diff --git a/.env.example b/.env.example index 8af31fb..197359e 100644 --- a/.env.example +++ b/.env.example @@ -127,7 +127,7 @@ GDRIVE_FILE_PATTERN=即時業績_當日 # ========================================== # Hermes 3 競價情報分析（Module 2 / ADR-012） # ========================================== -# [選填] Hermes Ollama 端點；留空時自動走 GCP-A → GCP-B → 111（ADR-028） +# [選填] Hermes Ollama 端點；留空時自動走 GCP-A → GCP-B（111 預設不承接 Hermes 批量分析） # 僅允許 http://34.143.170.20:11434、http://34.21.145.224:11434、http://192.168.0.111:11434 HERMES_URL= @@ -135,6 +135,8 @@ HERMES_URL= HERMES_TIMEOUT=120 # [預設 5m] Hermes runner 熱駐留；禁止回到 24h，避免 GCP-B/111 長駐高負載 HERMES_KEEP_ALIVE=5m +# [預設 false] 僅救急時才允許 Hermes LLM 落到 111；平時失敗交給規則/DB fallback +HERMES_ALLOW_111_FALLBACK=false # [選填] Embedding 服務主機；留空時自動走同一條 Ollama 三主機級聯 EMBEDDING_HOST= diff --git a/config.py b/config.py index 3c6cbbe..6340787 100644 --- a/config.py +++ b/config.py @@ -325,7 +325,7 @@ YOUTUBE_API_KEY = os.getenv('YOUTUBE_API_KEY', '') # ========================================== # 系統版本與路徑 # ========================================== -SYSTEM_VERSION = "V10.414" +SYSTEM_VERSION = "V10.415" LOG_FILE_PATH = os.path.join(BASE_DIR, 'logs/system.log') public_url = PUBLIC_URL # 用於模板顯示 diff --git a/docs/AI_INTELLIGENCE_MODULE_SOT.md b/docs/AI_INTELLIGENCE_MODULE_SOT.md index 897154b..a87951b 100644 --- a/docs/AI_INTELLIGENCE_MODULE_SOT.md +++ b/docs/AI_INTELLIGENCE_MODULE_SOT.md @@ -13,7 +13,7 @@ - Gemini 只能作為 Ollama 主路徑失敗後的備援；MCP Grounding、PPT/vision、週/月報、Code Review、EA HITL、複雜 SKU 升級等舊鎖定場景也必須先走 GCP-A → GCP-B → 111。 - 188 `192.168.0.188` 僅是 App / DB / scheduler / Telegram bot 容器宿主與 AutoHeal target，不可作為 Ollama 節點。 - 通用 AI 文案、關鍵字、商品洞察與 Telegram Q&A 第一響應不得 Gemini-first。 -- Hermes intent / analyst 路徑不得手刻 `/api/generate` 或只 resolve 單次 host；必須走 `OllamaService`，讓同一請求可依序 retry GCP-A → GCP-B → 111。 +- Hermes intent / analyst 路徑不得手刻 `/api/generate` 或只 resolve 單次 host；必須走 `OllamaService`。預設 `HERMES_ALLOW_111_FALLBACK=false`，同一請求只跑 GCP-A → GCP-B；兩台都失敗時回規則引擎或 DB 證據 fallback，不把批量價格分析轉嫁到 111。救急時才可顯式設 true 允許 111 接手。 - NemoTron qwen3 dispatch 的 `/api/chat` tool-calling 路徑也必須同一請求最多嘗試三台 Ollama，第一台失敗要 `mark_unhealthy()` 後再試下一台，最後才 fallback NIM。 - PPT vision、PPT 文案 final fallback、MCP 離線 final fallback 等特殊 Ollama 路徑也不得只打單一 host；如需 `/api/generate`，一律透過 `OllamaService.generate()`。 - Code Review pipeline 也必須 Ollama-first：Hermes scan 與 OpenClaw assessment 都走 `OllamaService` 三主機 retry；Gemini telemetry 只能以 `code_review_openclaw_gemini` 出現，表示 Ollama/可選 Claude 備援都失敗後才啟用。 @@ -28,7 +28,7 @@ - `docker-compose.yml` 的 `momo-app`、`scheduler`、`telegram-bot` 必須明確設定 `GEMINI_API_HARD_DISABLED=${GEMINI_API_HARD_DISABLED:-true}` 與 `GEMINI_FALLBACK_ENABLED=${GEMINI_FALLBACK_ENABLED:-false}`；`.env` 可保留 `GEMINI_API_KEY`，但不得因 key 存在就讓核心容器產生 Gemini 付費出站。 - Gemini 不可被任何狀態面板或 router 推薦為主提供者：`AIProviderService._get_recommended_provider()` 不得回傳 `gemini`，只能顯示為 fallback 狀態；`llm_model_router` 的 `ea_engine` 若收到 `gemini-*` default 必須改回 `hermes3:latest`，需要深推理時才升本地 `deepseek-r1:14b`。 - ElephantAlpha prompt / agent registry 不得再把 OpenClaw 描述為 Gemini 主模型；OpenClaw 是 `qwen2.5-coder:7b` / `qwen3:14b` Ollama-first 策略師，Gemini 僅能在 guard 顯式解鎖後作 emergency fallback。 -- 111 `192.168.0.111` 只是最後一道 Mac fallback，不承接 7B+、vision、long-context 模型長駐；`OllamaService.generate()` 落到 111 時會將 `qwen3`、`deepseek-r1`、`hermes3`、`qwen2.5*`、`gemma3`、`llava`、`minicpm-v` 與 7B+ 模型依 `OLLAMA_111_MODEL_DOWNGRADE_PATTERNS` 降級到 `OLLAMA_111_MODEL_FALLBACK=llama3.2:latest`，並以 `OLLAMA_111_KEEP_ALIVE=5m`、`OLLAMA_111_MAX_TIMEOUT=20`、`OLLAMA_111_NUM_CTX=4096`、`OLLAMA_111_NUM_PREDICT=512` 封頂。Hermes / OpenClaw 報告型路徑的業務 keep-alive 也預設 `5m`；Code Review 另以 `CODE_REVIEW_ALLOW_111_FALLBACK=false` 預設跳過 111，避免 16GB RAM 主機與 GCP-B 被長駐 runner、長輸出與 24h keep-alive 壓到高 load。 +- 111 `192.168.0.111` 只是最後一道 Mac fallback，不承接 7B+、vision、long-context 模型長駐；`OllamaService.generate()` 落到 111 時會將 `qwen3`、`deepseek-r1`、`hermes3`、`qwen2.5*`、`gemma3`、`llava`、`minicpm-v` 與 7B+ 模型依 `OLLAMA_111_MODEL_DOWNGRADE_PATTERNS` 降級到 `OLLAMA_111_MODEL_FALLBACK=llama3.2:latest`，並以 `OLLAMA_111_KEEP_ALIVE=5m`、`OLLAMA_111_MAX_TIMEOUT=20`、`OLLAMA_111_NUM_CTX=4096`、`OLLAMA_111_NUM_PREDICT=512` 封頂。OpenClaw 報告型路徑的業務 keep-alive 預設 `5m`；Code Review 以 `CODE_REVIEW_ALLOW_111_FALLBACK=false`、Hermes 以 `HERMES_ALLOW_111_FALLBACK=false` 預設跳過 111，避免 16GB RAM 主機與 GCP-B 被長駐 runner、長輸出與 24h keep-alive 壓到高 load。 - ElephantAlpha 的 `price_drop_alert` / `market_opportunity` Telegram HITL 告警必須把同款證據獨立呈現，至少包含 `match_type`、`price_basis`、`alert_tier` 與 `match_score`；沒有高信心同款與總價可比證據時，不得把 PChome/MOMO 價差寫成可直接跟價建議。 ## 一、四 AI Agent 路由架構 diff --git a/docs/memory/history_logs.md b/docs/memory/history_logs.md index 0e0c477..ac98bcc 100644 --- a/docs/memory/history_logs.md +++ b/docs/memory/history_logs.md @@ -13,6 +13,7 @@ ## 📅 詳細更新日誌 (考古存檔) ### 2026-05-24：PChome 近門檻身份回收第二輪 +- **V10.415 Hermes 預設不落 111 + 比對保護**: `OllamaService.generate()` 新增 `allow_111_fallback` 參數，預設維持三主機相容；Hermes intent / competitor analyst 改以 `HERMES_ALLOW_111_FALLBACK=false` 預設只跑 GCP-A → GCP-B，兩台都不可用時交給規則引擎或 DB 證據 fallback，不再把批量價格分析與意圖分類轉嫁到 111。同版 marketplace matcher 將防曬類列入 variant-sensitive，排除 SPF/PA/UVA/UVB 這類規格 token 被誤當型號，避免「兒童防曬乳」與「海洋友善保濕防曬乳」誤配；Recipe Box 兒童防曬氣墊粉餅保留精準同品線例外；另新增 `pack_quantity_difference`，讓 Beauty Foot 足膜 5入 vs 4入走 unit comparable，不再卡在低信心。 - **V10.414 MCP fetch run readiness gate**: 新增 `mcp_fetch_run_readiness` read-only builder、GET/POST endpoint、UI run readiness 審核面板與 deployment readiness smoke target，在 run package 後檢查 command preview、receipt path、artifact path、節流/timeout/dry-run-first 與操作員 shell-only 邊界；API/UI 不執行 CLI、不抓外站、不寫檔、不開 DB、不掛 scheduler，只放行到人工 shell dry-run 與後續 receipt gate。 - **V10.413 Code Review 預設保護 111 fallback**: production `ai_calls` 顯示 GCP-A 不可達時，Code Review OpenClaw 會先耗掉 primary timeout，再讓 GCP-B 撐到 60s，最後落到 111 `llama3.2` 成功，造成 111 與 GCP-B 高負載。新增 `CODE_REVIEW_ALLOW_111_FALLBACK=false` 預設：Code Review 的 Hermes LLM scan / OpenClaw assessment 只跑 GCP-A → GCP-B；只有明確設 true 才把部署後重分析丟給 111。