From 3dac5b95de2006c324139a2d8cf80defa9326b6c Mon Sep 17 00:00:00 2001
From: hub-gif <2487812171@qq.com>
Date: Wed, 22 Apr 2026 16:35:03 +0800
Subject: [PATCH] =?UTF-8?q?test(pipeline):=20=E7=9C=9F=E5=AE=9E=20run=5Fdi?=
 =?UTF-8?q?r=20=E4=B8=8A=E7=9F=A9=E9=98=B5/=E4=BF=83=E9=94=80=20LLM=20?=
 =?UTF-8?q?=E5=88=86=E5=9D=97=E8=BD=BD=E8=8D=B7=E7=BB=9F=E8=AE=A1=E4=B8=8E?=
 =?UTF-8?q?=E5=8F=AF=E9=80=89=E5=86=92=E7=83=9F?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

通过 MA_TEST_REAL_RUN_DIR 本地复现每细类请求规模；MA_TEST_REAL_LLM=1 可对首组发真实调用。不改动流水线。

Made-with: Cursor
---
 .../test_matrix_promo_llm_real_run_dir.py     | 139 ++++++++++++++++++
 1 file changed, 139 insertions(+)
 create mode 100644 backend/pipeline/tests/test_matrix_promo_llm_real_run_dir.py

