feat: replace pymupdf image extraction with pdf2image + Heron layout detection

PDF image extraction now renders pages with pdf2image, detects figure
regions using docling-layout-heron (RT-DETRv2), and crops only the
detected pictures. Removes full-page fallback for text-only pages.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

.gitignore

@@ -71,3 +71,6 @@ Desktop.ini
 htmlcov/
 
 officefile
+data/1法律
+data/data_backup
+

backend/pyproject.toml

@@ -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",
 ]
 

backend/scripts/test_pdf_image_extract.py

@@ -0,0 +1,140 @@
+"""测试 Heron 版面检测 + 裁剪 + VLM 图片描述流程"""
+import argparse
+import asyncio
+import logging
+import sys
+import time
+from pathlib import Path
+
+sys.path.append(str(Path(__file__).parent.parent))
+
+from PIL import Image
+from src.rag.document_loaders import PDFImageExtractor
+
+logging.basicConfig(level=logging.INFO, format="%(levelname)s %(message)s")
+logger = logging.getLogger(__name__)
+
+PROJECT_ROOT = Path(__file__).parent.parent.parent  # course-agent-od/
+TEST_PDF = (
+    PROJECT_ROOT
+    / "data" / "uploads" / "testuser" / "knowledge_bases" / "植物知识图谱"
+    / "ab810d4a-b030-45c0-94d2-bd6e2a23053b.pdf"
+)
+OUTPUT_DIR = PROJECT_ROOT / "data" / "images" / "_test_pdf_extract"
+
+
+def test_extract(max_pages: int | None = None):
+    """阶段1：Heron 版面检测 + 裁剪"""
+    if not TEST_PDF.exists():
+        logger.error(f"测试PDF不存在: {TEST_PDF}")
+        sys.exit(1)
+
+    # 清理旧输出
+    if OUTPUT_DIR.exists():
+        for f in OUTPUT_DIR.iterdir():
+            f.unlink()
+
+    print(f"\n{'='*60}")
+    print(f"[阶段1] Heron 版面检测 + 裁剪测试")
+    print(f"  PDF: {TEST_PDF.name}")
+    print(f"  输出: {OUTPUT_DIR}")
+    print(f"{'='*60}\n")
+
+    t0 = time.time()
+    images = PDFImageExtractor.extract_images(
+        str(TEST_PDF), str(OUTPUT_DIR),
+    )
+    elapsed = time.time() - t0
+
+    if not images:
+        print("  ❌ 未提取到任何图片")
+        return []
+
+    if max_pages:
+        images = [img for img in images if img["page"] <= max_pages]
+
+    # 按页分组统计
+    from collections import defaultdict
+    by_page = defaultdict(list)
+    for img in images:
+        by_page[img["page"]].append(img)
+
+    total_size = 0
+    fig_count = 0
+    full_count = 0
+    for page_num in sorted(by_page.keys()):
+        page_imgs = by_page[page_num]
+        for img in page_imgs:
+            path = Path(img["path"])
+            with Image.open(path) as pil:
+                w, h = pil.size
+            size_kb = img["size"] / 1024
+            total_size += img["size"]
+            is_full = "_full." in img["filename"]
+            tag = "FULL" if is_full else "FIG"
+            if is_full:
+                full_count += 1
+            else:
+                fig_count += 1
+            print(
+                f"  page {img['page']:>3d}  [{tag}] {img['filename']:<25s}  "
+                f"{w}x{h}  {size_kb:>7.1f} KB"
+            )
+
+    print(
+        f"\n  汇总: {len(images)} 张 ({fig_count} 裁剪 + {full_count} 整页兜底), "
+        f"{total_size/1024/1024:.1f} MB, 耗时 {elapsed:.1f}s\n"
+    )
+    return images
+
+
+async def test_vlm(images: list[dict], max_images: int = 3):
+    """阶段2：VLM图片描述"""
+    from src.llm.siliconflow import get_llm_client
+
+    # 优先选裁剪图
+    fig_images = [img for img in images if "_full." not in img["filename"]]
+    targets = (fig_images or images)[:max_images]
+
+    print(f"\n{'='*60}")
+    print(f"[阶段2] VLM图片描述测试 ({len(targets)} 张)")
+    print(f"{'='*60}\n")
+
+    client = get_llm_client()
+    sem = asyncio.Semaphore(3)
+
+    async def describe_one(idx: int, img: dict):
+        async with sem:
+            t0 = time.time()
+            desc = await client.describe_image(img["path"], img.get("context_text", ""))
+            elapsed = time.time() - t0
+            return idx, desc, elapsed
+
+    tasks = [describe_one(i, img) for i, img in enumerate(targets)]
+    results = await asyncio.gather(*tasks, return_exceptions=True)
+
+    for r in results:
+        if isinstance(r, Exception):
+            print(f"  ❌ 失败: {r}\n")
+            continue
+        idx, desc, elapsed = r
+        img = targets[idx]
+        print(f"  --- page {img['page']} ({img['filename']}) {elapsed:.1f}s ---")
+        print(f"  {desc}\n")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="测试PDF图片提取")
+    parser.add_argument("--vlm", action="store_true", help="启用VLM图片描述")
+    parser.add_argument("--max-pages", type=int, default=None, help="限制提取页数")
+    parser.add_argument("--max-vlm", type=int, default=3, help="VLM描述最大图片数")
+    args = parser.parse_args()
+
+    images = test_extract(max_pages=args.max_pages)
+
+    if args.vlm and images:
+        asyncio.run(test_vlm(images, max_images=args.max_vlm))
+
+
+if __name__ == "__main__":
+    main()

