fix: route telegram vision through ollama first

.env.example

@@ -315,6 +315,9 @@ RAG_EMBED_NORMALIZE=true
 PPT_VISION_ENABLED=false
 PPT_VISION_MODEL=minicpm-v:latest
 PPT_VISION_TIMEOUT=60
+PPT_AUTO_GENERATION_ENABLED=true
+PPT_AUTO_REPORT_TYPES=all
+PPT_AUTO_DEFAULT_CATEGORY=美妝保養
 DEEPSEEK_DIRECT_ENABLED=false
 DEEPSEEK_API_KEY=
 DEEPSEEK_BASE_URL=https://api.deepseek.com/v1
@@ -344,6 +347,9 @@ OLLAMA_EMBED_TIMEOUT=45
 OPENCLAW_QA_OLLAMA_FIRST=true
 OPENCLAW_QA_OLLAMA_MODEL=qwen3:14b
 OPENCLAW_QA_OLLAMA_TIMEOUT=60
+OPENCLAW_IMAGE_VISION_MODEL=minicpm-v:latest
+OPENCLAW_IMAGE_OLLAMA_TIMEOUT=45
+OPENCLAW_IMAGE_GEMINI_MODEL=gemini-1.5-flash
 NEMOTRON_OLLAMA_FIRST=true
 NEMOTRON_OLLAMA_MODEL=qwen3:14b
 NEMOTRON_OLLAMA_TIMEOUT=180

app.py

@@ -29,7 +29,7 @@ except OSError as e:
 
 # ================= 🔧 2. 核心模組導入 =================
 try:
-    from flask import Flask, render_template, jsonify, request, send_file, redirect, url_for, send_from_directory, flash, session
+    from flask import Flask, render_template, jsonify, request, send_file, redirect, url_for, flash, session
     from werkzeug.utils import secure_filename
     from pyngrok import ngrok, conf
     import schedule
@@ -146,16 +146,6 @@ def add_static_asset_version(endpoint, values):
         values['v'] = SYSTEM_VERSION
 
 
-@app.route('/favicon.ico')
-def favicon():
-    """使用既有品牌圖示回應瀏覽器預設 favicon 探測，避免全站 404 噪音。"""
-    return send_from_directory(
-        os.path.join(STATIC_DIR, 'images'),
-        'logo_circle.svg',
-        mimetype='image/svg+xml',
-        max_age=604800,
-    )
-
 # ==========================================
 # 🔒 Flask 安全配置
 # ==========================================

docs/AI_INTELLIGENCE_MODULE_SOT.md

@@ -18,6 +18,7 @@
 - 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。
 - OpenClaw Telegram Q&A 主路徑也不得綁單一 host：`_call_qwen3_qa()` 必須透過 `OllamaService` 跑 GCP-A → GCP-B → 111，並把實際落點寫入 `ai_calls.provider`。
+- OpenClaw Telegram 圖片商品辨識也必須 Ollama-first：`_identify_product_name_with_ollama_vision()` 透過 `OllamaService` 嘗試 GCP-A → GCP-B → 111；Gemini 只允許以 `openclaw_bot_image_gemini` caller 作為失敗後備援。
 
