fix: adapt frontend image paths for backend port 8002 and use blob download

Backend port changed from 8000 to 8002 due to port conflict (frpc on Windows). - Add resolveImageUrl() to convert relative image URLs to full backend URLs - Fix all spatial page image src and download handlers to use backend address - Fix chat message image rendering to use correct port - Switch download to blob-based approach to support cross-origin file saving Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-06-02 20:52:48 +08:00
parent a6987ff996
commit bbf6b921af
10 changed files with 171 additions and 127 deletions
@@ -43,6 +43,7 @@ dependencies = [
    "docx2txt>=0.9",
    "pypdf>=6.12.0",
    "pymupdf>=1.27.2.3",
+    "pdf2image>=1.17.0",
    "opencv-python-headless>=4.13.0.92",
 ]

@@ -128,6 +128,10 @@ class Settings(BaseSettings):
 # 全局配置实例
 settings = Settings()

+# 将 HF_ENDPOINT 写入环境变量，供 sentence_transformers/huggingface_hub 使用
+if settings.hf_endpoint and not os.environ.get("HF_ENDPOINT"):
+    os.environ["HF_ENDPOINT"] = settings.hf_endpoint
+

 def get_settings() -> Settings:
    """获取配置实例"""
@@ -2,11 +2,11 @@
 LangChain 1.0 文档加载器封装 + PDF图片提取
 """
 import logging
-import io
 from typing import List, Optional, Dict
 from pathlib import Path
-import fitz  # pymupdf
+import fitz  # pymupdf — 仅用于提取页面文本
 from PIL import Image
+from pdf2image import convert_from_path
 from langchain_community.document_loaders import (
    PyPDFLoader,
    Docx2txtLoader,
@@ -19,100 +19,92 @@ logger = logging.getLogger(__name__)


 class PDFImageExtractor:
-    """使用pymupdf从PDF中提取内嵌图片"""
+    """使用pdf2image将每页PDF渲染为图片，用pymupdf提取页面文本"""
+
+    DEFAULT_DPI = 250
+    DEFAULT_TARGET_WIDTH = 2500
+    JPEG_QUALITY = 90

    @staticmethod
-    def extract_images(file_path: str, output_dir: str) -> List[dict]:
-        """提取PDF中所有图片，返回图片元数据列表
+    def extract_images(
+        file_path: str,
+        output_dir: str,
+        dpi: int = None,
+        target_width: int = None,
+    ) -> List[dict]:
+        """将PDF每页渲染为一张图片

-        对正常页面提取内嵌图片，对瓦片式页面（>50个小图片）渲染整页截图。
+        Args:
+            file_path: PDF文件路径
+            output_dir: 图片输出目录
+            dpi: 渲染DPI（默认250）
+            target_width: 缩放目标宽度像素（默认2500）
+
+        Returns:
+            图片元数据列表，每个dict包含 path, filename, page, context_text, size
        """
+        dpi = dpi or PDFImageExtractor.DEFAULT_DPI
+        target_width = target_width or PDFImageExtractor.DEFAULT_TARGET_WIDTH
+        quality = PDFImageExtractor.JPEG_QUALITY
+
        Path(output_dir).mkdir(parents=True, exist_ok=True)
+
+        # 用pymupdf提取每页文本（作为VLM上下文）
+        page_texts: list[str] = []
+        try:
+            doc = fitz.open(file_path)
+            for page_num in range(len(doc)):
+                page_texts.append(doc[page_num].get_text("text"))
+            doc.close()
+        except Exception as e:
+            logger.warning(f"[ImageExtractor] pymupdf文本提取失败: {e}")
+
+        # 用pdf2image渲染所有页面为图片
+        try:
+            pil_images = convert_from_path(file_path, dpi=dpi, fmt="jpeg")
+        except Exception as e:
+            logger.error(f"[ImageExtractor] pdf2image渲染失败: {e}")
+            return []
+
        images = []
-        doc = fitz.open(file_path)
+        for page_num, pil_img in enumerate(pil_images):
+            try:
+                w, h = pil_img.size
+                if w > target_width:
+                    ratio = target_width / w
+                    new_h = int(h * ratio)
+                    pil_img = pil_img.resize(
+                        (target_width, new_h), Image.Resampling.LANCZOS
+                    )

-        TILE_THRESHOLD = 50
-        vlm_formats = {"png", "jpg", "jpeg", "webp", "gif"}
-        dpi = 150  # 页面渲染 DPI
+                filename = f"page{page_num + 1}.jpg"
+                output_path = Path(output_dir) / filename
+                pil_img.save(str(output_path), format="JPEG", quality=quality)

-        for page_num in range(len(doc)):
-            page = doc[page_num]
-            page_text = page.get_text("text")
-            image_list = page.get_images(full=True)
+                context_text = ""
+                if page_num < len(page_texts):
+                    context_text = page_texts[page_num][:600].strip()