若 GCP-A/GCP-B 都失敗且 Claude/Gemini 未顯式開啟，改回 deterministic 本地降級摘要，不呼叫 Gemini，也不再用 111 承接非即時重分析。 - **V10.412 MCP fetch run package gate**: 新增 `mcp_fetch_run_package` read-only builder、獨立 route extension、GET/POST endpoint、UI run package 審核面板與 deployment readiness smoke target，將已通過的 target review 轉成操作員可覆核的 command argv preview 與 receipt path 契約；API/UI 不執行 CLI、不抓外站、不寫檔、不開 DB、不掛 scheduler，只放行到後續 run readiness review。 diff --git a/services/hermes_analyst_service.py b/services/hermes_analyst_service.py index 7f9bcd8..0a257f5 100644 --- a/services/hermes_analyst_service.py +++ b/services/hermes_analyst_service.py @@ -35,6 +35,9 @@ from config import HERMES_TIMEOUT HERMES_MODEL = "hermes3:latest" HERMES_KEEP_ALIVE = os.getenv("HERMES_KEEP_ALIVE", "5m") +HERMES_ALLOW_111_FALLBACK = os.getenv("HERMES_ALLOW_111_FALLBACK", "false").strip().lower() in ( + "1", "true", "yes", "on", +) TOP_N = 20 # 輸出前 N 個威脅，控制 NemoTron 每次消耗配額 @@ -280,6 +283,7 @@ class HermesAnalystService: temperature=0.1, timeout=HERMES_TIMEOUT, keep_alive=HERMES_KEEP_ALIVE, # ADR-012：避免冷啟動 timeout + allow_111_fallback=HERMES_ALLOW_111_FALLBACK, ) _ctx.set_provider(get_provider_tag(resp.host or '')) _ctx.set_model(resp.model or HERMES_MODEL) @@ -585,6 +589,7 @@ class HermesAnalystService: temperature=0.1, timeout=HERMES_TIMEOUT, keep_alive=HERMES_KEEP_ALIVE, + allow_111_fallback=HERMES_ALLOW_111_FALLBACK, ) _ctx.set_provider(get_provider_tag(resp.host or '')) _ctx.set_model(resp.model or HERMES_MODEL) diff --git a/services/marketplace_product_matcher.py b/services/marketplace_product_matcher.py index ad099e8..a26ab1a 100644 --- a/services/marketplace_product_matcher.py +++ b/services/marketplace_product_matcher.py @@ -523,6 +523,11 @@ VARIANT_SENSITIVE_KEYWORDS = { "粉底棒", "遮瑕棒", "修容打亮棒", + "防曬", + "防曬乳", + "防曬霜", + "防曬噴霧", + "防曬棒", } VARIANT_OPTION_COLOR_WORDS = { @@ -876,11 +881,21 @@ def _extract_model_tokens(text: str) -> set[str]: tokens: set[str] = set() for match in re.finditer(r"(?= 4 and re.search(r"[a-z]", compact) and re.search(r"\d", compact): tokens.add(compact) return tokens +def _is_spec_like_latin_token(token: str) -> bool: + return bool( + re.fullmatch(r"spf\d{1,3}[a-z]?", token) + or re.fullmatch(r"pa\d*", token) + or token in {"uva", "uvb", "uv", "spf"} + ) + + def _brand_alias_present(text: str, alias_norm: str, text_tokens: set[str]) -> bool: if not alias_norm: return False @@ -946,7 +961,7 @@ def _leading_brand_tokens(original: str, normalized: str) -> set[str]: if re.fullmatch(r"[\u4e00-\u9fff]{2,6}", first_token) and first_token not in GENERIC_TOKENS: tokens.add(first_token) for token in _tokenize(leading): - if re.fullmatch(r"[a-z][a-z0-9\-']{2,}", token): + if re.fullmatch(r"[a-z][a-z0-9\-']{2,}", token) and not _is_spec_like_latin_token(token): tokens.add(token) return tokens @@ -1260,6 +1275,30 @@ def _has_exact_count_alignment(left: ProductIdentity, right: ProductIdentity) -> return left_counts == right_counts +def _has_pack_quantity_difference(left: ProductIdentity, right: ProductIdentity) -> bool: + if not left.counts or not right.counts or _has_exact_count_alignment(left, right): + return False + + if left.total_piece_count and right.total_piece_count: + return left.total_piece_count != right.total_piece_count + + left_by_unit: dict[str, set[int]] = {} + right_by_unit: dict[str, set[int]] = {} + for count, unit in left.counts: + family = _count_unit_family(unit) + if family in COUNT_UNITS or unit in COUNT_UNITS: + left_by_unit.setdefault(family, set()).add(count) + for count, unit in right.counts: + family = _count_unit_family(unit) + if family in COUNT_UNITS or unit in COUNT_UNITS: + right_by_unit.setdefault(family, set()).add(count) + + for unit in set(left_by_unit) & set(right_by_unit): + if left_by_unit[unit] != right_by_unit[unit]: + return True + return False + + def _spec_score(left: ProductIdentity, right: ProductIdentity) -> tuple[float, bool, tuple[str, ...]]: volume_score, volume_conflict = _spec_component(left.volumes_ml, right.volumes_ml) weight_score, weight_conflict = _spec_component(left.weights_g, right.weights_g) @@ -1565,6 +1604,7 @@ def _is_unit_comparable_candidate( "multi_component_conflict", "count_conflict", "component_count_conflict", + "pack_quantity_difference", }) if not pack_difference: return False @@ -1644,6 +1684,8 @@ def _model_line_tokens(identity: ProductIdentity) -> set[str]: for token in identity.core_tokens: if token in GENERIC_TOKENS: continue + if _is_spec_like_latin_token(token): + continue if re.fullmatch(r"[a-z][a-z0-9-]{2,}", token): tokens.add(token) for match in re.finditer(r"([\u4e00-\u9fff]{2,})(?:系列)", token): @@ -1707,6 +1749,7 @@ def _build_evidence_flags( "variant_selection_review", "variant_option_conflict", "variant_descriptor_conflict", + "pack_quantity_difference", "count_conflict", "bundle_offer_conflict", "multi_component_conflict", @@ -1834,7 +1877,17 @@ def score_marketplace_match( catalog_count_omission = _allow_catalog_count_omission(left, right) if catalog_count_omission: reasons.append("catalog_count_omission") + if _has_pack_quantity_difference(left, right): + reasons.append("pack_quantity_difference") variant_descriptor_conflict = _has_variant_descriptor_conflict(left, right, shared_anchor) + sun_protection_line_conflict = ( + variant_descriptor_conflict + and left.product_type == right.product_type == "防曬" + and not shared_anchor + ) + if sun_protection_line_conflict: + reasons.append("variant_descriptor_conflict") + reasons.append("sun_protection_line_conflict") variant_option_conflict = _has_explicit_variant_option_conflict(left, right, shared_anchor) if variant_option_conflict: reasons.append("variant_option_conflict") @@ -1861,6 +1914,8 @@ def score_marketplace_match( hard_veto = True if left.product_type and right.product_type and left.product_type != right.product_type: hard_veto = True + if sun_protection_line_conflict: + hard_veto = True if