backend/src/api/knowledge_base.py

@@ -790,10 +790,16 @@ async def download_document(
         if not os.path.exists(document.file_path):
             raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="文件不存在")
 
+        import mimetypes
+        mime_type, _ = mimetypes.guess_type(document.original_filename or document.file_path)
+        if not mime_type:
+            mime_type = "application/octet-stream"
+
         return FileResponse(
             document.file_path,
             filename=document.original_filename,
-            media_type="application/octet-stream"
+            media_type=mime_type,
+            content_disposition_type="inline"
         )
 
     except HTTPException:

backend/src/core/config.py

@@ -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"""

backend/src/rag/document_loaders.py

@@ -1,12 +1,12 @@
 """
-LangChain 1.0 文档加载器封装 + PDF图片提取
+LangChain 1.0 文档加载器封装 + PDF图片提取（Heron 版面检测 + 裁剪）
 """
 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 +19,256 @@ logger = logging.getLogger(__name__)
 
 
 class PDFImageExtractor:
-    """使用pymupdf从PDF中提取内嵌图片"""
+    """使用 pdf2image 渲染页面 + Heron RT-DETR 检测 Picture 区域 + 裁剪"""
+
+    # 渲染参数
+    DEFAULT_DPI = 250
+    DEFAULT_TARGET_WIDTH = 2500
+    JPEG_QUALITY = 90
+    CROP_PADDING = 10
+
+    # Heron 检测参数（参考 ZDTL config）
+    HERON_MODEL = "docling-project/docling-layout-heron"
+    PICTURE_THRESHOLD = 0.35
+    NMS_IOU = 0.3
+    MAX_AREA_RATIO = 0.45
+    MIN_SIZE_W = 100
+    MIN_SIZE_H = 80
 
     @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 _nms(boxes: list[dict], iou_threshold: float) -> list[dict]:
+        """IoU-based Non-Maximum Suppression"""
+        if not boxes:
+            return boxes
 
