{"id":230,"date":"2025-05-04T23:11:03","date_gmt":"2025-05-04T15:11:03","guid":{"rendered":"https:\/\/knoka.vip\/?p=230"},"modified":"2025-05-06T09:43:03","modified_gmt":"2025-05-06T01:43:03","slug":"%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f","status":"publish","type":"post","link":"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/","title":{"rendered":"\u4e00\u6587\u5feb\u901f\u5b66\u4f1a\u57fa\u4e8eMilvus\u6784\u5efa\u591a\u6a21\u6001RAG\u95ee\u7b54\u7cfb\u7edf"},"content":{"rendered":"\n

0\u3001\u7cfb\u7edf\u76ee\u6807<\/h3>\n\n\n\n

\u672c\u7cfb\u7edf\u9884\u671f\u76ee\u6807\u662f\u6784\u5efa\u4e00\u4e2a\u591a\u6a21\u6001\u7684RAG\u7cfb\u7edf\uff0c\u6b64\u7cfb\u7edf\u53ef\u5b9e\u73b0\u591a\u79cd\u6a21\u6001\u7684\u6570\u636e\u5bfc\u5165\uff0c\u5305\u62ec\u5982PDF\u3001Word\u3001Excel\u3001\u56fe\u7247\u7b49\uff08\u4ec5\u7ed3\u6784\u5316\u6570\u636e\uff09\uff0c\u5e76\u53ef\u901a\u8fc7\u5bf9\u8bdd\u8fdb\u884c\u68c0\u7d22\uff0c\u8fdb\u884c\u6574\u7406\u548c\u8fd4\u56de\u7ed3\u679c\uff0c\u8fd4\u56de\u7ed3\u679c\u548c\u5305\u62ec\u56fe\u7247\u548c\u6587\u672c\u7b49\u3002<\/p>\n\n\n\n

1\u3001\u670d\u52a1\u7aef\u51c6\u5907<\/h3>\n\n\n\n

\u6211\u4eec\u8fd9\u91cc\u91c7\u7528Milvus\u4f5c\u4e3a\u6211\u4eec\u7684RAG\u57fa\u7840\uff0c\u5c06\u5176\u5728\u4e91\u670d\u52a1\u5668\u4e0a\u4e8b\u5148\u5b8c\u6210\u914d\u7f6e\uff0c\u5e76\u6d4b\u8bd5\u6210\u529f\u8fdb\u884c\u94fe\u63a5\uff0c\u6b64\u5904\u53ef\u53c2\u8003\u672c\u4eba\u7684\u4e0a\u4e00\u7ae0\u6559\u7a0b\uff0c\u8fd9\u91cc\u4e0d\u518d\u8be6\u7ec6\u9610\u8ff0<\/p>\n\n\n\n

\u4e00\u6587\u5b66\u4f1a\u4e91\u670d\u52a1\u5668\u914d\u7f6eMilvus\u5411\u91cf\u6570\u636e\u5e93 – \u4eca\u5929\u5f00\u59cb\u5b66AI<\/a><\/p>\n\n\n\n

2\u3001\u591a\u6a21\u6001RAG\u6784\u5efa<\/h3>\n\n\n\n

2.1 \u6a21\u578b\u51c6\u5907<\/h4>\n\n\n\n

\u9996\u5148\u6211\u4eec\u9700\u8981\u51c6\u5907\u6211\u4eec\u7684\u8bcd\u5d4c\u5165\u6a21\u578b\uff0c\u7531\u4e8e\u60f3\u5b9e\u73b0\u7684\u662f\u4e00\u4e2a\u591a\u6a21\u6001\u7684RAG\uff0c\u56e0\u6b64\u9700\u8981\u9009\u62e9\u5408\u9002\u7684\u5d4c\u5165\u6a21\u578b\uff0c\u8fd9\u91cc\u6240\u9009\u62e9\u7684\u6a21\u578b\u4e3a\u901a\u7528\u591a\u8bed\u8a00\u591a\u6a21\u6001\u5411\u91cf\u6a21\u578b \u2014 jina-clip-v2<\/code>\uff0c\u8be5\u6a21\u578b\u57fa\u4e8e jina-clip-v1<\/code> \u548c jina-embeddings-3<\/code> \u6784\u5efa\uff0c\u5e76\u5b9e\u73b0\u4e86\u591a\u9879\u5173\u952e\u6539\u8fdb\uff1a<\/p>\n\n\n\n

    \n
  • \u6027\u80fd\u63d0\u5347\uff1av2 \u5728\u6587\u672c-\u56fe\u50cf\u548c\u6587\u672c-\u6587\u672c\u68c0\u7d22\u4efb\u52a1\u4e2d\uff0c\u6027\u80fd\u8f83 v1 \u63d0\u5347\u4e86 3%\u3002\u6b64\u5916\uff0c\u4e0e v1 \u7c7b\u4f3c\uff0cv2 \u7684\u6587\u672c\u7f16\u7801\u5668\u4e5f\u80fd\u9ad8\u6548\u5730\u5e94\u7528\u4e8e\u591a\u8bed\u8a00\u957f\u6587\u672c\u5bc6\u96c6\u68c0\u7d22\u7d22\uff0c\u5176\u6027\u80fd\u53ef\u4e0e\u6211\u4eec\u76ee\u524d\u6700\u5148\u8fdb\u7684\u6a21\u578b \u2014\u2014 \u53c2\u6570\u91cf\u4f4e\u4e8e 1B \u7684\u6700\u4f73\u591a\u8bed\u8a00\u5411\u91cf\u6a21\u578b jina-embeddings-v3\uff08\u57fa\u4e8e MTEB \u6392\u884c\u699c\uff09\u2014\u2014 \u76f8\u5ab2\u7f8e\uff1b<\/li>\n\n\n\n
  • \u591a\u8bed\u8a00\u652f\u6301\uff1a\u4ee5 jina-embeddings-v3 \u4f5c\u4e3a\u6587\u672c\u5854\uff0cjina-clip-v2 \u652f\u6301 89 \u79cd\u8bed\u8a00\u7684\u591a\u8bed\u8a00\u56fe\u50cf\u68c0\u7d22\uff0c\u5e76\u5728\u8be5\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\u76f8\u6bd4 nllb-clip-large-siglip \u63d0\u5347\u4e86 4%\uff1b<\/li>\n\n\n\n
  • \u66f4\u9ad8\u56fe\u50cf\u5206\u8fa8\u7387\uff1av2 \u652f\u6301 512×512 \u50cf\u7d20\u7684\u8f93\u5165\u56fe\u50cf\u5206\u8fa8\u7387\uff0c\u76f8\u6bd4 v1 \u7684 224×224 \u6709\u4e86\u5927\u5e45\u63d0\u5347\u3002\u80fd\u591f\u66f4\u597d\u5730\u6355\u6349\u56fe\u50cf\u7ec6\u8282\uff0c\u63d0\u5347\u7279\u5f81\u63d0\u53d6\u7684\u7cbe\u5ea6\uff0c\u5e76\u66f4\u51c6\u786e\u5730\u8bc6\u522b\u7ec6\u7c92\u5ea6\u89c6\u89c9\u5143\u7d20\uff1b<\/li>\n\n\n\n
  • \u53ef\u53d8\u7ef4\u5ea6\u8f93\u51fa\uff1ajina-clip-v2 \u5f15\u5165\u4e86\u4fc4\u7f57\u65af\u5957\u5a03\u8868\u793a\u5b66\u4e60\uff08Matryoshka Representation Learning\uff0cMRL\uff09\u6280\u672f\uff0c\u53ea\u9700\u8bbe\u7f6e dimensions \u53c2\u6570\uff0c\u5373\u53ef\u83b7\u53d6\u6307\u5b9a\u7ef4\u5ea6\u7684\u5411\u91cf\u8f93\u51fa\uff0c\u4e14\u5728\u51cf\u5c11\u5b58\u50a8\u6210\u672c\u7684\u540c\u65f6\uff0c\u4fdd\u6301\u5f3a\u5927\u7684\u6027\u80fd\uff1b<\/li>\n<\/ul>\n\n\n\n

    \u8fd9\u91cc\u6211\u4eec\u9996\u5148\u5b89\u88c5modelscope\u5e93\uff0c\u63a5\u7740\u62c9\u53d6\u6a21\u578b\u5373\u53ef\uff0c\u6ce8\u610f\u8c03\u6574\u597d\u6a21\u578b\u8def\u5f84\uff08xlm-roberta-flash-implementation\u53ef\u80fd\u9700\u8981\u624b\u52a8\u8f6c\u79fb\u5230.cache\/huggingface\/modules\/transformers_modules\/xlm-roberta-flash-implementation\u5185\uff09<\/p>\n\n\n\n

    pip install modelscope\nmodelscope download --model jinaai\/jina-clip-v2 --local_dir .\/jinaai\/jina-clip-v2\nmodelscope download --model jinaai\/jina-clip-implementation --local_dir .\/jinaai\/jina-clip-implementation\nmodelscope download --model jinaai\/jina-embeddings-v3 --local_dir .\/jinaai\/jina-embeddings-v3\nmodelscope download --model jinaai\/xlm-roberta-flash-implementation --local_dir .\/jinaai\/xlm-roberta-flash-implementation<\/code><\/pre>\n\n\n
    \n
    \"\"<\/div><\/figure>\n<\/div>\n\n\n
    2.2 \u5411\u91cf\u6570\u636e\u5e93\u6784\u5efa<\/h4>\n\n\n\n
    \u8fd9\u91cc\u6211\u4eec\u91c7\u7528Python\u8fdb\u884c\uff0c\u9996\u5148\u8c03\u7528pymilvus\u5e93\uff0c\u901a\u8fc7connections\u8fde\u63a5\u6570\u636e\u5e93\uff0c\u4ee3\u7801\u5982\u4e0b<\/p>\n\n\n\n
    from pymilvus import utility, Collection\nfrom pymilvus import CollectionSchema, FieldSchema, DataType\nfrom pymilvus import connections\n\n# \u5c1d\u8bd5\u8fde\u63a5\u5230 Milvus \u670d\u52a1\u5668\ntry:\n    connections.connect(host='111.6.167.30', port=19530)\n    print(f\"\u6210\u529f\u8fde\u63a5\u5230 Milvus \u670d\u52a1\u5668\uff0c\u7aef\u53e3\u4e3a\uff1a{19530}\")\nexcept Exception as e:\n    print(f\"\u8fde\u63a5\u5931\u8d25\uff1a{e}\")<\/code><\/pre>\n\n\n\n
    \u63a5\u7740\u521b\u5efa\u4e24\u4e2a\u5b57\u6bb5\uff0c\u5206\u522b\u4e3a\u4e3b\u952e\u5b57\u6bb5\u548c\u5411\u91cf\u5b57\u6bb5<\/p>\n\n\n\n
      \n
    • \u4e3b\u952e\u5b57\u6bb5 \uff1a\u521b\u5efa\u540d\u4e3a id \u7684\u81ea\u589e\u4e3b\u952e\u5b57\u6bb5\uff0c\u7528\u4e8e\u552f\u4e00\u6807\u8bc6\u6bcf\u6761\u6570\u636e\u8bb0\u5f55\uff0c\u65e0\u9700\u624b\u52a8\u6307\u5b9a\u503c\uff1b<\/li>\n\n\n\n