variant_option_conflict: hard_veto = True @@ -2557,7 +2612,10 @@ def _shared_model_tokens(left: ProductIdentity, right: ProductIdentity) -> set[s return { token for token in left.core_tokens & right.core_tokens - if len(token) >= 4 and re.search(r"[a-z]", token) and re.search(r"\d", token) + if len(token) >= 4 + and re.search(r"[a-z]", token) + and re.search(r"\d", token) + and not _is_spec_like_latin_token(token) } @@ -2694,6 +2752,15 @@ def _has_baan_baby_lip_catalog_alignment(left: ProductIdentity, right: ProductId ) +def _has_recipe_box_child_sunscreen_cushion_alignment(left: ProductIdentity, right: ProductIdentity) -> bool: + brand_tokens = left.brand_tokens | right.brand_tokens + return ( + {"recipe", "box"} <= brand_tokens + and "兒童防曬氣墊粉餅" in left.searchable_name + and "兒童防曬氣墊粉餅" in right.searchable_name + ) + + def _has_pavaruni_40_scent_oil_alignment(left: ProductIdentity, right: ProductIdentity) -> bool: left_text = left.searchable_name right_text = right.searchable_name @@ -3068,6 +3135,8 @@ def _has_variant_descriptor_conflict(left: ProductIdentity, right: ProductIdenti return False if _has_baan_baby_lip_catalog_alignment(left, right): return False + if _has_recipe_box_child_sunscreen_cushion_alignment(left, right): + return False if _has_pavaruni_40_scent_oil_alignment(left, right): return False if _has_pavaruni_20_scent_candle_alignment(left, right): diff --git a/services/ollama_service.py b/services/ollama_service.py index 94115a0..faecf76 100644 --- a/services/ollama_service.py +++ b/services/ollama_service.py @@ -373,7 +373,8 @@ class OllamaService: system_prompt: str = None, temperature: float = 0.7, timeout: int = None, keep_alive: str = None, options: Optional[Dict[str, Any]] = None, - images: Optional[List[str]] = None) -> OllamaResponse: + images: Optional[List[str]] = None, + allow_111_fallback: bool = True) -> OllamaResponse: """ 生成文字 — 含三主機自動 retry（HOTFIX 2026-05-04） @@ -400,17 +401,26 @@ class OllamaService: attempted_hosts: List[str] = [] last_error: Optional[str] = None canonical_hosts = _canonical_host_chain() + allowed_hosts = [ + host for host in canonical_hosts + if allow_111_fallback or not _is_111_fallback_host(host) + ] + max_attempts = len(canonical_hosts) if allow_111_fallback else max(1, len(allowed_hosts)) - for attempt in range(3): + for attempt in range(max_attempts): current_host = _normalize_host(self.host) # property 每次 lazy resolve + if not allow_111_fallback and _is_111_fallback_host(current_host): + last_error = "111 fallback disabled; no approved GCP Ollama host available" + logger.warning("[Ollama] %s", last_error) + break if current_host in attempted_hosts: # 已試過同主機時，若是標準三主機鏈且 caller 沒指定 host， # 改走尚未嘗試的下一台。避免 request timeout(60s) 大於 # unhealthy TTL(30s) 時第三輪又 resolve 回 primary，導致 111 # final fallback 永遠沒被打到。 