-        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。
+        import torch
+
+        bboxes = torch.tensor([b["bbox"] for b in boxes], dtype=torch.float32)
+        scores = torch.tensor([b["score"] for b in boxes])
+
+        x1 = bboxes[:, 0]
+        y1 = bboxes[:, 1]
+        x2 = bboxes[:, 2]
+        y2 = bboxes[:, 3]
+        areas = (x2 - x1) * (y2 - y1)
+
+        _, order = scores.sort(descending=True)
+        keep = []
+        while order.numel() > 0:
+            if order.numel() == 1:
+                keep.append(order.item())
+                break
+            i = order[0].item()
+            keep.append(i)
+
+            xx1 = torch.max(x1[i], x1[order[1:]])
+            yy1 = torch.max(y1[i], y1[order[1:]])
+            xx2 = torch.min(x2[i], x2[order[1:]])
+            yy2 = torch.min(y2[i], y2[order[1:]])
+
+            inter = (xx2 - xx1).clamp(min=0) * (yy2 - yy1).clamp(min=0)
+            union = areas[i] + areas[order[1:]] - inter
+            iou = inter / union
+            mask = iou <= iou_threshold
+            order = order[1:][mask]
+
+        return [boxes[i] for i in keep]
+
+    @staticmethod
+    def _detect_pictures(
+        pil_images: list,
+        model_name: str = None,
+        picture_threshold: float = None,
+        nms_iou: float = None,
+        max_area_ratio: float = None,
+        min_size_w: int = None,
+        min_size_h: int = None,
+    ) -> dict[int, list[dict]]:
+        """用 Heron RT-DETR 检测每页中的 Picture 区域
+
+        Returns:
+            {page_index: [{bbox: [x1,y1,x2,y2], score: float}, ...]}
         """
+        import torch
+        from transformers import RTDetrV2ForObjectDetection, RTDetrImageProcessor
+
+        model_name = model_name or PDFImageExtractor.HERON_MODEL
+        picture_threshold = picture_threshold or PDFImageExtractor.PICTURE_THRESHOLD
+        nms_iou = nms_iou or PDFImageExtractor.NMS_IOU
+        max_area_ratio = max_area_ratio or PDFImageExtractor.MAX_AREA_RATIO
+        min_size_w = min_size_w or PDFImageExtractor.MIN_SIZE_W
+        min_size_h = min_size_h or PDFImageExtractor.MIN_SIZE_H
+
+        logger.info(f"[ImageExtractor] 加载 Heron 模型: {model_name}")
+        processor = RTDetrImageProcessor.from_pretrained(
+            model_name, local_files_only=True
+        )
+        model = RTDetrV2ForObjectDetection.from_pretrained(
+            model_name, local_files_only=True
+        )
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+        model.to(device)
+        model.eval()
+        logger.info(f"[ImageExtractor] 模型加载完成, device={device}")
+
+        results: dict[int, list[dict]] = {}
+
+        for idx, pil_img in enumerate(pil_images):
+            img = pil_img.convert("RGB")
+            W, H = img.size
+            inputs = processor(images=img, return_tensors="pt").to(device)
+
+            with torch.no_grad():
+                outputs = model(**inputs)
+
+            detections = processor.post_process_object_detection(
+                outputs, threshold=0.1, target_sizes=[(H, W)]
+            )[0]
+
+            pictures = []
+            for score, label, box in zip(
+                detections["scores"], detections["labels"], detections["boxes"]
+            ):
+                s = score.item()
+                l = label.item()
+                x1, y1, x2, y2 = box.tolist()
+                area = (x2 - x1) * (y2 - y1) / (W * H)
+                w, h = x2 - x1, y2 - y1
+
+                if area > max_area_ratio or w < min_size_w or h < min_size_h:
+                    continue
+
+                cls = model.config.id2label.get(l, str(l))
+                if cls == "picture" and s >= picture_threshold:
+                    pictures.append({
+                        "bbox": [round(v, 1) for v in [x1, y1, x2, y2]],
+                        "score": round(s, 3),
+                    })
+
+            # NMS + 按位置排序
+            pictures = PDFImageExtractor._nms(pictures, nms_iou)
+            pictures.sort(key=lambda b: (b["bbox"][1] // 200, b["bbox"][0]))