diff --git a/backend/pipeline/tests/test_matrix_promo_llm_real_run_dir.py b/backend/pipeline/tests/test_matrix_promo_llm_real_run_dir.py
new file mode 100644
index 0000000..41934b3
--- /dev/null
+++ b/backend/pipeline/tests/test_matrix_promo_llm_real_run_dir.py
@@ -0,0 +1,139 @@
+"""
+矩阵 / 促销 两组 LLM 的「读超时」多为网关侧：单次请求在 ``LLM_CHAT_TIMEOUT``（默认 600s）
+内未返回完整响应即失败。默认配置下 ``llm_group_summaries_chunk_by_matrix=True``，
+会为**每个细类矩阵分组各发 1 次** ``chat/completions``，串行执行；任一次超时都会导致
+对应的 ``matrix_groups_llm.json`` / ``promo_groups_llm.json`` 记录 error。
+
+本模块**不修改流水线**，仅用于本地用真实合并表做载荷统计或可选冒烟调用。
+
+用法（在 ``backend`` 目录、已配置 Django / .env）::
+
+    # 仅统计每个分块请求的 JSON 字符数与粗算输入 token（不调网关）
+    set MA_TEST_REAL_RUN_DIR=D:\\...\\pipeline_runs\\20260413_104252_低GI
+    python -m pytest pipeline/tests/test_matrix_promo_llm_real_run_dir.py -v -s
+
+    # 额外对「第一个矩阵分组」发 1 次真实请求（需 OPENAI_* / LLM_* 可用）
+    set MA_TEST_REAL_LLM=1
+    python -m pytest pipeline/tests/test_matrix_promo_llm_real_run_dir.py::test_first_matrix_group_llm_smoke_optional -v -s
+"""
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+
+import pytest
+
+from pipeline.competitor_report import jd_report as jcr
+from pipeline.competitor_report.llm_group_payloads import (
+    build_matrix_groups_llm_payload,
+    build_promo_groups_llm_payload,
+)
+from pipeline.csv.schema import MERGED_FIELD_TO_CSV_HEADER
+
+
+def _real_run_dir() -> Path | None:
+    raw = (os.environ.get("MA_TEST_REAL_RUN_DIR") or "").strip()
+    if not raw:
+        return None
+    p = Path(raw).expanduser().resolve()
+    return p if p.is_dir() else None
+
+
+def _load_merged_rows(run_dir: Path) -> list[dict[str, str]]:
+    merged = run_dir / "keyword_pipeline_merged.csv"
+    if not merged.is_file():
+        pytest.skip(f"无合并表: {merged}")
+    _, rows = jcr._read_csv_rows(merged)
+    return rows
+
+
+def test_matrix_promo_chunk_payload_metrics_from_real_run_dir() -> None:
+    """
+    对真实 run_dir：输出矩阵 / 促销 **按细类分块** 时每一次请求的用户 JSON 规模与粗算 tokens。
+    用于对照 ``AI_crawler.chat_completion_text`` 的读超时（与输入长、输出 max_tokens 上限、网关排队均相关）。
+    """
+    rd = _real_run_dir()
+    if rd is None:
+        pytest.skip("请设置环境变量 MA_TEST_REAL_RUN_DIR 为流水线目录（含 keyword_pipeline_merged.csv）")
+
+    rows = _load_merged_rows(rd)
+    sku_h = MERGED_FIELD_TO_CSV_HEADER["sku_id"]
+    title_h = MERGED_FIELD_TO_CSV_HEADER["title"]
+    kw = (os.environ.get("MA_TEST_REAL_KEYWORD") or "低GI").strip() or "低GI"
+
+    pl_mx = build_matrix_groups_llm_payload(rows, sku_header=sku_h, title_h=title_h)
+    pl_po = build_promo_groups_llm_payload(rows, sku_header=sku_h, title_h=title_h)
+
+    assert pl_mx, "矩阵分组载荷为空（无可用细类或合并表无类目路径）"
+    assert pl_po, "促销分组载荷为空"
+
+    from pipeline.llm.generate_group_summaries import (
+        MATRIX_GROUPS_SYSTEM,
+        MATRIX_GROUPS_USER_PREFIX,
+        PROMO_GROUPS_SYSTEM,
+        PROMO_GROUPS_USER_PREFIX,
+    )
+    from pipeline.llm.llm_client import estimate_chat_input_tokens
+
+    def _per_chunk_stats(
+        *,
+        groups: list[dict],
+        system: str,
+        user_prefix: str,
+        label: str,
+    ) -> None:
+        max_chars = max_est = 0
+        worst_name = ""
+        for g in groups:
+            raw = json.dumps({"keyword": kw, "groups": [g]}, ensure_ascii=False)
+            user = user_prefix + raw
+            est = estimate_chat_input_tokens(system, user)
+            if len(raw) > max_chars:
+                max_chars = len(raw)
+                max_est = est
+                worst_name = str(g.get("group") or "")
+        n = len(groups)
+        print(
+            f"\n[{label}] 分块数={n}（默认 chunk 模式下串行请求数≈{n}）\n"
+            f"  单次 user JSON（含 keyword+单组）最大字符数≈{max_chars}\n"
+            f"  对应粗算输入 tokens≈{max_est}（与 estimate_chat_input_tokens 一致）\n"
+            f"  最大块细类名: {worst_name!r}\n"
+            f"  说明: 每块仍可能申请较大 max_tokens；网关生成慢或排队时，单次即可触发 Read timeout。\n"
+        )
+
+    _per_chunk_stats(
+        groups=pl_mx,
+        system=MATRIX_GROUPS_SYSTEM,
+        user_prefix=MATRIX_GROUPS_USER_PREFIX,
+        label="第五章·矩阵归纳 matrix",
+    )
+    _per_chunk_stats(
+        groups=pl_po,
+        system=PROMO_GROUPS_SYSTEM,
+        user_prefix=PROMO_GROUPS_USER_PREFIX,
+        label="第六章·促销归纳 promo",
+    )
+
+
+def test_first_matrix_group_llm_smoke_optional() -> None:
+    """可选：只对第一个矩阵分组调用 1 次网关，验证连通性（默认跳过）。"""
+    if (os.environ.get("MA_TEST_REAL_LLM") or "").strip() != "1":
+        pytest.skip("仅当 MA_TEST_REAL_LLM=1 时调用真实网关")
+
+    rd = _real_run_dir()
+    if rd is None:
+        pytest.skip("请设置 MA_TEST_REAL_RUN_DIR")
+
+    rows = _load_merged_rows(rd)
+    sku_h = MERGED_FIELD_TO_CSV_HEADER["sku_id"]
+    title_h = MERGED_FIELD_TO_CSV_HEADER["title"]
+    kw = (os.environ.get("MA_TEST_REAL_KEYWORD") or "低GI").strip() or "低GI"
+
+    pl_mx = build_matrix_groups_llm_payload(rows, sku_header=sku_h, title_h=title_h)
+    assert pl_mx
+
+    from pipeline.llm.generate_group_summaries import generate_matrix_group_summaries_llm
+
+    out = generate_matrix_group_summaries_llm([pl_mx[0]], keyword=kw)
+    assert (out or "").strip(), "模型返回为空"