 ## 一、四 AI Agent 路由架構
 

docs/memory/claude_inventory_validation_20260513.md

@@ -56,6 +56,7 @@
 - PPT vision、PPT 文案 final fallback 與 MCP 離線 final fallback 已改走 `OllamaService.generate()`；`OllamaService.generate()` 支援 `options`、`keep_alive` 與 vision `images`，特殊 `/api/generate` 路徑同樣取得三主機 retry。
 - OpenClaw QA / daily Hermes template / NemoTron qwen3 的 flag 文件與測試已對齊 Ollama-first 預設 ON；顯式 `false` 才是 Gemini/NIM legacy 緊急退路。OpenClaw QA 已移除單一 `OPENCLAW_QA_OLLAMA_HOST` 主機覆寫，`_call_qwen3_qa()` 改走 `OllamaService` 的 GCP-A → GCP-B → 111 retry 並回寫實際 provider。
 - Code Review pipeline 已對齊 Ollama-first：`_hermes_scan()` 與 `_openclaw_assess()` 都先走 `OllamaService` 的 GCP-A → GCP-B → 111 retry；Gemini 僅在 Ollama（與可選 Claude）失敗後以 `code_review_openclaw_gemini` caller 記錄備援，不再以 `code_review_openclaw` 直接 Gemini-first。
+- Telegram 圖片商品辨識已對齊 Ollama-first：`routes/openclaw_bot_routes.py` 會先用 `OPENCLAW_IMAGE_VISION_MODEL` 透過 `OllamaService` retry GCP-A → GCP-B → 111；Gemini 只以 `openclaw_bot_image_gemini` caller 作為圖片辨識備援。
 - `.env.example` 已補齊 Python runtime 實際讀取的環境變數，`tests/test_phase3f_cleanup_contracts.py::test_env_example_documents_runtime_os_env_keys` 會掃 `app.py/config.py/scheduler.py/run_scheduler.py/routes/services/utils` 的 `os.getenv()` / `os.environ.get()`；只允許 `PYTEST_CURRENT_TEST` 與 `MOMO_ALLOW_INSECURE_CONFIG_FOR_TESTS` 兩個測試內部 key 不進範例。
 - `docker-compose*.yml` 使用的 `${VAR}` 也已納入 `.env.example` 契約，包含 MCP compose 的 `TAVILY_API_KEY`、`EXA_API_KEY`、`MCP_POSTGRES_PASSWORD`、`FIRECRAWL_AUTH_KEY`，以及 image tag / Grafana / pgAdmin / Metabase / Grist 變數；`test_env_example_documents_docker_compose_variables` 會守住。
 - Market Intel `seed_writer_cli_status` route 已補 API 層回歸：即使 `execute=true` 且環境有 `MARKET_INTEL_SEED_WRITE_APPROVAL`，API 仍不得回吐 token / `approval_token_hint` / 固定 token 文案，且不得 ready 或寫入；`tests/test_market_intel_skeleton.py::test_seed_writer_cli_status_route_never_leaks_approval_token` 會守住。

routes/openclaw_bot_routes.py

@@ -4,7 +4,7 @@
 OpenClaw Telegram 群組智能助理 v5
 ─────────────────────────────────────────
 核心功能：
-  • 群組自然對話，Gemini Flash 主引擎（2~5s）
+  • 群組自然對話，Ollama-first 三主機級聯，Gemini 僅備援
   • Inline Keyboard 15 個功能入口
   • 全商品查詢帶出商品ID
   • AI 分析強制比對內部DB + 外部MCP情報
@@ -91,7 +91,7 @@ except ImportError:
 
 # V-New: 引入 Ollama 探測機制
 try:
-    from services.ollama_service import OllamaService
+    from services.ollama_service import OllamaService, get_host_label, get_provider_tag
     _OLLAMA_AVAILABLE = True
 except ImportError:
     _OLLAMA_AVAILABLE = False
@@ -103,6 +103,11 @@ except ImportError:
 GEMINI_API_KEY  = os.getenv('GEMINI_API_KEY', '')
 GEMINI_BASE_URL = 'https://generativelanguage.googleapis.com/v1beta/models'
 GEMINI_MODEL    = 'gemini-2.0-flash'
+IMAGE_VISION_OLLAMA_MODEL = os.getenv(
+    'OPENCLAW_IMAGE_VISION_MODEL',
+    os.getenv('PPT_VISION_MODEL', 'minicpm-v:latest'),
+)
+IMAGE_VISION_GEMINI_MODEL = os.getenv('OPENCLAW_IMAGE_GEMINI_MODEL', 'gemini-1.5-flash')
 PPT_CACHE_TTL_HOURS = max(1, int(os.getenv('OPENCLAW_PPT_CACHE_TTL_HOURS', '24')))
 
 TAIPEI_TZ = timezone(timedelta(hours=8))
@@ -8678,6 +8683,123 @@ def _handle_event_ignore_callback(data: str, cq: dict, chat_id, message_id) -> N
     sys_log.info(f"[EA HITL] event_ignore event_id={event_id} by={user_label_raw}")
 