+            results[idx] = pictures
+
+            logger.info(
+                f"[ImageExtractor] page {idx + 1}: "
+                f"{len(pictures)} pictures detected"
+            )
+
+        return results
+
+    @staticmethod
+    def extract_images(
+        file_path: str,
+        output_dir: str,
+        dpi: int = None,
+        target_width: int = None,
+    ) -> List[dict]:
+        """将PDF每页渲染为图片，用 Heron 检测并裁剪出图表区域
+
+        若某页无检测到 Picture，则保存整页图作为兜底。
+
+        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
+        padding = PDFImageExtractor.CROP_PADDING
+
         Path(output_dir).mkdir(parents=True, exist_ok=True)
-        images = []
+
+        # Phase 1: pymupdf 提取页面文本
+        page_texts: list[str] = []
+        try:
             doc = fitz.open(file_path)
-
-        TILE_THRESHOLD = 50
-        vlm_formats = {"png", "jpg", "jpeg", "webp", "gif"}
-        dpi = 150  # 页面渲染 DPI
-
             for page_num in range(len(doc)):
-            page = doc[page_num]
-            page_text = page.get_text("text")
-            image_list = page.get_images(full=True)
-
-            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}")
-                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
-
+                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}")
+
+        # Phase 2: pdf2image 渲染页面
+        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 []
+
+        # Phase 3: 缩放页面图
+        scaled_images = []
+        for pil_img in pil_images:
+            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
+                )
+            scaled_images.append(pil_img)
+
+        # Phase 4: Heron 检测 Picture 区域
+        try:
+            detections = PDFImageExtractor._detect_pictures(scaled_images)
+        except Exception as e:
+            logger.error(f"[ImageExtractor] Heron检测失败: {e}")
+            return []
+
+        # Phase 5: 裁剪 + 保存
+        images = []
+        for page_idx, pil_img in enumerate(scaled_images):
+            page_num = page_idx + 1
+            W, H = pil_img.size
+            context_text = ""
+            if page_idx < len(page_texts):
+                context_text = page_texts[page_idx][:600].strip()
+
+            pics = detections.get(page_idx, [])
+
+            if pics:
+                # 裁剪检测到的 Picture 区域
+                for fig_idx, pic in enumerate(pics):
+                    try:
+                        x1, y1, x2, y2 = pic["bbox"]
+                        x1 = max(0, int(x1) - padding)
+                        y1 = max(0, int(y1) - padding)
+                        x2 = min(W, int(x2) + padding)
+                        y2 = min(H, int(y2) + padding)
+
+                        crop = pil_img.crop((x1, y1, x2, y2))
+                        filename = f"page{page_num}_fig{fig_idx + 1}.jpg"
+                        output_path = Path(output_dir) / filename
+                        crop.save(str(output_path), format="JPEG", quality=quality)
+
+                        images.append({
+                            "path": str(output_path),
+                            "filename": filename,
+                            "page": page_num,
+                            "context_text": context_text,
+                            "size": output_path.stat().st_size,
+                        })
+                        logger.info(
+                            f"[ImageExtractor] 裁剪 page={page_num} "
+                            f"fig={fig_idx + 1} bbox={pic['bbox']} "
+                            f"size={images[-1]['size']}"
+                        )
+                    except Exception as e:
+                        logger.warning(
+                            f"[ImageExtractor] 裁剪失败 page={page_num} "
+                            f"fig={fig_idx + 1}: {e}"
+                        )
+            else:
+                continue
+
+        logger.info(
+            f"[ImageExtractor] 完成: {len(images)} 张图片 from {file_path}"
+        )
         return images
 
