wooo/ewoooc
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

讓 PChome feeder 採納人工覆核回饋
All checks were successful
CD Pipeline / deploy (push) Successful in 1m8s
Details

Browse Source
This commit is contained in:
OoO
2026-05-20 10:06:46 +08:00
parent 756b01af66
commit 45e1aad308
5 changed files with 159 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
TODO_NEXT_STEPS.txt
View File

@@ -4,6 +4,7 @@
================================================================================
【已完成】
- V10.305 將 PChome 人工覆核回饋接回 feeder：下一輪搜尋若命中已被 `reject_identity` 否決的同一候選，會記錄 `manual_rejected` 並跳過正式寫入；已被標記 `unit_price_required` 的候選只保留單位價比較，不寫入正式總價差；人工 `accept_identity` 可保守覆蓋低分門檻但會打 `manual_review/manual_accept` 標籤，讓核心比價閉環可被後續報表與簡報追蹤。
- V10.304 補 PChome 比價人工覆核決策閉環：新增 `competitor_match_reviews`、`/api/pchome-review/<sku>/decision` 與商品看板覆核列「採用同款 / 否決候選 / 標記單位價」動作；只有人工採用同款才寫入 `competitor_prices` + `competitor_price_history`，否決與單位價標記只追加 manual attempt 並關閉本輪覆核，避免錯配污染核心價差。
- V10.302 補 PChome 比價覆核匯出與診斷原因：`filter=pchome_review` 每筆覆核把 matcher `reasons=` 翻成品牌不符、商品線不符、容量差異、組合差異、需單位價、價差極端等可行動標籤；新增 `/api/export/excel/pchome-review` 匯出完整覆核隊列、人工處置、候選 PChome、單位價比較與原始診斷，避免核心比價只停在籠統「待對比」。
- V10.301 補市場情報 candidate queue review AI summary Telegram dispatch gate：新增 `/api/market_intel/manual_sample_review/candidate_queue_review_ai_summary_persistence_telegram_dispatch_gate` 與 UI 按鈕，在 summary persistence closeout 後檢查 Telegram 訊息契約、channel label、artifact path、token 外洩風險與後續 run package promotion；API/UI 仍不讀 approval/Telegram token、不呼叫 LLM、不開 DB、不寫檔、不派送 Telegram、不掛 scheduler。

2
config.py
View File

@@ -320,7 +320,7 @@ YOUTUBE_API_KEY = os.getenv('YOUTUBE_API_KEY', '')
# ==========================================
# 系統版本與路徑
# ==========================================
SYSTEM_VERSION = "V10.304"
SYSTEM_VERSION = "V10.305"
LOG_FILE_PATH = os.path.join(BASE_DIR, 'logs/system.log')
public_url = PUBLIC_URL # 用於模板顯示