 
+def _clean_vision_product_name(raw: str) -> str:
+    """把 vision 模型回應收斂成可直接丟給比價查詢的商品名稱。"""
+    text = (raw or '').strip()
+    if not text:
+        return ''
+    text = re.sub(r"^```(?:text)?\s*", "", text, flags=re.IGNORECASE).strip()
+    text = re.sub(r"\s*```$", "", text).strip()
+    first_line = next((line.strip() for line in text.splitlines() if line.strip()), '')
+    first_line = re.sub(r"^(商品名稱|品名|辨識結果|結果)\s*[:：]\s*", "", first_line).strip()
+    first_line = first_line.strip("`*_ -　")
+    if not first_line:
+        return ''
+    refusal_patterns = ('無法辨識', '看不清', '無法確認', '不確定', 'unknown', 'not sure')
+    lowered = first_line.lower()
+    if any(pattern in lowered for pattern in refusal_patterns):
+        return ''
+    return first_line[:60]
+
+
+def _identify_product_name_with_ollama_vision(img_b64: str, request_id: str) -> str:
+    """圖片比價的主辨識路徑：Ollama vision 三主機級聯。"""
+    if not _OLLAMA_AVAILABLE:
+        return ''
+    prompt = (
+        "這是一張商品圖片。請辨識商品名稱，包含品牌、型號、規格。"
+        "只回商品名稱，不要解釋，不要 markdown，不超過 30 字；"
+        "如果是多個商品，只取最顯眼的一個。必須使用繁體中文。"
+    )
+    timeout = int(os.getenv('OPENCLAW_IMAGE_OLLAMA_TIMEOUT', '45'))
+    with log_ai_call(
+        caller='openclaw_bot_image',
+        provider='gcp_ollama',
+        model=IMAGE_VISION_OLLAMA_MODEL,
+        request_id=request_id,
+        meta={'route': 'ollama_first', 'task': 'image_product_recognition'},
+    ) as ctx:
+        try:
+            resp = OllamaService(model=IMAGE_VISION_OLLAMA_MODEL).generate(
+                prompt=prompt,
+                model=IMAGE_VISION_OLLAMA_MODEL,
+                temperature=0.1,
+                timeout=timeout,
+                options={'num_predict': 64},
+                images=[img_b64],
+            )
+            ctx.set_provider(get_provider_tag(resp.host or ''))
+            ctx.set_tokens(input=resp.input_tokens, output=resp.output_tokens)
+            ctx.add_meta('host', resp.host)
+            ctx.add_meta('host_label', get_host_label(resp.host or ''))
+            if not resp.success:
+                ctx.set_error(resp.error or 'ollama vision failed')
+                ctx.fallback_to_caller('openclaw_bot_image_gemini')
+                return ''
+            product_name = _clean_vision_product_name(resp.content)
+            if not product_name:
+                ctx.set_error('empty_or_unusable_vision_response')
+                ctx.fallback_to_caller('openclaw_bot_image_gemini')
+            return product_name
+        except Exception as exc:
+            ctx.set_error(f"{type(exc).__name__}: {exc}")
+            ctx.fallback_to_caller('openclaw_bot_image_gemini')
+            sys_log.warning(f"[VisionSearch] Ollama vision failed: {exc}")
+            return ''
+
+
+def _identify_product_name_with_gemini_vision(img_b64: str, request_id: str) -> str:
+    """圖片比價的雲端備援：只有 Ollama vision 失敗後才呼叫。"""
+    if not GEMINI_API_KEY:
+        return ''
+    vision_payload = {
+        'contents': [{
+            'parts': [
+                {'text': (
+                    '這是一張商品圖片。請辨識商品名稱（品牌、型號、規格），'
+                    '輸出格式：只回商品名稱，不超過 30 字，繁體中文。'
+                    '如果是多個商品，只取最顯眼的一個。'
+                )},
+                {'inline_data': {'mime_type': 'image/jpeg', 'data': img_b64}},
+            ],
+        }],
+    }
+    with log_ai_call(
+        caller='openclaw_bot_image_gemini',
+        provider='gemini',
+        model=IMAGE_VISION_GEMINI_MODEL,
+        request_id=request_id,
+        meta={'fallback_from': 'openclaw_bot_image', 'task': 'image_product_recognition'},
+    ) as ctx:
+        try:
+            vis_r = requests.post(
+                f"{GEMINI_BASE_URL}/{IMAGE_VISION_GEMINI_MODEL}:generateContent?key={GEMINI_API_KEY}",
+                json=vision_payload, timeout=20,
+            )
+            vis_r.raise_for_status()
+            body = vis_r.json()
+            usage = body.get('usageMetadata', {}) or {}
+            ctx.set_tokens(
+                input=usage.get('promptTokenCount', 0),
+                output=usage.get('candidatesTokenCount', 0),
+            )
+            raw = (
+                body
+                .get('candidates', [{}])[0]
+                .get('content', {})
+                .get('parts', [{}])[0]
+                .get('text', '')
+            )
+            product_name = _clean_vision_product_name(raw)
+            if not product_name:
+                ctx.set_error('empty_or_unusable_vision_response')
+            return product_name
+        except Exception as exc:
+            ctx.set_error(f"{type(exc).__name__}: {exc}")
+            sys_log.warning(f"[VisionSearch] Gemini vision fallback failed: {exc}")
+            return ''
+
+
 # ── Webhook ───────────────────────────────────────────────────
 @openclaw_bot_bp.route('/bot/telegram/webhook', methods=['POST'])
 def telegram_webhook():
@@ -8943,7 +9065,7 @@ def telegram_webhook():
             # 已通過授權的 private chat
             question = text_raw
 
