entari-plugin-hyw 4.0.0rc15__py3-none-any.whl → 4.0.0rc17__py3-none-any.whl

hyw_core/browser_control/service.py

@@ -12,6 +12,8 @@ from concurrent.futures import ThreadPoolExecutor
 from typing import Optional, Dict, Any, List
 from loguru import logger
 import trafilatura
+from PIL import Image
+from io import BytesIO
 
 # Import intelligent completeness checker
 from ..crawling.completeness import CompletenessChecker, trigger_lazy_load
@@ -541,13 +543,13 @@ class ScreenshotService:
         tasks = [self.screenshot_url(url, timeout=timeout, full_page=full_page) for url in urls]
         return await asyncio.gather(*tasks, return_exceptions=True)
 
-    async def screenshot_url(self, url: str, wait_load: bool = True, timeout: float = 15.0, full_page: bool = False, quality: int = 80) -> Optional[str]:
+    async def screenshot_url(self, url: str, wait_load: bool = True, timeout: float = 15.0, full_page: bool = False, quality: int = 90, scale: int = 1) -> Optional[str]:
         """Screenshot URL (Async wrapper for sync). Returns base64 string only."""
         loop = asyncio.get_running_loop()
         result = await loop.run_in_executor(
             self._executor,
             self._screenshot_sync,
-            url, wait_load, timeout, full_page, quality, False  # extract_content=False
+            url, wait_load, timeout, full_page, quality, scale, False  # extract_content=False
         )
         # Backward compatible: return just the screenshot for old callers
         if isinstance(result, dict):
@@ -569,7 +571,7 @@ class ScreenshotService:
         result = await loop.run_in_executor(
             self._executor,
             self._screenshot_sync,
-            url, True, timeout, False, 65, True  # quality=65 for balance, extract_content=True
+            url, True, timeout, False, 80, 2, True  # quality=80 for balance, scale=2, extract_content=True
         )
         
         if not isinstance(result, dict):
@@ -584,12 +586,12 @@ class ScreenshotService:
         return result
 
 
-    def _screenshot_sync(self, url: str, wait_load: bool, timeout: float, full_page: bool, quality: int, extract_content: bool = False) -> Any:
+    def _screenshot_sync(self, url: str, wait_load: bool, timeout: float, full_page: bool, quality: int, scale: int = 1, extract_content: bool = False) -> Any:
         """Synchronous screenshot. If extract_content=True, returns Dict else str."""
         if not url: 
             return {"screenshot_b64": None, "content": "", "title": "", "url": url} if extract_content else None
         tab = None
-        capture_width = 1440  # Higher resolution for readability
+        capture_width = 3000  # Increased for more comfortable page size while maintaining high resolution
         
         try:
             self._ensure_ready()
@@ -603,7 +605,7 @@ class ScreenshotService:
             # This eliminates the need for post-load resize and reflow
             try:
                 tab.run_cdp('Emulation.setDeviceMetricsOverride', 
-                            width=capture_width, height=900, deviceScaleFactor=1, mobile=False)
+                            width=capture_width, height=900, deviceScaleFactor=scale, mobile=False)
             except:
                 pass
             
@@ -876,6 +878,23 @@ class ScreenshotService:
             # Capture screenshot
             screenshot_b64 = tab.get_screenshot(as_base64='jpg', full_page=False)
             
+            # Use Pillow for intelligent compression
+            if screenshot_b64 and quality < 95: # Only compress if quality is not near maximum
+                try:
+                    img_bytes = base64.b64decode(screenshot_b64)
+                    img = Image.open(BytesIO(img_bytes))
+                    
+                    # Convert to RGB if not already (some images might be RGBA, which JPEG doesn't support)
+                    if img.mode in ("RGBA", "P"):
+                        img = img.convert("RGB")
+                    
+                    output_buffer = BytesIO()
+                    img.save(output_buffer, format="WebP", quality=quality, optimize=True) # Output as WebP format
+                    screenshot_b64 = base64.b64encode(output_buffer.getvalue()).decode()
+                    logger.debug(f"ScreenshotService: Applied Pillow compression with quality={quality}")
+                except Exception as e:
+                    logger.warning(f"ScreenshotService: Pillow compression failed: {e}")
+            
             # Extract content if requested
             if extract_content:
                 try:

hyw_core/core.py

@@ -301,28 +301,20 @@ class HywCore:
             