4
docs/AI_INTELLIGENCE_MODULE_SOT.md
View File

@@ -2,7 +2,7 @@
> **最後更新**: 2026-05-20 (台北時間)
> **狀態**: 🟢 四 AI Agent 自動化閉環已落地；LLM 路由紅線升級為 Ollama-first 三主機級聯，Gemini 僅備援 / 鎖定場景
> **適用版本**: V10.304
> **適用版本**: V10.305
---
@@ -56,7 +56,7 @@ SQL漏斗(~300筆)
- 比對覆蓋率補強入口：`POST /api/ai/pchome-match/backfill`，優先補抓仍無有效 PChome 配對的高價 ACTIVE 商品，完成後自動重算 AI 挑品清單。
- 排程閉環：`run_pchome_match_backfill_task` 每日 10:30 執行，補抓 PChome 待比對商品、寫入歷史價格，再重算 `strategy='product_pick'` 清單。
- PChome / MOMO 競價摘要出口 `services/competitor_intel_repository.py` 使用 30 分鐘共享快取（`COMPETITOR_INTEL_CACHE_TTL_SECONDS` 可調），避免 `/growth_analysis`、`/daily_sales`、PPT/AI 報表每次請求重跑昂貴覆蓋率與價差趨勢查詢；`run_competitor_price_feeder_task` 與 PChome backfill 完成後會主動清除快取。快取只包摘要輸出，不改 matcher 的高信心門檻與 identity_v2 準確性規則。
- 商品看板第一屏：`/` 的 V2 看板直接以 `products`、`price_records`、`competitor_prices`、`competitor_match_attempts`、`competitor_match_reviews`、`ai_price_recommendations` 顯示比對覆蓋率、PChome 優勢、MOMO 威脅、AI 挑品、待比對優先清單與 PChome 覆核隊列；`filter=ai_picks` 可查看 50 品 AI 挑品列表，`filter=pchome_review` 可直接查看需人工處理的比價覆核 SKU，並以 DB 分頁支援 search/category/status 後的完整隊列，不得只截前 50 筆。覆核狀態篩選必須至少包含全部、需單位價、身份否決、低信心、價格過期與找不到同款，讓人工可依 matcher 診斷類型分批處理。列內顯示候選 PChome 商品、候選價、match score、單位價換算摘要、人工動作與 matcher 診斷原因標籤（品牌不符、商品線不符、容量差異、組合差異、需單位價、價差極端等），不得只顯示籠統「待比對」。`/api/export/excel/pchome-review` 必須匯出同一套覆核隊列、人工處置、候選 PChome、單位價比較與原始診斷，讓人工覆核、簡報與後續 AI 分析共用同一份證據。`/api/pchome-review/<sku>/decision` 是人工閉環入口：`accept_identity` 才可把候選寫入 `competitor_prices` 與 `competitor_price_history` 並打上 `manual_review/manual_accept/identity_v2`；`reject_identity` 與 `unit_price_required` 只寫 `competitor_match_reviews` 並追加 manual attempt，不得把不同販售組合或否決候選灌入正式價差。商品看板深度快取同時寫入 `data/dashboard_full_cache.pkl`，供多個 Gunicorn worker 共用，避免部署後各 worker 重複重建 7,000+ 商品統計造成開頁變慢；所有資料異動與 AI 挑品重算都透過 `clear_dashboard_cache()` 同步清除記憶體與共享快取，手動重算 API 會立即預熱商品看板快取，避免第一位使用者承擔重建成本。
- 商品看板第一屏：`/` 的 V2 看板直接以 `products`、`price_records`、`competitor_prices`、`competitor_match_attempts`、`competitor_match_reviews`、`ai_price_recommendations` 顯示比對覆蓋率、PChome 優勢、MOMO 威脅、AI 挑品、待比對優先清單與 PChome 覆核隊列；`filter=ai_picks` 可查看 50 品 AI 挑品列表，`filter=pchome_review` 可直接查看需人工處理的比價覆核 SKU，並以 DB 分頁支援 search/category/status 後的完整隊列，不得只截前 50 筆。覆核狀態篩選必須至少包含全部、需單位價、身份否決、低信心、價格過期與找不到同款，讓人工可依 matcher 診斷類型分批處理。列內顯示候選 PChome 商品、候選價、match score、單位價換算摘要、人工動作與 matcher 診斷原因標籤（品牌不符、商品線不符、容量差異、組合差異、需單位價、價差極端等），不得只顯示籠統「待比對」。`/api/export/excel/pchome-review` 必須匯出同一套覆核隊列、人工處置、候選 PChome、單位價比較與原始診斷，讓人工覆核、簡報與後續 AI 分析共用同一份證據。`/api/pchome-review/<sku>/decision` 是人工閉環入口：`accept_identity` 才可把候選寫入 `competitor_prices` 與 `competitor_price_history` 並打上 `manual_review/manual_accept/identity_v2`；`reject_identity` 與 `unit_price_required` 只寫 `competitor_match_reviews` 並追加 manual attempt，不得把不同販售組合或否決候選灌入正式價差。PChome feeder 後續搜尋同一候選時必須讀取 `competitor_match_reviews`：已否決候選寫 `manual_rejected` 並跳過正式寫入；已標記單位價候選寫 `manual_unit_price_required`；已採用候選可保守補到最低門檻並保留 `manual_review/manual_accept` 標籤。商品看板深度快取同時寫入 `data/dashboard_full_cache.pkl`，供多個 Gunicorn worker 共用，避免部署後各 worker 重複重建 7,000+ 商品統計造成開頁變慢；所有資料異動與 AI 挑品重算都透過 `clear_dashboard_cache()` 同步清除記憶體與共享快取，手動重算 API 會立即預熱商品看板快取，避免第一位使用者承擔重建成本。
| 角色 | 模型 | 主機 | 成本 | 每日限額 |
|------|------|------|------|---------|