    • \u5411\u91cf\u5b57\u6bb5 \uff1a\u5b9a\u4e49\u540d\u4e3a vector \u7684\u6d6e\u70b9\u5411\u91cf\u5b57\u6bb5\uff0c\u7ef4\u5ea6\u4e3a 1024\uff08\u6a21\u578b\u51b3\u5b9a\uff09\uff0c\u7528\u4e8e\u5b58\u50a8\u6587\u672c\/\u56fe\u50cf\u7684\u5d4c\u5165\u5411\u91cf\uff1b<\/li>\n<\/ul>\n\n\n\n
      \u5e76\u8bbe\u7f6e\u52a8\u6001\u6269\u5c55\uff0c\u542f\u7528 enable_dynamic_field \u7279\u6027\uff0c\u5141\u8bb8\u5728\u63d2\u5165\u6570\u636e\u65f6\u52a8\u6001\u6dfb\u52a0\u672a\u9884\u5148\u5b9a\u4e49\u7684\u5b57\u6bb5\uff08\u5982 text \u6587\u672c\u5185\u5bb9\u6216 image_url \u56fe\u7247\u5730\u5740\uff09\uff1b<\/p>\n\n\n\n
      \u96c6\u5408\u540d\u79f0\u8bbe\u7f6e\u4e3amultimodal_rag_demo\uff0c\u5206\u7247\u7b56\u7565\u8bbe\u7f6e\u4e3ashards_num=2\uff0c\u4f18\u5316\u6570\u636e\u5b58\u50a8\u548c\u67e5\u8be2\u7684\u5e76\u53d1\u6027\u80fd\uff0c\u7d22\u5f15\u7c7b\u578b\u4e3aIVF_FLAT\uff08\u57fa\u4e8e\u805a\u7c7b\u7684\u5012\u6392\u7d22\u5f15\uff09\uff0c\u5176\u9002\u5408\u5927\u89c4\u6a21\u6570\u636e\uff08\u767e\u4e07\u7ea7\u4ee5\u4e0a\uff09\uff0c\u901a\u8fc7\u5206\u7c07\u52a0\u901f\u641c\u7d22\uff0c\u5e76\u5c06nlist=1024<\/code>\uff0c\u5c06\u6570\u636e\u5212\u5206\u4e3a 1024 \u4e2a\u805a\u7c7b\u4e2d\u5fc3\uff0c\u5e73\u8861\u641c\u7d22\u901f\u5ea6\u4e0e\u7cbe\u5ea6\uff0cmetric_type=COSINE<\/code>\uff1a\u4f7f\u7528\u4f59\u5f26\u76f8\u4f3c\u5ea6\u8861\u91cf\u5411\u91cf\u95f4\u76f8\u4f3c\u6027\uff1b<\/p>\n\n\n\n
      collection_name=\"multimodal_rag_2\"\n# \u5b9a\u4e49\u96c6\u5408\u53c2\u6570\nid_field = FieldSchema(\n    name=\"id\",\n    dtype=DataType.INT64,\n    is_primary=True,\n    auto_id=True\n)\n\nvector_field = FieldSchema(\n    name=\"vector\",\n    dtype=DataType.FLOAT_VECTOR,\n    dim=1024\n)\n\n# \u521b\u5efa\u96c6\u5408\u6a21\u5f0f\uff08\u542f\u7528\u52a8\u6001\u5b57\u6bb5\uff09\nschema = CollectionSchema(\n    fields=[id_field, vector_field],\n    description=\"\u591a\u6a21\u6001RAG\u5411\u91cf\u5b58\u50a8\",\n    enable_dynamic_field=True  # \u5141\u8bb8\u63d2\u5165\u989d\u5916\u5b57\u6bb5\uff08\u5982text\/image_url\uff09\n)\n\n# \u521b\u5efa\u96c6\u5408\ncollection = Collection(\n    name=collection_name,\n    schema=schema,\n    shards_num=2\n)\n\n# \u521b\u5efa\u7d22\u5f15\nindex_params = {\n    \"metric_type\": \"COSINE\",\n    \"index_type\":\"IVF_FLAT\",\n    \"params\":{\"nlist\":1024}\n}\n\ncollection.create_index(\n    field_name=\"vector\",\n    index_params=index_params\n)\n\nprint(\"\u7d22\u5f15\u521b\u5efa\u6210\u529f\")<\/code><\/pre>\n\n\n\n
      \u521b\u5efa\u5b8c\u6210\u540e\u6211\u4eec\u53ef\u5c1d\u8bd5\u8fdb\u884c\u6570\u636e\u63d2\u5165\u6d4b\u8bd5\uff0c\u793a\u4f8b\u5982\u4e0b<\/p>\n\n\n\n
      # \u63d2\u5165\u6570\u636e\uff08\u793a\u4f8b\u6570\u636e\uff09\ndata_to_insert = [\n    {\"text\": \"\u793a\u4f8b\u6587\u672c1\", \"vector\": [0.1]*1024},\n    {\"image_url\": \"http:\/\/example.com\/img1.jpg\", \"vector\": [0.2]*1024},\n    # \u6dfb\u52a0\u66f4\u591a\u6570\u636e...\n]\n\ntry:\n    insert_result = collection.insert(data_to_insert)\n    print(f\"\u63d2\u5165\u6210\u529f\uff0c\u63d2\u5165\u884c\u6570: {insert_result.insert_count}\")\nexcept MilvusException as e:\n    print(f\"\u63d2\u5165\u5931\u8d25: {e}\")\n\n# \u52a0\u8f7d\u96c6\u5408\ncollection.load()\nprint(\"\u96c6\u5408\u52a0\u8f7d\u5b8c\u6210\")\n\n# \u9a8c\u8bc1\u6570\u636e\nprint(f\"\u96c6\u5408\u884c\u6570: {collection.num_entities}\")\n<\/code><\/pre>\n\n\n\n
      3\u3001\u591a\u6a21\u6001RAG\u6570\u636e\u63d2\u5165<\/h3>\n\n\n\n
      \u9996\u5148\u4e0b\u8f7d\u5982\u4e0b\u4f9d\u8d56\u5e93<\/p>\n\n\n\n
      pip install transformers torch timm<\/code><\/pre>\n\n\n\n
      \u7136\u540e\u901a\u8fc7\u5982\u4e0b\u4ee3\u7801\uff0c\u52a0\u8f7d\u5d4c\u5165\u6a21\u578b\u5e76\u8fdb\u884c\u5b9e\u4f8b\u5316\uff0c\u8fd9\u91cc\u5b9a\u4e49\u4e86\u8bcd\u5d4c\u5165\u548c\u56fe\u50cf\u5d4c\u5165\u7684\u65b9\u6cd5<\/p>\n\n\n\n
      from transformers import AutoModel\nfrom pymilvus import MilvusClient\nimport torch\nimport os\nos.environ['HF_ENDPOINT'] = 'https:\/\/hf-mirror.com'\n# Define Encoder class to handle text and image embedding generation\nclass Encoder:\n    def __init__(self):\n        # Initialize the model (AutoModel from transformers instead of SentenceTransformer)\n        # self.model = AutoModel.from_pretrained(model_name, trust_remote_code=True)\n        self.model = AutoModel.from_pretrained('.\/jinaai\/jina-clip-v2', trust_remote_code=True)\n\n    def encode_text(self, text: list[str]) -> list[float]:     # Generate embeddings for text only\n        with torch.no_grad():\n            text_emb = self.model.encode_text(text)\n        return text_emb\n    def encode_image(self, image_urls: list[str]) -> list[list[float]]:\n         # Generate embeddings for images only\n        with torch.no_grad():\n            image_emb = self.model.encode_image(image_urls)\n        return image_emb\n\nencoder = Encoder()\n<\/code><\/pre>\n\n\n\n
      \u6570\u636e\u793a\u4f8b\u5982\u4e0b\uff0c\u67e5\u770b\u5d4c\u5165\u7ef4\u5ea6\u662f1024<\/p>\n\n\n\n
      from modelscope import AutoModel\nmodel = AutoModel.from_pretrained('.\/jinaai\/jina-clip-v2', trust_remote_code=True)\nsentences = ['A blue cat', 'A red cat', 'A red cat generated by AI', 'A dog biting on a stick']\n\nimage_urls = [\n    'https:\/\/www.helloimg.com\/i\/2025\/04\/27\/680db7185383b.png',\n    'https:\/\/www.helloimg.com\/i\/2025\/04\/27\/680db74bacdb7.png'\n]\n\n# Generate embeddings for text and images\ntext_embeddings = encoder.encode_text(sentences)\nimage_embeddings = encoder.encode_image(image_urls)\n\n# Ensure consistent dimensions\ndim = len(image_embeddings[0])<\/code><\/pre>\n\n\n\n
      \u6267\u884c\u63d2\u5165\u6d41\u7a0b\u5982\u4e0b\uff0c\u6d41\u7a0b\u4e3a\u5148\u8fde\u63a5\u4e0aMilvus \u670d\u52a1\u5668\u518d\u8fde\u63a5multimodal_rag_2\uff0c\u63a5\u7740\u5bf9\u793a\u4f8b\u6570\u636e\u4f9d\u6b21\u8fdb\u884c\u6587\u672c\u5d4c\u5165\uff0c\u5f97\u5230\u5176\u5d4c\u5165\u5411\u91cf\u548c\u52a8\u6001\u5b57\u6bb5\uff0c\u518d\u901a\u8fc7insert\u65b9\u6cd5\u8fdb\u884c\u63d2\u5165\u6570\u636e<\/p>\n\n\n\n