-        # ── 圖片訊息：Gemini Vision 商品辨識 ─────────────────────
+        # ── 圖片訊息：Ollama-first Vision 商品辨識 ─────────────────
         if not question and msg.get('photo'):
             send_typing(chat_id)
             try:
@@ -8962,43 +9084,22 @@ def telegram_webhook():
                 img_data = requests.get(img_url, timeout=15).content
                 import base64 as _b64
                 img_b64  = _b64.b64encode(img_data).decode()
-                # Gemini Vision 辨識商品名稱
-                vision_payload = {
-                    'contents': [{
-                        'parts': [
-                            {'text': (
-                                '這是一張商品圖片。請辨識商品名稱（品牌、型號、規格），'
-                                '輸出格式：只回商品名稱，不超過 30 字，繁體中文。'
-                                '如果是多個商品，只取最顯眼的一個。'
-                            )},
-                            {'inline_data': {'mime_type': 'image/jpeg', 'data': img_b64}}
-                        ]
-                    }]
-                }
-                vis_r = requests.post(
-                    f"{GEMINI_BASE_URL}/gemini-1.5-flash:generateContent?key={GEMINI_API_KEY}",
-                    json=vision_payload, timeout=20
-                )
-                if vis_r.ok:
-                    product_name = (
-                        vis_r.json()
-                        .get('candidates', [{}])[0]
-                        .get('content', {})
-                        .get('parts', [{}])[0]
-                        .get('text', '').strip()
-                    )
-                    if product_name:
-                        send_message(chat_id,
-                                     f"🔍 辨識到商品：*{product_name}*\n正在搜尋 momo 比價...",
-                                     msg_id, parse_mode='Markdown')
-                        # 直接執行比價
-                        handle_cmd('competitor', product_name, chat_id, msg_id)
-                    else:
-                        send_message(chat_id, "⚠️ 無法辨識圖片中的商品，請嘗試更清晰的圖片", msg_id)
+
+                req_id = f"img-{chat_id or 0}-{msg_id or 0}"
+                product_name = _identify_product_name_with_ollama_vision(img_b64, req_id)
+                if not product_name:
+                    product_name = _identify_product_name_with_gemini_vision(img_b64, req_id)
+
+                if product_name:
+                    send_message(chat_id,
+                                 f"🔍 辨識到商品：*{product_name}*\n正在搜尋 momo 比價...",
+                                 msg_id, parse_mode='Markdown')
+                    # 直接執行比價
+                    handle_cmd('competitor', product_name, chat_id, msg_id)
                 else:
                     send_message(
                         chat_id,
-                        "⚠️ 圖片辨識失敗，請直接輸入商品名稱搜尋",
+                        "⚠️ 無法辨識圖片中的商品，請直接輸入商品名稱搜尋",
                         msg_id,
                         [_row(('🔍 文字搜尋', 'await:search_compare'))],
                     )