93
services/competitor_price_feeder.py
View File

@@ -716,6 +716,37 @@ class CompetitorPriceFeeder:
f"incoming_score={match_score:.3f}",
)
def _fetch_latest_manual_review_for_candidate(
self,
sku: str,
competitor_product_id: str,
source: str = "pchome",
) -> Optional[dict]:
"""Read the latest human review for this exact candidate, if the table exists."""
if not competitor_product_id:
return None
from sqlalchemy import text
try:
with self.engine.connect() as conn:
row = conn.execute(text("""
SELECT review_action, review_reason, reviewer_identity, reviewed_at
FROM competitor_match_reviews
WHERE sku = :sku
AND source = :source
AND candidate_product_id = :candidate_id
ORDER BY reviewed_at DESC, id DESC
LIMIT 1
"""), {
"sku": sku,
"source": source,
"candidate_id": competitor_product_id,
}).mappings().first()
except Exception:
return None
return dict(row) if row else None
def _run_sku_items(self, skus: list, source: str = "pchome", label: str = "PChome 競品價格") -> FeederResult:
start = time.time()
@@ -777,7 +808,57 @@ class CompetitorPriceFeeder:
continue
best_product, score, diagnostics = result
if getattr(diagnostics, "comparison_mode", "") == "unit_comparable":
manual_review = self._fetch_latest_manual_review_for_candidate(
sku,
getattr(best_product, "product_id", None),
source=source,
)
manual_action = (manual_review or {}).get("review_action")
if manual_action == "reject_identity":
logger.info(
f"[Feeder] {sku} 候選已被人工否決，跳過正式寫入 | "
f"candidate={getattr(best_product, 'product_id', None)}"
)
self._record_match_attempt(
sku,
momo_name,
momo_product_id=momo_product_id,
momo_price=momo_price,
search_terms=search_terms,
candidate_count=len(products),
attempt_status="manual_rejected",
best_product=best_product,
best_score=score,
error_message=f"manual_review_rejected; {_format_match_diagnostics(diagnostics)}",
source=source,
)
attempts_written += 1
skipped_low += 1
continue
if manual_action == "unit_price_required":
logger.info(
f"[Feeder] {sku} 候選已被人工標記為單位價比較，不寫正式總價差 | "
f"candidate={getattr(best_product, 'product_id', None)}"
)
self._record_match_attempt(
sku,
momo_name,
momo_product_id=momo_product_id,
momo_price=momo_price,
search_terms=search_terms,
candidate_count=len(products),
attempt_status="manual_unit_price_required",
best_product=best_product,
best_score=score,
error_message=f"manual_review_unit_price_required; {_format_match_diagnostics(diagnostics)}",
source=source,
)
attempts_written += 1
skipped_low += 1
continue
manual_accept_override = manual_action == "accept_identity"
if getattr(diagnostics, "comparison_mode", "") == "unit_comparable" and not manual_accept_override:
logger.info(
f"[Feeder] {sku} 候選屬單位價可比但非同販售組合，"
f"不寫入正式價差 | {_format_match_diagnostics(diagnostics)}"
@@ -799,7 +880,7 @@ class CompetitorPriceFeeder:
skipped_low += 1
continue
if score < MIN_MATCH_SCORE:
if score < MIN_MATCH_SCORE and not manual_accept_override:
logger.debug(
f"[Feeder] {sku} 比對分數過低 ({score:.3f} < {MIN_MATCH_SCORE})，"
f"{_format_match_diagnostics(diagnostics)}"
@@ -821,10 +902,15 @@ class CompetitorPriceFeeder:
skipped_low += 1
continue
if manual_accept_override:
score = max(score, MIN_MATCH_SCORE)
tags = _extract_tags(best_product)
tags.extend(getattr(diagnostics, "tags", []))
for reason in getattr(diagnostics, "reasons", ()) or ():
tags.append(f"match_{reason}")
if manual_accept_override:
tags.extend(["manual_review", "manual_accept"])
tags = [tag for tag in tags if tag != "identity_veto"]
tags = list(dict.fromkeys(tags))
should_write, write_reason = self._should_upsert_competitor_price(
sku,
@@ -832,6 +918,9 @@ class CompetitorPriceFeeder:
score,
source=source,
)
if manual_accept_override and not should_write:
should_write = True
write_reason = "manual_accept_override"
if not should_write:
logger.info(f"[Feeder] {sku} 進入人工覆核，不覆蓋既有配對 | {write_reason}")
self._record_match_attempt(