      from pymilvus import utility, Collection\nfrom pymilvus import CollectionSchema, FieldSchema, DataType\nfrom pymilvus import connections\n\n# \u5c1d\u8bd5\u8fde\u63a5\u5230 Milvus \u670d\u52a1\u5668\ntry:\n    connections.connect(host='111.6.167.30', port=19530)\n    print(f\"\u6210\u529f\u8fde\u63a5\u5230 Milvus \u670d\u52a1\u5668\uff0c\u7aef\u53e3\u4e3a\uff1a{19530}\")\nexcept Exception as e:\n    print(f\"\u8fde\u63a5\u5931\u8d25\uff1a{e}\")\ncollection = Collection(\"multimodal_rag_2\")\ncollection.load()\n# \u63d2\u5165\u6570\u636e\ndata_to_insert = []\n\n# \u63d2\u5165\u6587\u672c\u5d4c\u5165\uff08\u5305\u542b\u52a8\u6001\u5b57\u6bb5text\uff09\nfor idx, txt_emb in enumerate(text_embeddings):\n    data_to_insert.append({\n        \"vector\": txt_emb,          # \u5fc5\u586b\u5b57\u6bb5\uff08Schema\u5b9a\u4e49\uff09\n        \"text\": sentences[idx]      # \u52a8\u6001\u5b57\u6bb5\uff08\u81ea\u52a8\u6dfb\u52a0\uff09\n    })\n\n# \u63d2\u5165\u56fe\u50cf\u5d4c\u5165\uff08\u5305\u542b\u52a8\u6001\u5b57\u6bb5image_url\uff09\nfor idx, img_emb in enumerate(image_embeddings):\n    data_to_insert.append({\n        \"vector\": img_emb,          # \u5fc5\u586b\u5b57\u6bb5\uff08Schema\u5b9a\u4e49\uff09\n        \"image_url\": image_urls[idx]  # \u52a8\u6001\u5b57\u6bb5\uff08\u81ea\u52a8\u6dfb\u52a0\uff09\n    })\n\ntry:\n    # \u6267\u884c\u63d2\u5165\u64cd\u4f5c\n    insert_result = collection.insert(data_to_insert)\n    print(f\"\u63d2\u5165\u6210\u529f\uff01\u5171\u63d2\u5165 {insert_result.insert_count} \u6761\u6570\u636e\")\n    \n    # \u663e\u5f0f\u52a0\u8f7d\u96c6\u5408\uff08\u786e\u4fdd\u53ef\u67e5\u8be2\uff09\n    collection.load()\n    print(f\"\u96c6\u5408\u5df2\u52a0\u8f7d\uff0c\u5f53\u524d\u5b9e\u4f53\u6570\u91cf: {collection.num_entities}\")\n    \nexcept MilvusException as e:\n    print(f\"\u63d2\u5165\u5931\u8d25: {e}\")\n    print(f\"\u9519\u8bef\u4ee3\u7801: {e.code}, \u9519\u8bef\u4fe1\u606f: {e.message}\")<\/code><\/pre>\n\n\n\n
      \"\"<\/div><\/figure>\n\n\n\n
      4\u3001\u5d4c\u5165\u540e\u67e5\u8be2<\/h3>\n\n\n\n
      \u6d4b\u8bd5\u811a\u672c\u5982\u4e0b\uff0c\u901a\u8fc7\u8fde\u63a5 Milvus \u5411\u91cf\u6570\u636e\u5e93\u5e76\u52a0\u8f7d\u4e00\u4e2a\u591a\u6a21\u6001\u6a21\u578b\uff08Jina-CLIP-V2\uff09\uff0c\u5206\u522b\u5bf9\u8f93\u5165\u7684\u56fe\u50cf URL \u548c\u6587\u672c\u67e5\u8be2\u751f\u6210\u5bf9\u5e94\u7684\u5d4c\u5165\u5411\u91cf\uff0c\u7136\u540e\u5728 Milvus \u4e2d\u6267\u884c\u57fa\u4e8e\u4f59\u5f26\u76f8\u4f3c\u5ea6\u7684\u8fd1\u90bb\u641c\u7d22\uff0c\u68c0\u7d22\u51fa\u4e0e\u67e5\u8be2\u5185\u5bb9\u6700\u76f8\u5173\u7684\u56fe\u6587\u6570\u636e\uff0c\u5e76\u5bf9\u641c\u7d22\u7ed3\u679c\u8fdb\u884c\u89e3\u6790\u548c\u5c55\u793a\uff0c\u4ece\u800c\u5b9e\u73b0\u4e00\u4e2a\u7ed3\u5408\u56fe\u50cf\u4e0e\u6587\u672c\u7684\u591a\u6a21\u6001\u68c0\u7d22\u7cfb\u7edf\u3002
      <\/p>\n\n\n\n
      from pymilvus import utility, Collection\nfrom pymilvus import CollectionSchema, FieldSchema, DataType\nfrom pymilvus import connections\n\n# \u5c1d\u8bd5\u8fde\u63a5\u5230 Milvus \u670d\u52a1\u5668\ntry:\n    connections.connect(host='111.6.167.30', port=19530)\n    print(f\"\u6210\u529f\u8fde\u63a5\u5230 Milvus \u670d\u52a1\u5668\uff0c\u7aef\u53e3\u4e3a\uff1a{19530}\")\nexcept Exception as e:\n    print(f\"\u8fde\u63a5\u5931\u8d25\uff1a{e}\")\n\nfrom transformers import AutoModel\nfrom pymilvus import MilvusClient\nimport torch\nimport os\nos.environ['HF_ENDPOINT'] = 'https:\/\/hf-mirror.com'\n# Define Encoder class to handle text and image embedding generation\nclass Encoder:\n    def __init__(self):\n        # Initialize the model (AutoModel from transformers instead of SentenceTransformer)\n        # self.model = AutoModel.from_pretrained(model_name, trust_remote_code=True)\n        self.model = AutoModel.from_pretrained('.\/jinaai\/jina-clip-v2', trust_remote_code=True)\n\n    def encode_text(self, text: list[str]) -> list[float]:     # Generate embeddings for text only\n        with torch.no_grad():\n            text_emb = self.model.encode_text(text)\n        return text_emb\n    def encode_image(self, image_urls: list[str]) -> list[list[float]]:\n         # Generate embeddings for images only\n        with torch.no_grad():\n            image_emb = self.model.encode_image(image_urls)\n        return image_emb\n\nencoder = Encoder()\n\ncollection = Collection(\"multimodal_rag_2\")\ncollection.load()\n\n# \u591a\u6a21\u6001\u641c\u7d22\uff08\u56fe\u50cf+\u6587\u672c\u67e5\u8be2\u7ec4\u5408\uff09\nquery_image_url = 'https:\/\/www.helloimg.com\/i\/2025\/04\/27\/680db74bacdb7.png'\nquery_text = \"Steap Audio\"\n\n# \u751f\u6210\u67e5\u8be2\u5d4c\u5165\uff08\u56fe\u50cf+\u6587\u672c\uff09\nquery_embedding_image = encoder.encode_image([query_image_url])[0]\nquery_embedding_text = encoder.encode_text([query_text])[0]\n\n# \u5b9a\u4e49\u641c\u7d22\u53c2\u6570\nsearch_params = {\n    \"metric_type\": \"COSINE\",\n    \"params\": {\"nprobe\": 10}  # \u4ec5\u5bf9IVF\u7c7b\u7d22\u5f15\u6709\u6548\n}\n\n# \u6267\u884c\u56fe\u50cf\u5d4c\u5165\u641c\u7d22\ntry:\n    image_search_result = collection.search(\n        data=[query_embedding_image],    # \u67e5\u8be2\u5411\u91cf\n        anns_field=\"vector\",             # \u5411\u91cf\u5b57\u6bb5\u540d\n        param=search_params,             # \u641c\u7d22\u53c2\u6570\n        limit=2,                         # \u8fd4\u56de\u7ed3\u679c\u6570\u91cf\n        output_fields=[\"text\", \"image_url\"],  # \u9700\u8981\u8fd4\u56de\u7684\u52a8\u6001\u5b57\u6bb5\n        expr=None                        # \u8fc7\u6ee4\u8868\u8fbe\u5f0f\uff08\u53ef\u9009\uff09\n    )\nexcept MilvusException as e:\n    print(f\"\u56fe\u50cf\u641c\u7d22\u5931\u8d25: {e}\")\n    image_search_result = []\n\n# \u6267\u884c\u6587\u672c\u5d4c\u5165\u641c\u7d22\ntry:\n    text_search_result = collection.search(\n        data=[query_embedding_text],\n        anns_field=\"vector\",\n        param=search_params,\n        limit=2,\n        output_fields=[\"text\", \"image_url\"]\n    )\nexcept MilvusException as e:\n    print(f\"\u6587\u672c\u641c\u7d22\u5931\u8d25: {e}\")\n    text_search_result = []\n\n# \u7ed3\u679c\u89e3\u6790\u51fd\u6570\ndef parse_search_result(results, source_type):\n    parsed = []\n    for hits in results:\n        for hit in hits:\n            item = {\n                \"score\": hit.score,\n                \"source\": source_type,\n                \"text\": hit.entity.get(\"text\"),\n                \"image_url\": hit.entity.get(\"image_url\")\n            }\n            parsed.append(item)\n    return parsed\n\n# \u89e3\u6790\u5e76\u5408\u5e76\u7ed3\u679c\nimage_results = parse_search_result(image_search_result, \"image_query\")\ntext_results = parse_search_result(text_search_result, \"text_query\")\n\n# \u6253\u5370\u7ed3\u679c\uff08\u793a\u4f8b\u683c\u5f0f\uff09\nprint(\"\u56fe\u50cf\u67e5\u8be2\u7ed3\u679c:\")\nfor res in image_results:\n    print(f\"\u6765\u6e90: {res['source']} | \u5f97\u5206: {res['score']:.4f}\")\n    if res['image_url']: print(f\"\u56fe\u7247\u5730\u5740: {res['image_url']}\")\n    if res['text']: print(f\"\u5173\u8054\u6587\u672c: {res['text']}\")\n    print(\"-\"*40)\n\nprint(\"\\n\u6587\u672c\u67e5\u8be2\u7ed3\u679c:\")\nfor res in text_results:\n    print(f\"\u6765\u6e90: {res['source']} | \u5f97\u5206: {res['score']:.4f}\")\n    if res['image_url']: print(f\"\u56fe\u7247\u5730\u5740: {res['image_url']}\")\n    if res['text']: print(f\"\u5173\u8054\u6587\u672c: {res['text']}\")\n    print(\"-\"*40)<\/code><\/pre>\n\n\n\n