 

backend/uv.lock

@@ -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"

web/src/app/(main)/knowledge/[id]/page.tsx

@@ -33,7 +33,7 @@ import {
   Settings
 } from "lucide-react";
 import { formatFileSize, formatDate } from "@/lib/utils";
-import { knowledgeBaseAPI } from "@/lib/api";
+import { knowledgeBaseAPI, resolveImageUrl } from "@/lib/api";
 import { KnowledgeBaseDetail, Document } from "@/types";
 
 export default function KnowledgeBaseDetailPage() {
@@ -158,13 +158,12 @@ export default function KnowledgeBaseDetailPage() {
   const handleViewDocument = async (doc: Document) => {
     try {
       const token = localStorage.getItem("auth_token");
-      const res = await fetch(`/api/knowledge-bases/documents/${doc.id}/download`, {
+      const res = await fetch(resolveImageUrl(`/knowledge-bases/documents/${doc.id}/download`), {
         headers: { Authorization: `Bearer ${token}` },
       });
-      if (!res.ok) throw new Error("下载失败");
+      if (!res.ok) throw new Error("获取文档失败");
       const blob = await res.blob();
 
-      // 从 Content-Disposition 提取文件名，或用 doc 信息拼接
       const disposition = res.headers.get("Content-Disposition");
       let filename = `${doc.title}${doc.file_type}`;
       if (disposition) {
@@ -172,15 +171,11 @@ export default function KnowledgeBaseDetailPage() {
         if (match) filename = decodeURIComponent(match[1].replace(/["']/g, ""));
       }
 
-      // 触发浏览器下载
       const url = URL.createObjectURL(blob);
-      const a = document.createElement("a");
-      a.href = url;
-      a.download = filename;
-      a.click();
-      URL.revokeObjectURL(url);
+      window.open(url, '_blank');
+      setTimeout(() => URL.revokeObjectURL(url), 60000);
     } catch {
-      window.open(`/api/knowledge-bases/documents/${doc.id}/download`, '_blank');
+      window.open(resolveImageUrl(`/knowledge-bases/documents/${doc.id}/download`), '_blank');
     }
   };
 

web/src/app/(main)/spatial/image-to-image/page.tsx

@@ -12,7 +12,7 @@ import { Badge } from "@/components/ui/badge";
 import { Separator } from "@/components/ui/separator";
 import { Slider } from "@/components/ui/slider";
 import { Loader2, Download, Copy, Upload, Edit3, Wand2, Expand, Palette, Image as ImageIcon, X } from "lucide-react";
-import { imageAPI } from "@/lib/api";
+import { imageAPI, resolveImageUrl } from "@/lib/api";
 import { toast } from "sonner";
 
 interface EditResult {
@@ -312,14 +312,22 @@ export default function ImageToImagePage() {
     }
   };
 
-  const handleDownload = (imageUrl: string, imageId: string) => {
+  const handleDownload = async (imageUrl: string, imageId: string) => {
+    try {
+      const res = await fetch(imageUrl);
+      const blob = await res.blob();
+      const url = URL.createObjectURL(blob);
       const link = document.createElement("a");
-    link.href = imageUrl;
+      link.href = url;
       link.download = `edited-image-${imageId}.png`;
       document.body.appendChild(link);
       link.click();
       document.body.removeChild(link);
+      URL.revokeObjectURL(url);
       toast.success("图像下载成功");
+    } catch {
+      toast.error("图像下载失败");
+    }
   };
 
   return (
@@ -595,7 +603,7 @@ export default function ImageToImagePage() {
                   <div className="space-y-4">
                     <div className="relative group">
                       <img
-                        src={editResult.url}
+                        src={resolveImageUrl(editResult.url)}
                         alt="Edited image"
                         className="w-full h-64 object-cover rounded-lg"
                       />
@@ -604,7 +612,7 @@ export default function ImageToImagePage() {
                         <Button
                           size="sm"
                           variant="secondary"
-                          onClick={() => handleDownload(editResult.url, editResult.id)}
+                          onClick={() => handleDownload(resolveImageUrl(editResult.url), editResult.id)}
                         >
                           <Download className="h-3 w-3" />
                         </Button>
@@ -648,7 +656,7 @@ export default function ImageToImagePage() {
                       >
                         <div className="relative">
                           <img
-                            src={variation.url}
+                            src={resolveImageUrl(variation.url)}
                             alt={`Variation ${index + 1}`}
                             className="w-full h-48 object-cover rounded-lg"
                           />
@@ -657,7 +665,7 @@ export default function ImageToImagePage() {
                             <Button
                               size="sm"
                               variant="secondary"
-                              onClick={() => handleDownload(variation.url, variation.id)}
+                              onClick={() => handleDownload(resolveImageUrl(variation.url), variation.id)}
                             >
                               <Download className="h-3 w-3" />
                             </Button>

web/src/app/(main)/spatial/page.tsx

@@ -32,7 +32,7 @@ import {
   X
 } from "lucide-react";
 import LoadingSpinner from "@/components/ui/loading-spinner";
-import { imageAPI } from "@/lib/api";
+import { imageAPI, resolveImageUrl } from "@/lib/api";
 import { toast } from "sonner";
 import { motion } from "framer-motion";
 