routes/system_public_routes.py

@@ -27,6 +27,18 @@ sys_log = SystemLogger("SystemPublicRoutes").get_logger()
 TAIPEI_TZ = timezone(timedelta(hours=8))
 LOG_FILE_PATH = os.path.join(BASE_DIR, 'logs/system.log')
 public_url = os.getenv('PUBLIC_URL', '服務啟動中...')
+STATIC_DIR = os.path.join(BASE_DIR, 'web/static')
+
+
+@system_public_bp.route('/favicon.ico')
+def favicon():
+    """使用既有品牌圖示回應瀏覽器預設 favicon 探測，避免全站 404 噪音。"""
+    return send_from_directory(
+        os.path.join(STATIC_DIR, 'images'),
+        'logo_circle.svg',
+        mimetype='image/svg+xml',
+        max_age=604800,
+    )
 
 
 @system_public_bp.route('/health')

services/llm_caller_registry.py

@@ -80,6 +80,8 @@ CALLER_REGISTRY: frozenset = frozenset({
     'tg_bot_copy_v2',           # second copy entrance
     'openclaw_bot_main',        # OpenClaw Bot 主鏈 Ollama
     'openclaw_bot_gemini',      # Bot Gemini fallback
+    'openclaw_bot_image',       # Bot 圖片商品辨識 Ollama-first
+    'openclaw_bot_image_gemini',  # Bot 圖片商品辨識 Gemini fallback
     'openclaw_bot_nim',         # Bot NIM fallback
 
     # 其他

tests/test_caller_registry.py

@@ -41,6 +41,7 @@ def test_registry_contains_core_callers():
         'sales_copy', 'trend_match', 'trend_qa', 'product_insights',
         # Bot
         'openclaw_bot_main', 'openclaw_bot_gemini', 'openclaw_bot_nim',
+        'openclaw_bot_image', 'openclaw_bot_image_gemini',
     }
 
     missing = must_have - CALLER_REGISTRY

tests/test_openclaw_bot_routes_webhook.py

@@ -149,6 +149,124 @@ def test_is_authorized_private_mode_switch(monkeypatch):
     assert bot._is_authorized("private", 777, 42) is False
 