      \"\"<\/div><\/figure>\n\n\n\n
      5\u3001\u670d\u52a1\u5316\u4f7f\u7528<\/h3>\n\n\n\n
      5.1 \u6570\u636e\u63d2\u5165\u670d\u52a1\u5316<\/h4>\n\n\n\n
      \u9996\u5148\u5b89\u88c5\u5982\u4e0b\u4f9d\u8d56\u5e93<\/p>\n\n\n\n
      !pip install langchain-community --user\n!pip install langchain --user\n!pip install docx2txt\n!pip install fastAPI\n!pip install pypdf<\/code><\/pre>\n\n\n\n
      \u63a5\u7740\u76f4\u63a5\u8fd0\u884c\u5982\u4e0b\u811a\u672c\uff0c\u5373\u53ef\u901a\u8fc7\u63a5\u53e3\u8fdb\u884c\u8c03\u7528<\/p>\n\n\n\n
      \u8be5\u4ee3\u7801\u5b9e\u73b0\u4e86\u4e00\u4e2a\u57fa\u4e8e FastAPI \u7684\u591a\u6a21\u6001 RAG \u670d\u52a1\uff0c\u901a\u8fc7\u6574\u5408 Milvus \u5411\u91cf\u6570\u636e\u5e93\u548c Jina-CLIP-V2 \u591a\u6a21\u6001\u6a21\u578b\uff0c\u63d0\u4f9b\u4e86\u652f\u6301\u6587\u672c\u3001\u56fe\u7247 URL \u548c\u6587\u6863\u6587\u4ef6\u7684\u4e0a\u4f20\u4e0e\u5411\u91cf\u5316\u63a5\u53e3\uff0c\u5e76\u5c06\u5904\u7406\u540e\u7684\u6587\u672c\u5185\u5bb9\u6216\u56fe\u7247\u5d4c\u5165\u5b58\u5165 Milvus \u96c6\u5408\u4e2d\uff0c\u7528\u4e8e\u540e\u7eed\u7684\u591a\u6a21\u6001\u68c0\u7d22\u5e94\u7528\u3002<\/p>\n\n\n\n
      from fastapi import FastAPI, File, UploadFile, HTTPException\nfrom fastapi.responses import JSONResponse\nfrom transformers import AutoModel\nimport torch\nfrom pymilvus import Collection, connections\nfrom langchain_community.document_loaders import (\n    PyPDFLoader,\n    TextLoader,\n    Docx2txtLoader,\n    CSVLoader,\n    UnstructuredMarkdownLoader\n)\nfrom langchain.text_splitter import RecursiveCharacterTextSplitter\nimport os\nimport shutil\nfrom tempfile import NamedTemporaryFile\nfrom fastapi import Body\nfrom pydantic import BaseModel\nfrom typing import List\n\napp = FastAPI()\n\n# Milvus\u914d\u7f6e\nMILVUS_HOST = '111.6.167.30'\nMILVUS_PORT = 19530\nCOLLECTION_NAME = \"multimodal_rag_2\"\n# \u56fe\u7247URL\u63a5\u53e3\nclass ImageUrls(BaseModel):\n    urls: List[str]\n# \u521d\u59cb\u5316Milvus\u548c\u6a21\u578b\n@app.on_event(\"startup\")\nasync def startup():\n    try:\n        # \u8fde\u63a5Milvus\n        connections.connect(host=MILVUS_HOST, port=MILVUS_PORT)\n        global collection\n        collection = Collection(COLLECTION_NAME)\n        collection.load()\n        \n        # \u52a0\u8f7d\u7f16\u7801\u6a21\u578b\n        global encoder\n        encoder = AutoModel.from_pretrained(\n            '.\/jinaai\/jina-clip-v2', \n            trust_remote_code=True\n        )\n    except Exception as e:\n        raise HTTPException(status_code=500, detail=f\"\u521d\u59cb\u5316\u5931\u8d25: {str(e)}\")\n\n# \u6587\u6863\u5904\u7406\u51fd\u6570\ndef process_document(file_path: str) -> list:\n    if file_path.endswith(\".pdf\"):\n        loader = PyPDFLoader(file_path)\n    elif file_path.endswith(\".txt\"):\n        loader = TextLoader(file_path)\n    elif file_path.endswith(\".docx\"):\n        loader = Docx2txtLoader(file_path)\n    elif file_path.endswith(\".csv\"):\n        loader = CSVLoader(file_path)\n    elif file_path.endswith(\".md\"):\n        loader = UnstructuredMarkdownLoader(file_path)\n    else:\n        raise ValueError(\"\u4e0d\u652f\u6301\u7684\u6587\u4ef6\u7c7b\u578b\")\n    \n    documents = loader.load()\n    \n    splitter = RecursiveCharacterTextSplitter(\n        chunk_size=400,\n        chunk_overlap=50,\n        length_function=len\n    )\n    return [doc.page_content for doc in splitter.split_documents(documents)]\n\n# \u6587\u4ef6\u4e0a\u4f20\u63a5\u53e3\n@app.post(\"\/upload\/file\/\")\nasync def upload_file(file: UploadFile = File(...)):\n    try:\n        # \u4fdd\u5b58\u4e34\u65f6\u6587\u4ef6\n        with NamedTemporaryFile(delete=False, suffix=file.filename) as tmp:\n            shutil.copyfileobj(file.file, tmp)\n            tmp_path = tmp.name\n        \n        # \u5904\u7406\u6587\u6863\n        texts = process_document(tmp_path)\n        embeddings = encoder.encode_text(texts).tolist()\n        \n        # \u63d2\u5165Milvus\n        data = [\n            {\"vector\": emb, \"text\": text} \n            for emb, text in zip(embeddings, texts)\n        ]\n        result = collection.insert(data)\n        collection.load()\n        \n        return JSONResponse({\n            \"status\": \"success\",\n            \"inserted\": result.insert_count,\n            \"entities\": collection.num_entities\n        })\n    \n    except Exception as e:\n        raise HTTPException(status_code=500, detail=str(e))\n    \n    finally:\n        # \u6e05\u7406\u4e34\u65f6\u6587\u4ef6\n        if os.path.exists(tmp_path):\n            os.remove(tmp_path)\n\n# \u4fee\u6539\u56fe\u7247\u63a5\u53e3\n@app.post(\"\/upload\/image-urls\/\")\nasync def upload_image_urls(data: ImageUrls):  # \u4f7f\u7528\u6a21\u578b\u63a5\u6536\n    try:\n        urls = data.urls  # \u83b7\u53d6\u5217\u8868\n        embeddings = encoder.encode_image(urls).tolist()\n        \n        data_to_insert = [\n            {\"vector\": emb, \"text\": url} \n            for emb, url in zip(embeddings, urls)\n        ]\n        result = collection.insert(data_to_insert)\n        collection.load()\n        \n        return JSONResponse({\n            \"status\": \"success\",\n            \"inserted\": result.insert_count,\n            \"entities\": collection.num_entities\n        })\n    except Exception as e:\n        raise HTTPException(status_code=500, detail=str(e))\n\n# \u6587\u672c\u63a5\u53e3\n@app.post(\"\/upload\/text\/\")\nasync