feat: multimodal RAG with PDF image extraction and display

Extract images from PDFs using pymupdf, generate descriptions via
Qwen3-VL-8B, store in ChromaDB alongside text chunks, and render
images in chat answers. Includes image proxy rewrite, force re-process
endpoint, and VLM API timeout.

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

backend/src/rag/vector_store.py

@@ -65,6 +65,17 @@ class VectorStore:
         print(f"[DEBUG-VectorStore] 返回结果数量: {len(result)}")
         return result
     
+    def delete_by_document_id(self, document_id: int) -> bool:
+        """删除指定文档的所有向量数据"""
+        try:
+            self.vectorstore._collection.delete(
+                where={"document_id": document_id}
+            )
+            return True
+        except Exception as e:
+            print(f"删除向量数据失败: {str(e)}")
+            return False
+
     def max_marginal_relevance_search(
         self,
         query: str,