-            if len(image_list) > TILE_THRESHOLD:
-                # 瓦片式页面：渲染整页为一张完整图片
-                try:
-                    pix = page.get_pixmap(dpi=dpi)
-                    img_bytes = pix.tobytes("png")
-                    filename = f"page{page_num+1}_full.png"
-                    output_path = Path(output_dir) / filename
-                    output_path.write_bytes(img_bytes)
-
-                    images.append({
-                        "path": str(output_path),
-                        "filename": filename,
-                        "page": page_num + 1,
-                        "context_text": page_text[:600].strip(),
-                        "size": len(img_bytes),
-                    })
-                    logger.info(f"[ImageExtractor] 瓦片页面渲染为整图 page={page_num+1} size={len(img_bytes)}")
-                except Exception as e:
-                    logger.warning(f"[ImageExtractor] 页面渲染失败 page={page_num+1}: {e}")
+                images.append({
+                    "path": str(output_path),
+                    "filename": filename,
+                    "page": page_num + 1,
+                    "context_text": context_text,
+                    "size": output_path.stat().st_size,
+                })
+                logger.info(
+                    f"[ImageExtractor] 页面渲染完成 page={page_num + 1} "
+                    f"size={images[-1]['size']}"
+                )
+            except Exception as e:
+                logger.warning(
+                    f"[ImageExtractor] 页面处理失败 page={page_num + 1}: {e}"
+                )
                continue

-            # 正常页面：提取内嵌图片
-            for img_idx, img_info in enumerate(image_list):
-                xref = img_info[0]
-                try:
-                    base_image = doc.extract_image(xref)
-                    image_bytes = base_image["image"]
-                    ext = base_image["ext"]
-
-                    # 过滤太小的碎片/图标/装饰
-                    if len(image_bytes) < 2048:
-                        continue
-                    w = base_image.get("width", 0)
-                    h = base_image.get("height", 0)
-                    if w < 150 or h < 150:
-                        continue
-
-                    # 图片周围文本
-                    context_start = max(0, page_text.find(
-                        page_text[:len(page_text)//2]) if len(page_text) > 600
-                        else 0)
-                    context_text = page_text[context_start:context_start+600].strip()
-
-                    # 不支持的格式（jpx/jpeg2000等）转为 PNG
-                    if ext.lower() not in vlm_formats:
-                        try:
-                            img = Image.open(io.BytesIO(image_bytes))
-                            if img.mode in ("CMYK", "P"):
-                                img = img.convert("RGB")
-                            buf = io.BytesIO()
-                            img.save(buf, format="PNG")
-                            image_bytes = buf.getvalue()
-                            ext = "png"
-                        except Exception as e:
-                            logger.warning(f"[ImageExtractor] 格式转换失败 page={page_num+1} img={img_idx}: {e}")
-                            continue
-
-                    filename = f"page{page_num+1}_img{img_idx+1}.{ext}"
-                    output_path = Path(output_dir) / filename
-                    output_path.write_bytes(image_bytes)
-
-                    images.append({
-                        "path": str(output_path),
-                        "filename": filename,
-                        "page": page_num + 1,
-                        "context_text": context_text,
-                        "size": len(image_bytes),
-                    })
-                except Exception as e:
-                    logger.warning(f"[ImageExtractor] 提取图片失败 page={page_num+1} img={img_idx}: {e}")
-                    continue
-
-        doc.close()
+        logger.info(
+            f"[ImageExtractor] 共渲染 {len(images)} 页 from {file_path}"
+        )
        return images


@@ -1,5 +1,5 @@
 version = 1
-revision = 2
+revision = 3
 requires-python = ">=3.12"
 resolution-markers = [
    "python_full_version >= '3.13'",
@@ -441,6 +441,7 @@ dependencies = [
    { name = "opencv-python-headless" },
    { name = "pandas" },
    { name = "passlib", extra = ["bcrypt"] },
+    { name = "pdf2image" },
    { name = "psycopg", extra = ["binary"] },
    { name = "pydantic", extra = ["email"] },
    { name = "pydantic-settings" },
@@ -492,6 +493,7 @@ requires-dist = [
    { name = "opencv-python-headless", specifier = ">=4.13.0.92" },
    { name = "pandas", specifier = ">=2.2.3" },
    { name = "passlib", extras = ["bcrypt"], specifier = ">=1.7.4" },
+    { name = "pdf2image", specifier = ">=1.17.0" },
    { name = "psycopg", extras = ["binary"], specifier = ">=3.1.0" },
    { name = "pydantic", extras = ["email"], specifier = ">=2.5.0" },
    { name = "pydantic-settings", specifier = ">=2.1.0" },
@@ -2396,6 +2398,18 @@ wheels = [
    { url = "https://mirrors.aliyun.com/pypi/packages/cc/20/ff623b09d963f88bfde16306a54e12ee5ea43e9b597108672ff3a408aad6/pathspec-0.12.1-py3-none-any.whl", hash = "sha256:a0d503e138a4c123b27490a4f7beda6a01c6f288df0e4a8b79c7eb0dc7b4cc08" },
 ]

+[[package]]
+name = "pdf2image"
+version = "1.17.0"
+source = { registry = "https://mirrors.aliyun.com/pypi/simple/" }
+dependencies = [
+    { name = "pillow" },
+]
+sdist = { url = "https://mirrors.aliyun.com/pypi/packages/00/d8/b280f01045555dc257b8153c00dee3bc75830f91a744cd5f84ef3a0a64b1/pdf2image-1.17.0.tar.gz", hash = "sha256:eaa959bc116b420dd7ec415fcae49b98100dda3dd18cd2fdfa86d09f112f6d57" }
+wheels = [
+    { url = "https://mirrors.aliyun.com/pypi/packages/62/33/61766ae033518957f877ab246f87ca30a85b778ebaad65b7f74fa7e52988/pdf2image-1.17.0-py3-none-any.whl", hash = "sha256:ecdd58d7afb810dffe21ef2b1bbc057ef434dabbac6c33778a38a3f7744a27e2" },
+]
+
 [[package]]
 name = "pillow"
 version = "11.3.0"