def upload_text(data: dict = Body(...)):  # \u4fee\u6539\u6b64\u5904\n    try:\n        text = data.get(\"text\")  # \u4ece\u8bf7\u6c42\u4f53\u4e2d\u83b7\u53d6 text\n        if not text:\n            raise HTTPException(400, \"text \u5b57\u6bb5\u4e0d\u80fd\u4e3a\u7a7a\")\n        \n        splitter = RecursiveCharacterTextSplitter(\n            chunk_size=400,\n            chunk_overlap=50,\n            length_function=len\n        )\n        texts = splitter.split_text(text)\n        embeddings = encoder.encode_text(texts).tolist()\n        \n        data_to_insert = [\n            {\"vector\": emb, \"text\": txt} \n            for emb, txt in zip(embeddings, texts)\n        ]\n        result = collection.insert(data_to_insert)\n        collection.load()\n        \n        return JSONResponse({\n            \"status\": \"success\",\n            \"inserted\": result.insert_count,\n            \"entities\": collection.num_entities\n        })\n    except Exception as e:\n        raise HTTPException(status_code=500, detail=str(e))\n\nif __name__ == \"__main__\":\n    import uvicorn\n    uvicorn.run(app, host=\"0.0.0.0\", port=8000)<\/code><\/pre>\n\n\n\n
      \u8c03\u7528\u793a\u4f8b\u5982\u4e0b<\/p>\n\n\n\n
      import requests\nimport os\n\n# # \u6587\u4ef6\u4e0a\u4f20\u63a5\u53e3\uff08Python \u7248\uff09\n# def upload_file(file_path):\n#     url = \"http:\/\/localhost:8000\/upload\/file\/\"\n#     headers = {\"accept\": \"application\/json\"}\n    \n#     # \u786e\u4fdd\u6587\u4ef6\u5b58\u5728\n#     if not os.path.exists(file_path):\n#         raise FileNotFoundError(f\"\u6587\u4ef6 {file_path} \u4e0d\u5b58\u5728\")\n    \n#     # \u53d1\u9001\u8bf7\u6c42\n#     with open(file_path, \"rb\") as f:\n#         files = {\"file\": f}\n#         response = requests.post(\n#             url,\n#             headers=headers,\n#             files=files\n#         )\n    \n#     # \u5904\u7406\u54cd\u5e94\n#     if response.ok:\n#         print(\"\u4e0a\u4f20\u6210\u529f:\", response.json())\n#     else:\n#         print(\"\u4e0a\u4f20\u5931\u8d25:\", response.status_code, response.text)\n\n# # \u4f7f\u7528\u793a\u4f8b\n# upload_file(\"\u516c\u4f17\u53f7\u63a8\u65872.pdf\")\n\ndef upload_image_urls(urls):\n    url = \"http:\/\/localhost:8000\/upload\/image-urls\/\"\n    headers = {\"Content-Type\": \"application\/json\"}\n    data = {\"urls\": urls}  # urls \u9700\u662f\u5217\u8868\n    \n    response = requests.post(url, json=data, headers=headers)\n    \n    if response.ok:\n        print(f\"\u6210\u529f\u63d2\u5165 {response.json()['inserted']} \u5f20\u56fe\u7247\u5411\u91cf\")\n        print(f\"\u5f53\u524d\u603b\u5b9e\u4f53\u6570\uff1a{response.json()['entities']}\")\n    else:\n        print(f\"\u63d2\u5165\u5931\u8d25\uff1a{response.status_code}\")\n        print(response.text)\n\n# \u4f7f\u7528\u793a\u4f8b\nimage_urls = [\n    \"https:\/\/www.helloimg.com\/i\/2025\/04\/27\/680df2fb24e44.png\",\n    \"https:\/\/www.helloimg.com\/i\/2025\/04\/27\/680df32016474.png\"\n]\nupload_image_urls(image_urls)\n\n# def upload_text(text):\n#     url = \"http:\/\/localhost:8000\/upload\/text\/\"\n#     headers = {\"Content-Type\": \"application\/json\"}\n    \n#     # \u6784\u9020\u8bf7\u6c42\u4f53\uff08\u663e\u5f0f\u6307\u5b9a\u5b57\u6bb5\u540d\uff09\n#     data = {\"text\": text}  # \u5fc5\u987b\u4e0e\u63a5\u53e3\u5b9a\u4e49\u7684\u5b57\u6bb5\u540d\u4e00\u81f4\n    \n#     response = requests.post(url, json=data, headers=headers)\n    \n#     if response.ok:\n#         print(f\"\u6210\u529f\u63d2\u5165 {response.json()['inserted']} \u4e2a\u6587\u672c\u5757\")\n#     else:\n#         print(f\"\u63d2\u5165\u5931\u8d25\uff1a{response.status_code}\")\n#         print(\"\u9519\u8bef\u4fe1\u606f\uff1a\", response.text)\n\n# # \u4f7f\u7528\u793a\u4f8b\n# sample_text = \"\u8fd9\u662f\u9700\u8981\u63d2\u5165\u7684\u6587\u672c\u5185\u5bb9\"\n# upload_text(sample_text)  # \u786e\u4fdd\u4f20\u9012\u975e\u7a7a\u6587\u672c<\/code><\/pre>\n\n\n\n
      \"\"<\/div><\/figure>\n\n\n\n
      5.2 \u4ea4\u4e92\u667a\u80fd\u4f53\u6784\u5efa<\/h4>\n\n\n\n
      \u8be5\u4ee3\u7801\u5b9e\u73b0\u4e86\u4e00\u4e2a\u57fa\u4e8e\u591a\u6a21\u6001\u68c0\u7d22\u4e0e\u5927\u6a21\u578b\u751f\u6210\u7684\u667a\u80fd\u95ee\u7b54\u7cfb\u7edf\uff0c\u901a\u8fc7\u8fde\u63a5 Milvus \u5411\u91cf\u6570\u636e\u5e93\u548c\u4f7f\u7528 Jina-CLIP-V2 \u591a\u6a21\u6001\u7f16\u7801\u5668\uff0c\u7ed3\u5408 DashScope \u7684 Qwen-VL-Plus \u591a\u6a21\u6001\u5927\u6a21\u578b\uff0c\u5b8c\u6210\u5bf9\u7528\u6237\u6587\u672c\u6216\u56fe\u50cf\u67e5\u8be2\u7684\u8bed\u4e49\u641c\u7d22\u3001\u76f8\u5173\u56fe\u6587\u5185\u5bb9\u53ec\u56de\u3001\u56fe\u50cf\u5185\u5bb9\u89e3\u6790\uff0c\u5e76\u6700\u7ec8\u7531\u5927\u6a21\u578b\u751f\u6210\u7ed3\u6784\u5316\u56de\u7b54\uff0c\u5b9e\u73b0\u4e86\u4ece\u67e5\u8be2\u5230\u7ed3\u679c\u751f\u6210\u7684\u7aef\u5230\u7aef\u591a\u6a21\u6001 RAG\uff08Retrieval-Augmented Generation\uff09\u6d41\u7a0b\u3002<\/p>\n\n\n\n
      \"\"<\/div><\/figure>\n\n\n\n
      import json\nfrom pymilvus import Collection, MilvusException\nfrom pymilvus import connections\nfrom transformers import AutoModel\nimport torch\nimport os\nfrom openai import OpenAI\nimport requests\nfrom urllib.parse import urlparse\n\n# \u8fde\u63a5Milvus\u670d\u52a1\u5668\ntry:\n    connections.connect(host='111.6.167.30', port=19530)\n    print(f\"\u6210\u529f\u8fde\u63a5\u5230 Milvus \u670d\u52a1\u5668\uff0c\u7aef\u53e3\u4e3a\uff1a{19530}\")\nexcept Exception as e:\n    print(f\"\u8fde\u63a5\u5931\u8d25\uff1a{e}\")\nos.environ['HF_ENDPOINT'] = 'https:\/\/hf-mirror.com'\n\nclass Encoder:\n    def __init__(self):\n        self.model = AutoModel.from_pretrained(\n            '.