 
+def test_photo_message_uses_ollama_vision_before_gemini(monkeypatch):
+    from routes import openclaw_bot_routes as bot
+
+    sent = []
+    handled = []
+
+    class FakeResponse:
+        def __init__(self, json_data=None, content=b"fake-image"):
+            self._json_data = json_data or {}
+            self.content = content
+
+        def json(self):
+            return self._json_data
+
+    def fake_get(url, **_kwargs):
+        if "getFile" in url:
+            return FakeResponse({"result": {"file_path": "photos/product.jpg"}})
+        return FakeResponse(content=b"fake-image")
+
+    monkeypatch.setattr(bot.requests, "get", fake_get)
+    monkeypatch.setattr(bot, "_is_authorized", lambda _chat_type, _chat_id, _uid: True)
+    monkeypatch.setattr(bot, "send_typing", lambda _chat_id: None)
+    monkeypatch.setattr(
+        bot,
+        "send_message",
+        lambda *args, **kwargs: sent.append((args, kwargs)),
+    )
+    monkeypatch.setattr(
+        bot,
+        "handle_cmd",
+        lambda cmd, arg, chat_id, reply_to: handled.append((cmd, arg, chat_id, reply_to)),
+    )
+    monkeypatch.setattr(
+        bot,
+        "_identify_product_name_with_ollama_vision",
+        lambda img_b64, request_id: "理膚寶水 B5 修復霜",
+    )
+    monkeypatch.setattr(
+        bot,
+        "_identify_product_name_with_gemini_vision",
+        lambda img_b64, request_id: (_ for _ in ()).throw(AssertionError("Gemini should not run first")),
+    )
+
+    app = _build_request_app()
+    payload = {
+        "update_id": 10030,
+        "message": {
+            "message_id": 80,
+            "chat": {"id": 777, "type": "private"},
+            "from": {"id": 777777},
+            "photo": [{"file_id": "small"}, {"file_id": "large"}],
+        },
+    }
+
+    with app.test_request_context("/bot/telegram/webhook", method="POST", json=payload):
+        bot.telegram_webhook()
+
+    assert handled == [("competitor", "理膚寶水 B5 修復霜", 777, 80)]
+    assert "理膚寶水 B5 修復霜" in sent[0][0][1]
+
+
+def test_photo_message_falls_back_to_gemini_when_ollama_empty(monkeypatch):
+    from routes import openclaw_bot_routes as bot
+
+    handled = []
+    calls = []
+
+    class FakeResponse:
+        def __init__(self, json_data=None, content=b"fake-image"):
+            self._json_data = json_data or {}
+            self.content = content
+
+        def json(self):
+            return self._json_data
+
+    def fake_get(url, **_kwargs):
+        if "getFile" in url:
+            return FakeResponse({"result": {"file_path": "photos/product.jpg"}})
+        return FakeResponse(content=b"fake-image")
+
+    def fake_ollama(_img_b64, _request_id):
+        calls.append("ollama")
+        return ""
+
+    def fake_gemini(_img_b64, _request_id):
+        calls.append("gemini")
+        return "飛利浦 Sonicare"
+
+    monkeypatch.setattr(bot.requests, "get", fake_get)
+    monkeypatch.setattr(bot, "_is_authorized", lambda _chat_type, _chat_id, _uid: True)
+    monkeypatch.setattr(bot, "send_typing", lambda _chat_id: None)
+    monkeypatch.setattr(bot, "send_message", lambda *args, **kwargs: None)
+    monkeypatch.setattr(
+        bot,
+        "handle_cmd",
+        lambda cmd, arg, chat_id, reply_to: handled.append((cmd, arg, chat_id, reply_to)),
+    )
+    monkeypatch.setattr(bot, "_identify_product_name_with_ollama_vision", fake_ollama)
+    monkeypatch.setattr(bot, "_identify_product_name_with_gemini_vision", fake_gemini)
+
+    app = _build_request_app()
+    payload = {
+        "update_id": 10031,
+        "message": {
+            "message_id": 81,
+            "chat": {"id": 777, "type": "private"},
+            "from": {"id": 777777},
+            "photo": [{"file_id": "small"}, {"file_id": "large"}],
+        },
+    }
+
+    with app.test_request_context("/bot/telegram/webhook", method="POST", json=payload):
+        bot.telegram_webhook()
+
+    assert calls == ["ollama", "gemini"]
+    assert handled == [("competitor", "飛利浦 Sonicare", 777, 81)]
+
+
 def test_obs_heal_audit_uses_current_callback_user(monkeypatch):
     from types import SimpleNamespace
     from routes import openclaw_bot_routes as bot

tests/test_phase3f_cleanup_contracts.py

@@ -142,6 +142,9 @@ def test_env_example_documents_runtime_and_ai_automation_variables():
         "OPENCLAW_DAILY_HERMES_TEMPLATE",
         "OPENCLAW_OLLAMA_MODEL",
         "OPENCLAW_PPT_CACHE_TTL_HOURS",
+        "OPENCLAW_IMAGE_GEMINI_MODEL",
+        "OPENCLAW_IMAGE_OLLAMA_TIMEOUT",
+        "OPENCLAW_IMAGE_VISION_MODEL",
         "OPENCLAW_QA_OLLAMA_FIRST",
         "OPENCLAW_QA_OLLAMA_MODEL",
         "OPENCLAW_QA_OLLAMA_TIMEOUT",