64
tests/test_competitor_match_attempts_persistence.py
View File

@@ -59,6 +59,9 @@ def test_competitor_match_review_service_closes_human_review_loop():
assert "INSERT INTO competitor_price_history" in service_source
assert "manual_review" in service_source
assert "manual_accept" in service_source
assert "_fetch_latest_manual_review_for_candidate" in (ROOT / "services/competitor_price_feeder.py").read_text(encoding="utf-8")
assert "manual_review_rejected" in (ROOT / "services/competitor_price_feeder.py").read_text(encoding="utf-8")
assert "manual_accept_override" in (ROOT / "services/competitor_price_feeder.py").read_text(encoding="utf-8")
assert "CREATE TABLE IF NOT EXISTS competitor_match_reviews" in migration
assert "review_action" in migration
assert "reviewer_identity" in migration
@@ -68,6 +71,67 @@ def test_competitor_match_review_service_closes_human_review_loop():
assert "/api/pchome-review/" in dashboard_js
def test_competitor_feeder_respects_manual_rejected_candidate(monkeypatch):
from services.competitor_price_feeder import CompetitorPriceFeeder
from services.pchome_crawler import PChomeProduct
product = PChomeProduct(
product_id="DDAB01-REJECTED",
name="舒特膚 AD 乳液 200ml",
price=899,
original_price=999,
discount=10,
image_url="",
product_url="https://24h.pchome.com.tw/prod/DDAB01-REJECTED",
stock=20,
store="24h",
rating=4.7,
review_count=8,
is_on_sale=True,
crawled_at=datetime.now(),
)
class FakeCrawler:
def __init__(self, *_args, **_kwargs):
pass
def search_products(self, *_args, **_kwargs):
return True, "ok", [product]
monkeypatch.setattr("services.pchome_crawler.PChomeCrawler", FakeCrawler)
feeder = CompetitorPriceFeeder(engine=object())
attempts = []
writes = []
monkeypatch.setattr(
feeder,
"_fetch_latest_manual_review_for_candidate",
lambda *_args, **_kwargs: {"review_action": "reject_identity"},
)
monkeypatch.setattr(
feeder,
"_record_match_attempt",
lambda *args, **kwargs: attempts.append(kwargs),
)
monkeypatch.setattr(
feeder,
"_upsert_competitor_price",
lambda *args, **kwargs: writes.append((args, kwargs)),
)
result = feeder._run_sku_items([{
"sku": "A003",
"name": "舒特膚 AD 乳液 200ml",
"product_id": 3,
"momo_price": 980,
}])
assert result.matched == 0
assert result.skipped_low_score == 1
assert writes == []
assert attempts[0]["attempt_status"] == "manual_rejected"
assert "manual_review_rejected" in attempts[0]["error_message"]
def test_competitor_feeder_logs_keyword_parser_fallback(monkeypatch, caplog):
from services import competitor_price_feeder
from services import marketplace_product_matcher