@@ -350,14 +350,22 @@ export default function SpatialPage() {
     }
   };
 
-  const handleDownload = (imageUrl: string, imageId: string) => {
+  const handleDownload = async (imageUrl: string, imageId: string) => {
+    try {
+      const res = await fetch(imageUrl);
+      const blob = await res.blob();
+      const url = URL.createObjectURL(blob);
       const link = document.createElement("a");
-    link.href = imageUrl;
+      link.href = url;
       link.download = `image-${imageId}.png`;
       document.body.appendChild(link);
       link.click();
       document.body.removeChild(link);
+      URL.revokeObjectURL(url);
       toast.success("图像下载成功");
+    } catch {
+      toast.error("图像下载失败");
+    }
   };
 
   if (authLoading) {
@@ -580,7 +588,7 @@ export default function SpatialPage() {
                                 <Card className="overflow-hidden">
                                   <div className="aspect-square relative bg-muted">
                                     <img
-                                      src={image.url}
+                                      src={resolveImageUrl(image.url)}
                                       alt={`Generated image ${index + 1}`}
                                       className="w-full h-full object-cover"
                                       loading="lazy"
@@ -599,7 +607,7 @@ export default function SpatialPage() {
                                       variant="outline"
                                       size="sm"
                                       className="w-full"
-                                      onClick={() => handleDownload(image.url, image.id)}
+                                      onClick={() => handleDownload(resolveImageUrl(image.url), image.id)}
                                     >
                                       <Download className="h-4 w-4 mr-2" />
                                       下载图片
@@ -852,7 +860,7 @@ export default function SpatialPage() {
                           <div className="space-y-4">
                             <div className="relative group bg-muted rounded-lg overflow-hidden">
                               <img
-                                src={editResult.url}
+                                src={resolveImageUrl(editResult.url)}
                                 alt="Edited image"
                                 className="w-full h-64 object-contain"
                                 loading="lazy"
@@ -871,7 +879,7 @@ export default function SpatialPage() {
                                 variant="outline"
                                 size="sm"
                                 className="w-full"
-                                onClick={() => handleDownload(editResult.url, editResult.id)}
+                                onClick={() => handleDownload(resolveImageUrl(editResult.url), editResult.id)}
                               >
                                 <Download className="h-4 w-4 mr-2" />
                                 下载图片
@@ -907,7 +915,7 @@ export default function SpatialPage() {
                                 <Card className="overflow-hidden">
                                   <div className="relative bg-muted">
                                     <img
-                                      src={variation.url}
+                                      src={resolveImageUrl(variation.url)}
                                       alt={`Variation ${index + 1}`}
                                       className="w-full h-48 object-cover"
                                       loading="lazy"
@@ -922,7 +930,7 @@ export default function SpatialPage() {
                                       variant="outline"
                                       size="sm"
                                       className="w-full"
-                                      onClick={() => handleDownload(variation.url, variation.id)}
+                                      onClick={() => handleDownload(resolveImageUrl(variation.url), variation.id)}
                                     >
                                       <Download className="h-4 w-4 mr-2" />
                                       下载图片

web/src/app/(main)/spatial/text-to-image/page.tsx

@@ -11,7 +11,7 @@ import { Textarea } from "@/components/ui/textarea";
 import { Badge } from "@/components/ui/badge";
 import { Separator } from "@/components/ui/separator";
 import { Loader2, Download, Copy, RefreshCw, Sparkles, Image as ImageIcon } from "lucide-react";
-import { imageAPI } from "@/lib/api";
+import { imageAPI, resolveImageUrl } from "@/lib/api";
 import { toast } from "sonner";
 
 interface GeneratedImage {
@@ -95,14 +95,22 @@ export default function TextToImagePage() {
     }
   };
 
