for /api/generate.",
)
parser.add_argument(
"--timeout-seconds",
type=float,
default=DEFAULT_TIMEOUT_SECONDS,
help="Max wait per prompt before giving up.",
)
parser.add_argument(
"--poll-interval-seconds",
type=float,
default=DEFAULT_POLL_INTERVAL_SECONDS,
help="Seconds between /api/slides polls.",
)
parser.add_argument(
"--theme",
default="Editorial",
help="Theme to pass to /api/generate.",
)
search_group = parser.add_mutually_exclusive_group()
search_group.add_argument(
"--search-web",
dest="search_web",
action="store_true",
help="Pass search_web=true to /api/generate (default).",
)
search_group.add_argument(
"--no-search-web",
dest="search_web",
action="store_false",
help="Pass search_web=false to /api/generate.",
)
parser.set_defaults(search_web=True)
parser.add_argument(
"--slide-count",
type=int,
default=8,
help="slide_count to pass to /api/generate.",
)
args = parser.parse_args(argv)
_refuse_if_not_opted_in()
prompts = load_prompts()
if args.prompt_id:
prompts = [get_prompt_spec(args.prompt_id, prompts)]
client = (client_factory or _build_default_client)()
written: list[Path] = []
failures: list[tuple[str, str]] = []
try:
for spec in prompts:
# Phase 6U: prefer the per-prompt evidence target so each
# prompt's required source count drives the harvest depth.
evidence = spec.get("expected_evidence") or {}
spec_min_sources = 0
try:
spec_min_sources = int(evidence.get("min_sources", 0) or 0)
except (TypeError, ValueError):
spec_min_sources = 0
try:
deck = run_live_generation(
spec["prompt"],
base_url=args.base_url,
theme=args.theme,
search_web=args.search_web,
slide_count=args.slide_count,
timeout_seconds=args.timeout_seconds,
poll_interval_seconds=args.poll_interval_seconds,
client=client,
min_sources=spec_min_sources,
)
except (LiveGenerationError, LiveGenerationTimeout) as exc:
failures.append((spec["id"], str(exc)))
sys.stderr.write(f"[run_live_eval] {spec['id']} FAILED: {exc}\n")
continue
record = evaluate_deck(deck, spec, ran_live=True)
written.append(write_record(record))
finally:
try:
client.close()
except Exception:
pass
for path in written:
sys.stdout.write(f"wrote {path}\n")
if failures:
sys.stderr.write(
f"[run_live_eval] {len(failures)} prompt(s) failed: "
+ ", ".join(f"{pid}={err}" for pid, err in failures)
+ "\n"
)
return 1
return 0
if __name__ == "__main__": # pragma: no cover
raise SystemExit(main())
|