\/jinaai\/jina-clip-v2', \n            trust_remote_code=True\n        )\n\n    def encode_text(self, texts: list[str]) -> list[float]:\n        with torch.no_grad():\n            return self.model.encode_text(texts).tolist()\n\n    def encode_image(self, image_urls: list[str]) -> list[float]:\n        with torch.no_grad():\n            return self.model.encode_image(image_urls).tolist()\n\nencoder = Encoder()\ncollection = Collection(\"multimodal_rag_2\")\ncollection.load()\n\n# \u521d\u59cb\u5316OpenAI\u5ba2\u6237\u7aef\nclient = OpenAI(\n    api_key=\"sk-7cc2531e8b3c425e803eed8296b5072d\",\n    base_url=\"https:\/\/dashscope.aliyuncs.com\/compatible-mode\/v1\",\n)\n\ndef is_valid_url(url):\n    try:\n        headers = {'User-Agent': 'Mozilla\/5.0'}  # \u6dfb\u52a0UA\u907f\u514d\u88ab\u62e6\u622a\n        response = requests.head(url, headers=headers, timeout=5, allow_redirects=True)\n        content_type = response.headers.get('Content-Type', '')\n        return response.status_code == 200 and content_type.startswith('image\/')\n    except requests.RequestException:\n        return False\n\nclass FinancialAssistant:\n    def __init__(self, encoder, collection):\n        self.encoder = encoder\n        self.collection = collection\n    # \u591a\u6a21\u6001\u63d0\u53d6\u56fe\u7247\u4fe1\u606f\n    def extract_image_info(self, image_urls):\n        system_image_prompt = \"\"\"\n        \u4f60\u662f\u4e00\u4e2a\u56fe\u7247\u89e3\u6790\u52a9\u624b\uff0c\u7528\u6237\u4f1a\u5411\u4f60\u63d0\u4f9b\u4e00\u5f20\u6216\u8005\u591a\u5f20\u56fe\u7247\uff0c\u8bf7\u6839\u636e\u56fe\u7247\u5185\u5bb9\u8fdb\u884c\uff0c\u5c3d\u53ef\u80fd\u7684\u8fdb\u884c\u63cf\u8ff0\n        \u4f60\u9700\u8981\u6839\u636e\u56fe\u7247\u5185\u5bb9\u751f\u6210\u7b80\u6d01\u7684\u63cf\u8ff0\uff0c\u63cf\u8ff0\u4e2d\u9700\u8981\u5305\u542b\u56fe\u7247\u7684\u4e3b\u8981\u5143\u7d20\u548c\u7279\u5f81\u3002\n        \u8fd4\u56dejson\u683c\u5f0f\uff0c\u5305\u542b\u4ee5\u4e0b\u5b57\u6bb5\uff1a\n        [\n        {{\n            \"text\": \u56fe\u7247\u7684\u63cf\u8ff0\u4fe1\u606f,\n            \"images\": \u56fe\u50cfURL\n        }},\n        \/\/ \u5176\u4ed6\u6761\u76ee...\n        ]\n        \"\"\"\n        user_content = []\n        valid_urls = []\n        for image_url in image_urls:\n            if is_valid_url(image_url):\n                valid_urls.append(image_url)\n                # \u4fee\u6b63\uff1a\u5206\u4e24\u6b21\u6dfb\u52a0\u5143\u7d20\n                user_content.append(\n                    {\"type\": \"text\", \"text\": f\"\u8fd9\u5f20\u56fe\u7247url\u662f{image_url}\"}\n                )\n                user_content.append(\n                    {\"type\": \"image_url\", \"image_url\": {\"url\": image_url}}\n                )\n                print(f\"\u6709\u6548\u7684\u56fe\u50cfURL: {image_url}\")\n            else:\n                print(f\"\u65e0\u6548\u7684\u56fe\u50cfURL: {image_url}\")\n        \n        if not valid_urls:\n            return []\n        \n        messages = [\n            {\"role\": \"system\", \"content\": system_image_prompt},\n            {\"role\": \"user\", \"content\": user_content}\n        ]\n        \n        try:\n            completion = client.chat.completions.create(\n                model=\"qwen-vl-plus\", \n                messages=messages\n            )\n            # print(completion)\n            return completion.choices[0].message.content\n        except Exception as e:\n            print(f\"\u56fe\u7247\u89e3\u6790\u5931\u8d25: {e}\")\n            return []\n\n    def search_and_respond(self, query_text=None, query_image_url=None):\n        results = []\n        images_urls = []\n        text_images_urls=[]\n        valid_urls = set()\n        image_results = []\n        text_images_results = []\n\n        # \u56fe\u50cf\u67e5\u8be2\n        if query_image_url and is_valid_url(query_image_url):\n            image_embedding = self.encoder.encode_image([query_image_url])[0]\n            image_params = {\n                \"metric_type\": \"COSINE\",\n                \"params\": {\"nprobe\": 10},\n                \"expr\": \"text like 'http%'\"  # \u8fc7\u6ee4\u975eURL\u5185\u5bb9\n            }\n            try:\n                image_res = self.collection.search(\n                    data=[image_embedding],\n                    anns_field=\"vector\",\n                    param=image_params,\n                    limit=3,\n                    output_fields=[\"text\"]\n                )\n                images_urls.extend([\n                    {\"type\": \"image\", \"content\": hit.entity.get(\"text\")}\n                    for hits in image_res for hit in hits\n                ])\n            except MilvusException as e:\n                print(f\"\u56fe\u50cf\u641c\u7d22\u5931\u8d25: {e}\")\n\n        # \u6587\u672c\u67e5\u8be2(\u4e5f\u53ef\u67e5\u8be2\u51fa\u56fe\u50cf)\n        if query_text:\n            text_embedding = self.encoder.encode_text([query_text])[0]\n            text_params = {\n                \"metric_type\": \"COSINE\",\n                \"params\": {\"nprobe\": 10}\n            }\n            try:\n                text_res = self.collection.search(\n                    data=[text_embedding],\n                    anns_field=\"vector\",\n                    param=text_params,\n                    limit=5,\n                    output_fields=[\"text\"]\n                )\n                text_images_urls.extend([\n                    {\"type\": \"text\", \"content\": hit.entity.get(\"text\")}\n                    if not hit.entity.get(\"text\").startswith(('http', 'https')) \n                    else {\"type\": \"image\", \"content\": hit.entity.get(\"text\")}\n                    for hits in text_res for hit in hits\n                ])\n            except MilvusException as e:\n                print(f\"\u6587\u672c\u641c\u7d22\u5931\u8d25: {e}\")\n        # \u5904\u7406\u56fe\u50cf\u641c\u7d22\u7ed3\u679c\n        if images_urls:\n            image_results = self.extract_image_info(images_urls)\n        print(\"\u56fe\u641c\u56fe\u7247\u68c0\u7d22\u5185\u5bb9\u63d0\u53d6----------------------------------\")\n        print(image_results)\n        print(\"----------------------------------------------------\")\n        # \u5904\u7406\u6587\u672c\u79cd\u56fe\u50cf\u641c\u7d22\u7ed3\u679c\u5b57\u6bb5\n        if text_images_urls:\n            for item in text_images_urls:\n                if item[\"type\"] == \"image\":\n                    text_images_results.append(self.extract_image_info([item[\"content\"]]))\n                else:\n                    results.append({\n                        \"type\": \"text\",\n                        \"content\": item[\"content\"]\n                    })\n        print(\"\u6587\u641c\u56fe\u7247\u68c0\u7d22\u5185\u5bb9\u63d0\u53d6----------------------------------\")\n        print(text_images_results)\n        print(\"----------------------------------------------------\")\n\n        # \u6784\u9020API\u8bf7\u6c42\uff08\u65b0\u589e\u7cfb\u7edf\u8bbe\u5b9a\uff09\n        