             # Build visible results list (excluding hidden items)
             visible_results = [r for r in web_results if not r.get("_hidden")]
-            
-            # Parse markdown to find which citations are used (pattern: [number])
-            citation_pattern = re.compile(r'\[(\d+)\]')
-            cited_ids = set()
-            for match in citation_pattern.finditer(content):
-                cited_ids.add(int(match.group(1)))
-            
-            # Only include cited references, in order of first appearance
+
+            # Pass ALL visible results to frontend so [N] citations map correctly to N-th item
+            # App.vue handles reordering used vs unused citations
             references = []
-            for idx in sorted(cited_ids):
-                # idx is 1-based in markdown
-                if 1 <= idx <= len(visible_results):
-                    r = visible_results[idx - 1]
-                    references.append({
-                        "title": r.get("title", ""),
-                        "url": r.get("url", ""),
-                        "snippet": r.get("content", "")[:300] if r.get("content") else "",
-                        "images": r.get("images", []),
-                        "is_fetched": r.get("_type") == "page",
-                        "raw_screenshot_b64": r.get("screenshot_b64"),
-                    })
-            
+            for r in visible_results:
+                references.append({
+                    "title": r.get("title", ""),
+                    "url": r.get("url", ""),
+                    "snippet": r.get("content", "")[:300] if r.get("content") else "",
+                    "images": r.get("images", []),
+                    "is_fetched": r.get("_type") == "page",
+                    "raw_screenshot_b64": r.get("screenshot_b64"),
+                })
+
             # Build response
             response = QueryResponse(
                 success=True,

hyw_core/definitions.py

@@ -120,6 +120,7 @@ AGENT_SYSTEM_PROMPT = """# 你是一个智能助手 (Agent), 你的职责是使
 ## 任务
 理解用户意图分配给你的任务.
 如果用户没有明确分配任务, 则默认任务为解释用户问题中的关键词.
+分辨用户消息的语意, 提取出用户最想了解的核心内容, 作为任务的核心.
 
 ## 核心原则
 最小限度使用自身知识, 尽可能使用 web_tool 获取信息.
@@ -128,16 +129,32 @@ AGENT_SYSTEM_PROMPT = """# 你是一个智能助手 (Agent), 你的职责是使
 - 并行调用工具
     - 网页搜索: 可以同时调用3次, 其中URL截图消耗较大, 最多同时调用1个
 - 积极使用 web_tool 获取信息
-    - 搜索时, 关键词保证简单、指向准确、利于传统搜索引擎.
-    - 不要尝试通过搜索引擎反推出角色、任务、地点, 搜索引擎没有这个能力
+    - 搜索时, 关键词保证单一、简单、指向准确、利于传统搜索引擎, 通常只搜索1个词或短语.
+        - 建议搜索: "minecraft create"; 不搜索 "create 是什么 百科"
+        - 建议搜索: "opnecode"; 不搜索 "open code 怎么配置"
+        - 建议搜索: "Bypass permissions"; 不搜索 "Bypass permissions 软件 选项"
+    - 本搜索不支持高级搜索、不支持引号、不支持减号等复杂语法
+    - 不要尝试通过搜索引擎描述如何尼尔反推出角色、任务、地点, 搜索引擎没有这个能力
     - 禁止搜索可能导致一切潜在推销广告的内容, 不出现“是什么”、“怎么办”等容易产生广告的内容
     - 禁止搜索任何敏感内容(galgame之类的除外), 禁止搜索政治、成人色情、暴力等内容
-    - 获取页面截图时, 只使用官方性较强的 wiki、官方网站、资源站等等, 不使用第三方转载新闻网站.
-- 最多可调用2轮工具, 之后必须给出最终回答
+    - 获取页面截图时, 只挑选官方性较强的 wiki、官方网站、资源站等等, 不使用第三方转载新闻网站.
+- 最多可调用3轮工具, 但请适度保持快速, 3轮之后必须给出最终回答.
 - 适当时候调用 `refuse_answer`
 - 对于具体任务, 如果是转述、格式化、翻译等, 请直接给出最终回答, 不再调用工具
 - 遇到计算、js代码、算法任务, 积极使用 js_executor 工具完成计算任务.
 
+## 抓重点原则
+搜索结果中往往混杂大量信息，你需要：
+- 主动识别与用户问题最匹配的结果，大胆引用，不要因为信息混在众多结果中就忽略它
+- 即使只有一条结果明确匹配，也要优先使用该结果，而非泛泛而谈
+
+## 图文融合原则
+当用户同时提供图片和文字时：
+- 先理解用户真正想知道什么（识图？查资料？对比分析？）
+- 图片是"锚点"，搜索是"扩展"——围绕图片内容组织搜索信息
+- 行文自然流畅，让图片分析和搜索结果无缝衔接
+- 例如："图中展示的是 XX（识别结果），这是一款...（搜索扩展）"
+
 ## 回答格式
 - 字数: 尽可能少, 有多少获取到的信息、需要解释的内容, 就写多少, 减少无意义输出, 足够完成用户分配给你的任务 / 解释关键词即可.
 - `# ` 大标题约 8-10 个字