-  const handleDownload = (imageUrl: string, imageId: string) => {
+  const handleDownload = async (imageUrl: string, imageId: string) => {
+    try {
+      const res = await fetch(imageUrl);
+      const blob = await res.blob();
+      const url = URL.createObjectURL(blob);
       const link = document.createElement("a");
-    link.href = imageUrl;
+      link.href = url;
       link.download = `generated-image-${imageId}.png`;
       document.body.appendChild(link);
       link.click();
       document.body.removeChild(link);
+      URL.revokeObjectURL(url);
       toast.success("图像下载成功");
+    } catch {
+      toast.error("图像下载失败");
+    }
   };
 
   const handleCopyPrompt = (prompt: string) => {
@@ -316,7 +324,7 @@ export default function TextToImagePage() {
                         <Card className="overflow-hidden">
                           <div className="aspect-square relative">
                             <img
-                              src={image.url}
+                              src={resolveImageUrl(image.url)}
                               alt={`Generated image ${index + 1}`}
                               className="w-full h-full object-cover"
                             />
@@ -326,7 +334,7 @@ export default function TextToImagePage() {
                                 <Button
                                   size="sm"
                                   variant="secondary"
-                                  onClick={() => handleDownload(image.url, image.id)}
+                                  onClick={() => handleDownload(resolveImageUrl(image.url), image.id)}
                                 >
                                   <Download className="h-3 w-3" />
                                 </Button>

web/src/components/chat/message-item.tsx

@@ -113,7 +113,18 @@ export default function MessageItem({ message, selectedModel }: MessageItemProps
 
   const handleCopy = async () => {
     try {
+      if (navigator.clipboard && window.isSecureContext) {
         await navigator.clipboard.writeText(message.content);
+      } else {
+        const textarea = document.createElement("textarea");
+        textarea.value = message.content;
+        textarea.style.position = "fixed";
+        textarea.style.left = "-9999px";
+        document.body.appendChild(textarea);
+        textarea.select();
+        document.execCommand("copy");
+        document.body.removeChild(textarea);
+      }
       toast.success("已复制");
     } catch {
       toast.error("复制失败");
@@ -254,7 +265,7 @@ export default function MessageItem({ message, selectedModel }: MessageItemProps
                     ),
                     img: ({ src, alt }: any) => {
                       const resolvedSrc = src && src.startsWith("/")
-                        ? `${process.env.NEXT_PUBLIC_API_URL || `${window.location.protocol}//${window.location.hostname}:8000`}${src}`
+                        ? `${process.env.NEXT_PUBLIC_API_URL || `${window.location.protocol}//${window.location.hostname}:8002`}${src}`
                         : src;
                       return (
                       <a href={resolvedSrc} target="_blank" rel="noopener noreferrer" className="block my-3 group">

web/src/lib/api.ts

@@ -12,7 +12,14 @@ import type {
 } from "@/types";
 
 // API基础配置
-const API_BASE_URL = process.env.NEXT_PUBLIC_API_URL || (typeof window !== 'undefined' ? `${window.location.protocol}//${window.location.hostname}:8000` : "http://127.0.0.1:8000");
+const API_BASE_URL = process.env.NEXT_PUBLIC_API_URL || (typeof window !== 'undefined' ? `${window.location.protocol}//${window.location.hostname}:8002` : "http://127.0.0.1:8002");
+
+// 将后端返回的相对路径图片URL转为完整的后端地址
+export function resolveImageUrl(url: string): string {
+  if (!url) return url;
+  if (url.startsWith("http://") || url.startsWith("https://") || url.startsWith("data:")) return url;
+  return `${API_BASE_URL}${url.startsWith("/") ? "" : "/"}${url}`;
+}
 
 // 请求拦截器
 async function apiRequest<T>(

web/src/store/chat.ts

@@ -339,6 +339,15 @@ export const useChatStore = create<ChatStore>((set, get) => ({
           scheduleFlush();
         },
         (_sessionId: number, messageId?: number, userMessageId?: number) => {
+          // 刷出缓冲区中剩余的内容
+          if (rafId !== null) {
+            cancelAnimationFrame(rafId);
+            rafId = null;
+          }
+          if (contentBuffer) {
+            flushBuffer();
+          }
+
           if (messageId) {
             set((state) => ({
               messages: state.messages.map(msg => {