rag_content = []\n        # \u6dfb\u52a0\u641c\u7d22\u7ed3\u679c\u5230\u7528\u6237\u5185\u5bb9\n        for item in results:\n            if item[\"type\"] == \"image\":\n                if is_valid_url(item[\"content\"]) and item[\"content\"] not in valid_urls:\n                    rag_content.append({\n                        \"type\": \"image_url\",\n                        \"image_url\": {\"url\": item[\"content\"]}\n                    })\n                    valid_urls.add(item[\"content\"])\n            else:\n                rag_content.append({\n                    \"type\": \"text\",\n                    \"text\": item[\"content\"]\n                })\n        rag_content.extend(text_images_results)\n        rag_content.extend(image_results)\n        print(f\"RAG\u5185\u5bb9: {rag_content}\")\n        print(\"----------------------------------------------------\")\n\n        system_prompt = \"\"\"\n        \u4f60\u662f\u4e00\u4e2a\u667a\u80fd\u52a9\u624b\uff0c\u7528\u6237\u4f1a\u5411\u4f60\u63d0\u95ee\u3002\u8bf7\u6839\u636e\u7528\u6237\u7684\u95ee\u9898\u548c\u63d0\u4f9b\u7684\u53c2\u8003\u4fe1\u606f\u8fdb\u884c\u56de\u7b54\u3002\n        \u4f60\u9700\u8981\u8fc7\u6ee4\u4e0e\u95ee\u9898\u65e0\u5173\u7684\u4fe1\u606f\uff0c\u9488\u5bf9\u7528\u6237\u7684\u6838\u5fc3\u95ee\u9898\u63d0\u4f9b\u7b80\u6d01\u7684\u56de\u7b54\u3002\n        \u8bf7\u6ce8\u610f\uff0c\u7528\u6237\u53ef\u80fd\u4f1a\u63d0\u4f9b\u56fe\u50cf\u6216\u6587\u672c\u4f5c\u4e3a\u53c2\u8003\u4fe1\u606f,\u5176\u4e2d\u56fe\u50cf\u90e8\u5206\u6bcf\u4e2a\u5747\u4f1a\u6709URL\u53ca\u5176\u76f8\u5173\u63cf\u8ff0\u3002\n        \u4f60\u9700\u8981\u6839\u636e\u8fd9\u4e9b\u4fe1\u606f\u6765\u751f\u6210\u56de\u7b54\uff0c\u6ce8\u610f\u89c2\u5bdf\u5982\u679c\u53c2\u8003\u4fe1\u606f\u5185\u591a\u4e2a\u56fe\u7247url\u5b8c\u5168\u4e00\u6837\uff0c\u90a3\u4ec5\u56de\u590d\u5176\u4e2d\u4e00\u4e2a\u5373\u53ef\u3002\n        \u56de\u7b54\u65b9\u5f0f\u9700\u8981\u4fdd\u8bc1\u8db3\u591f\u7b80\u6d01\u548c\u51c6\u786e\u3002\n        \u4f60\u53ef\u4ee5\u4f7f\u7528\u63d0\u4f9b\u7684\u53c2\u8003\u4fe1\u606f\u6765\u589e\u5f3a\u4f60\u7684\u56de\u7b54\uff0c\u4f46\u8bf7\u786e\u4fdd\u56de\u7b54\u7684\u6838\u5fc3\u662f\u7528\u6237\u7684\u95ee\u9898\u3002\n        \u56de\u7b54\u683c\u5f0f\u4e3ajson\uff0c\u5305\u542b\u4ee5\u4e0b\u5b57\u6bb5\uff1a\n        {\n            \"text\":\u9488\u5bf9\u7528\u6237\u95ee\u9898\u7684\u56de\u7b54\uff0c\u4e0e\u7528\u6237\u4ea4\u4e92\u7684\u6587\u672c\u4fe1\u606f,\n            \"images\": [\u56fe\u50cfURL\u5217\u8868],\n        }\n        \"\"\"\n\n        # \u6784\u9020\u7528\u6237\u6d88\u606f\n        user_content = f\"\"\"\"\n        \u4f60\u7684\u8bbe\u5b9a\u662f\uff1a{system_prompt}\\n\n        \u5f53\u524d\u7528\u6237\u95ee\u9898\u4e3a:{query_text}\\n\n        \u53c2\u8003\u4fe1\u606f\u4e3a:{str(rag_content)}\n        \"\"\"\n\n        messages = [\n            {\"role\": \"system\", \"content\": system_prompt},\n            {\"role\": \"user\", \"content\": user_content}\n        ]\n\n        # \u8c03\u7528OpenAI API\n        try:\n            print(f\"\u8bf7\u6c42\u6d88\u606f: {messages}\")\n            print(\"----------------------------------------------------\")\n            response = client.chat.completions.create(\n                model=\"qwen-vl-plus\",\n                messages=messages\n            )\n            return response\n        except Exception as e:\n            print(f\"API\u8c03\u7528\u5931\u8d25: {e}\")\n            return None\n\n# \u4f7f\u7528\u793a\u4f8b\nassistant = FinancialAssistant(encoder, collection)\nresponse = assistant.search_and_respond(\n    query_text=\"\u84dd\u8272\u5934\u53d1\u5c11\u5973\u6709\u5417\uff1f\",\n)\n# print(response.choices[0].message.content)\nresult = response.choices[0].message.content\njson_result=json.loads(result)\nprint(json_result.get(\"text\"))\nprint(json_result.get(\"images\"))<\/code><\/pre>\n\n\n\n
      \"\"<\/div><\/figure>\n","protected":false},"excerpt":{"rendered":"
      0\u3001\u7cfb\u7edf\u76ee\u6807 \u672c\u7cfb\u7edf\u9884\u671f\u76ee\u6807\u662f\u6784\u5efa\u4e00\u4e2a\u591a\u6a21\u6001\u7684RAG\u7cfb\u7edf\uff0c\u6b64\u7cfb\u7edf\u53ef\u5b9e\u73b0\u591a\u79cd\u6a21\u6001\u7684\u6570\u636e\u5bfc\u5165\uff0c\u5305\u62ec\u5982PDF\u3001Wor […]<\/p>\n","protected":false},"author":1,"featured_media":231,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[3],"tags":[],"class_list":["post-230","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-rag"],"yoast_head":"\n\u4e00\u6587\u5feb\u901f\u5b66\u4f1a\u57fa\u4e8eMilvus\u6784\u5efa\u591a\u6a21\u6001RAG\u95ee\u7b54\u7cfb\u7edf - \u4eca\u5929\u5f00\u59cb\u5b66AI<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/\u4e00\u6587\u6559\u4f60\u5b66\u4f1a\u57fa\u4e8emilvus\u6784\u5efa\u591a\u6a21\u6001rag\u95ee\u7b54\u7cfb\u7edf\/\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u4e00\u6587\u5feb\u901f\u5b66\u4f1a\u57fa\u4e8eMilvus\u6784\u5efa\u591a\u6a21\u6001RAG\u95ee\u7b54\u7cfb\u7edf - \u4eca\u5929\u5f00\u59cb\u5b66AI\" \/>\n<meta property=\"og:description\" content=\"0\u3001\u7cfb\u7edf\u76ee\u6807 \u672c\u7cfb\u7edf\u9884\u671f\u76ee\u6807\u662f\u6784\u5efa\u4e00\u4e2a\u591a\u6a21\u6001\u7684RAG\u7cfb\u7edf\uff0c\u6b64\u7cfb\u7edf\u53ef\u5b9e\u73b0\u591a\u79cd\u6a21\u6001\u7684\u6570\u636e\u5bfc\u5165\uff0c\u5305\u62ec\u5982PDF\u3001Wor […]\" \/>\n<meta property=\"og:url\" content=\"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/\u4e00\u6587\u6559\u4f60\u5b66\u4f1a\u57fa\u4e8emilvus\u6784\u5efa\u591a\u6a21\u6001rag\u95ee\u7b54\u7cfb\u7edf\/\" \/>\n<meta property=\"og:site_name\" content=\"\u4eca\u5929\u5f00\u59cb\u5b66AI\" \/>\n<meta property=\"article:published_time\" content=\"2025-05-04T15:11:03+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2025-05-06T01:43:03+00:00\" \/>\n<meta property=\"og:image\" content=\"http:\/\/knoka.vip\/wp-content\/uploads\/2025\/05\/R-C.png\" \/>\n\t<meta property=\"og:image:width\" content=\"900\" \/>\n\t<meta property=\"og:image:height\" content=\"506\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/png\" \/>\n<meta name=\"author\" content=\"Knoka\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"Knoka\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"3 