next_host = None - if self._explicit_host is None and current_host in canonical_hosts: - next_host = next((host for host in canonical_hosts if host not in attempted_hosts), None) + if self._explicit_host is None and current_host in allowed_hosts: + next_host = next((host for host in allowed_hosts if host not in attempted_hosts), None) if not next_host: # 非標準 host 或 explicit host 維持原行為：跳出避免無限迴圈。 break @@ -434,8 +444,9 @@ class OllamaService: payload["keep_alive"] = keep_alive logger.info( - "[Ollama] 嘗試 #%s/3 host=%s model=%s timeout=%ss keep_alive=%s", + "[Ollama] 嘗試 #%s/%s host=%s model=%s timeout=%ss keep_alive=%s", attempt + 1, + max_attempts, current_host, effective_model, effective_timeout, diff --git a/tests/test_hermes_ollama_cascade.py b/tests/test_hermes_ollama_cascade.py index f339ba2..9b03206 100644 --- a/tests/test_hermes_ollama_cascade.py +++ b/tests/test_hermes_ollama_cascade.py @@ -77,7 +77,7 @@ def test_hermes_intent_uses_ollama_service_and_logs_actual_host(monkeypatch, res monkeypatch, content='{"intent":"query_sales","confidence":0.9,"complexity_score":0.8,' '"requires_data_fetch":true,"preliminary_answer":""}', - host='http://192.168.0.111:11434', + host='http://34.21.145.224:11434', ) svc = hermes_mod.HermesAnalystService() @@ -88,12 +88,13 @@ def test_hermes_intent_uses_ollama_service_and_logs_actual_host(monkeypatch, res call_kwargs = fake_service.instances[0].generate_calls[0] assert call_kwargs['model'] == hermes_mod.HERMES_MODEL assert call_kwargs['keep_alive'] == hermes_mod.HERMES_KEEP_ALIVE + assert call_kwargs['allow_111_fallback'] is False assert _wait_for(reset_ai_logger, 1) rec = reset_ai_logger[0] assert rec['caller'] == 'hermes_intent' - assert rec['provider'] == 'ollama_111' - assert rec['meta']['host_label'] == '111 備援' + assert rec['provider'] == 'ollama_secondary' + assert rec['meta']['host_label'] == 'GCP-SSD-2' def test_hermes_batch_analyze_uses_ollama_service_and_logs_secondary(monkeypatch, reset_ai_logger): @@ -136,6 +137,7 @@ def test_hermes_batch_analyze_uses_ollama_service_and_logs_secondary(monkeypatch call_kwargs = fake_service.instances[0].generate_calls[0] assert call_kwargs['system_prompt'] == svc.SYSTEM_PROMPT assert call_kwargs['keep_alive'] == hermes_mod.HERMES_KEEP_ALIVE + assert call_kwargs['allow_111_fallback'] is False assert _wait_for(reset_ai_logger, 1) rec = reset_ai_logger[0] @@ -158,3 +160,7 @@ def test_hermes_candidate_sql_only_joins_direct_price_alert_matches(): def test_hermes_keep_alive_defaults_to_short_runner_residency(): assert hermes_mod.HERMES_KEEP_ALIVE == "5m" + + +def test_hermes_disables_111_fallback_by_default(): + assert hermes_mod.HERMES_ALLOW_111_FALLBACK is False diff --git a/tests/test_marketplace_product_matcher.py b/tests/test_marketplace_product_matcher.py index 19e6d29..ce4e5d5 100644 --- a/tests/test_marketplace_product_matcher.py +++ b/tests/test_marketplace_product_matcher.py @@ -151,6 +151,23 @@ def test_unit_price_comparison_builds_normalized_evidence(): assert comparison["unit_gap_pct"] < 0 +def test_marketplace_matcher_routes_same_base_different_piece_pack_to_unit_comparable(): + from services.marketplace_product_matcher import score_marketplace_match + + diagnostics = score_marketplace_match( + "【日本Beauty Foot】去角質足膜25mlx2枚入 5入組(一般尺寸、大尺寸可選)", + "【日本Beauty Foot 】煥膚足膜(25ml*2枚入)四入組", + momo_price=1290, + competitor_price=989, + ) + + assert diagnostics.comparison_mode == "unit_comparable" + assert diagnostics.match_type == "same_product_different_pack" + assert diagnostics.price_basis == "unit_price" + assert "pack_quantity_difference" in diagnostics.reasons + assert "unit_comparable" in diagnostics.reasons + + def test_marketplace_matcher_does_not_unit_compare_multi_component_set(): from services.marketplace_product_matcher import score_marketplace_match @@ -1313,6 +1330,10 @@ def test_marketplace_matcher_keeps_high_variant_low_score_lines_outside_focused_ "【Solone】持久眼線筆(眼線膠 超防暈推薦)", "Solone 斜角眉筆 0.35g", ) + sunscreen_line_gap = score_marketplace_match( + "【我的心機】溫和寶貝兒童防曬乳35ml(SPF50+ PA+++)", + "我的心機 海洋友善保濕高效防曬乳35ml(SPF50+PA++++)", + ) for diagnostics in ( lush, @@ -1328,10 +1349,14 @@ def test_marketplace_matcher_keeps_high_variant_low_score_lines_outside_focused_ romand_line_gap, summer_eve_variant_gap, solone_type_gap, + sunscreen_line_gap, ): assert diagnostics.score < 0.76 assert not any(reason.startswith("focused_exact_identity_") for reason in diagnostics.reasons) + assert sunscreen_line_gap.hard_veto is True + assert "variant_descriptor_conflict" in sunscreen_line_gap.reasons + def test_marketplace_matcher_rejects_refill_core_vs_case_only_pack(): from services.marketplace_product_matcher import score_marketplace_match diff --git a/tests/test_ollama_retry_chain.py b/tests/test_ollama_retry_chain.py index 740e5b9..bb4d5a6 100644 --- a/tests/test_ollama_retry_chain.py +++ b/tests/test_ollama_retry_chain.py @@ -163,6 +163,30 @@ def test_generate_forces_final_fallback_when_unhealthy_ttl_expires_mid_request() assert 'all 3 hosts failed' in (resp.error or '') +def test_generate_can_disable_111_fallback_for_batch_llm_work(): + """批量 LLM 任務可選擇只跑 GCP-A/GCP-B，避免 111 承接長分析。""" + import requests + from services import ollama_service as oss + from services.ollama_service import OllamaService + + svc = OllamaService() + hosts = [ + oss.OLLAMA_HOST_SECONDARY, + oss.OLLAMA_HOST_FALLBACK, + ] + + with patch('services.ollama_service.resolve_ollama_host', side_effect=hosts), \ + patch('services.ollama_service.requests.post', + side_effect=requests.Timeout('secondary timeout')) as mock_post: + resp = svc.generate('test', allow_111_fallback=False) + + posted_hosts = [call.args[0].split('/api/generate')[0] for call in mock_post.call_args_list] + assert resp.success is False + assert posted_hosts == [oss.OLLAMA_HOST_SECONDARY] + assert oss.OLLAMA_HOST_FALLBACK not in posted_hosts + assert '111 fallback disabled' in (resp.error or '') + + def test_generate_token_parsing_phase13(): """Phase 13 補強：OllamaResponse 解 prompt_eval_count + eval_count""" from services.ollama_service import OllamaService