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/#article\",\"isPartOf\":{\"@id\":\"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/\"},\"author\":{\"name\":\"Knoka\",\"@id\":\"https:\/\/knoka.vip\/#\/schema\/person\/59bf442e4e4c21f6969b984b8c5bfc84\"},\"headline\":\"\u4e00\u6587\u5feb\u901f\u5b66\u4f1a\u57fa\u4e8eMilvus\u6784\u5efa\u591a\u6a21\u6001RAG\u95ee\u7b54\u7cfb\u7edf\",\"datePublished\":\"2025-05-04T15:11:03+00:00\",\"dateModified\":\"2025-05-06T01:43:03+00:00\",\"mainEntityOfPage\":{\"@id\":\"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/\"},\"wordCount\":88,\"commentCount\":1,\"publisher\":{\"@id\":\"https:\/\/knoka.vip\/#\/schema\/person\/59bf442e4e4c21f6969b984b8c5bfc84\"},\"image\":{\"@id\":\"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/#primaryimage\"},\"thumbnailUrl\":\"http:\/\/knoka.vip\/wp-content\/uploads\/2025\/05\/R-C.png\",\"articleSection\":[\"\u5411\u91cf\u6570\u636e\u5e93\"],\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/\",\"url\":\"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/\",\"name\":\"\u4e00\u6587\u5feb\u901f\u5b66\u4f1a\u57fa\u4e8eMilvus\u6784\u5efa\u591a\u6a21\u6001RAG\u95ee\u7b54\u7cfb\u7edf - \u4eca\u5929\u5f00\u59cb\u5b66AI\",\"isPartOf\":{\"@id\":\"https:\/\/knoka.vip\/#website\"},\"primaryImageOfPage\":{\"@id\":\"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/#primaryimage\"},\"image\":{\"@id\":\"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/#primaryimage\"},\"thumbnailUrl\":\"http:\/\/knoka.vip\/wp-content\/uploads\/2025\/05\/R-C.png\",\"datePublished\":\"2025-05-04T15:11:03+00:00\",\"dateModified\":\"2025-05-06T01:43:03+00:00\",\"breadcrumb\":{\"@id\":\"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/#primaryimage\",\"url\":\"http:\/\/knoka.vip\/wp-content\/uploads\/2025\/05\/R-C.png\",\"contentUrl\":\"http:\/\/knoka.vip\/wp-content\/uploads\/2025\/05\/R-C.png\",\"width\":900,\"height\":506},{\"@type\":\"BreadcrumbList\",\"@id\":\"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/knoka.vip\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u4e00\u6587\u5feb\u901f\u5b66\u4f1a\u57fa\u4e8eMilvus\u6784\u5efa\u591a\u6a21\u6001RAG\u95ee\u7b54\u7cfb\u7edf\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/knoka.vip\/#website\",\"url\":\"https:\/\/knoka.vip\/\",\"name\":\"\u4eca\u5929\u5f00\u59cb\u5b66AI\",\"description\":\"\u77e5\u8bc6\u5206\u4eab\",\"publisher\":{\"@id\":\"https:\/\/knoka.vip\/#\/schema\/person\/59bf442e4e4c21f6969b984b8c5bfc84\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/knoka.vip\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":[\"Person\",\"Organization\"],\"@id\":\"https:\/\/knoka.vip\/#\/schema\/person\/59bf442e4e4c21f6969b984b8c5bfc84\",\"name\":\"Knoka\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/knoka.vip\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.pho.ink\/avatar\/a81dd1238e551ec6c470995487946913852d59837b557d5166ae60169466da05?s=96&r=g&d=404\",\"contentUrl\":\"https:\/\/gravatar.pho.ink\/avatar\/a81dd1238e551ec6c470995487946913852d59837b557d5166ae60169466da05?s=96&r=g&d=404\",\"caption\":\"Knoka\"},\"logo\":{\"@id\":\"https:\/\/knoka.vip\/#\/schema\/person\/image\/\"},\"sameAs\":[\"http:\/\/118.178.172.39\"],\"url\":\"http:\/\/knoka.vip\/index.php\/author\/admin\/\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u4e00\u6587\u5feb\u901f\u5b66\u4f1a\u57fa\u4e8eMilvus\u6784\u5efa\u591a\u6a21\u6001RAG\u95ee\u7b54\u7cfb\u7edf - \u4eca\u5929\u5f00\u59cb\u5b66AI","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/\u4e00\u6587\u6559\u4f60\u5b66\u4f1a\u57fa\u4e8emilvus\u6784\u5efa\u591a\u6a21\u6001rag\u95ee\u7b54\u7cfb\u7edf\/","og_locale":"zh_CN","og_type":"article","og_title":"\u4e00\u6587\u5feb\u901f\u5b66\u4f1a\u57fa\u4e8eMilvus\u6784\u5efa\u591a\u6a21\u6001RAG\u95ee\u7b54\u7cfb\u7edf - \u4eca\u5929\u5f00\u59cb\u5b66AI","og_description":"0\u3001\u7cfb\u7edf\u76ee\u6807 \u672c\u7cfb\u7edf\u9884\u671f\u76ee\u6807\u662f\u6784\u5efa\u4e00\u4e2a\u591a\u6a21\u6001\u7684RAG\u7cfb\u7edf\uff0c\u6b64\u7cfb\u7edf\u53ef\u5b9e\u73b0\u591a\u79cd\u6a21\u6001\u7684\u6570\u636e\u5bfc\u5165\uff0c\u5305\u62ec\u5982PDF\u3001Wor […]","og_url":"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/\u4e00\u6587\u6559\u4f60\u5b66\u4f1a\u57fa\u4e8emilvus\u6784\u5efa\u591a\u6a21\u6001rag\u95ee\u7b54\u7cfb\u7edf\/","og_site_name":"\u4eca\u5929\u5f00\u59cb\u5b66AI","article_published_time":"2025-05-04T15:11:03+00:00","article_modified_time":"2025-05-06T01:43:03+00:00","og_image":[{"width":900,"height":506,"url":"http:\/\/knoka.vip\/wp-content\/uploads\/2025\/05\/R-C.png","type":"image\/png"}],"author":"Knoka","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"Knoka","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"3 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/#article","isPartOf":{"@id":"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/"},"author":{"name":"Knoka","@id":"https:\/\/knoka.vip\/#\/schema\/person\/59bf442e4e4c21f6969b984b8c5bfc84"},"headline":"\u4e00\u6587\u5feb\u901f\u5b66\u4f1a\u57fa\u4e8eMilvus\u6784\u5efa\u591a\u6a21\u6001RAG\u95ee\u7b54\u7cfb\u7edf","datePublished":"2025-05-04T15:11:03+00:00","dateModified":"2025-05-06T01:43:03+00:00","mainEntityOfPage":{"@id":"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/"},"wordCount":88,"commentCount":1,"publisher":{"@id":"https:\/\/knoka.vip\/#\/schema\/person\/59bf442e4e4c21f6969b984b8c5bfc84"},"image":{"@id":"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/#primaryimage"},"thumbnailUrl":"http:\/\/knoka.vip\/wp-content\/uploads\/2025\/05\/R-C.png","articleSection":["\u5411\u91cf\u6570\u636e\u5e93"],"inLanguage":"zh-Hans","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/#respond"]}]},{"@type":"WebPage","@id":"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/","url":"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/","name":"\u4e00\u6587\u5feb\u901f\u5b66\u4f1a\u57fa\u4e8eMilvus\u6784\u5efa\u591a\u6a21\u6001RAG\u95ee\u7b54\u7cfb\u7edf - \u4eca\u5929\u5f00\u59cb\u5b66AI","isPartOf":{"@id":"https:\/\/knoka.vip\/#website"},"primaryImageOfPage":{"@id":"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/#primaryimage"},"image":{"@id":"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/#primaryimage"},"thumbnailUrl":"http:\/\/knoka.vip\/wp-content\/uploads\/2025\/05\/R-C.png","datePublished":"2025-05-04T15:11:03+00:00","dateModified":"2025-05-06T01:43:03+00:00","breadcrumb":{"@id":"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/#primaryimage","url":"http:\/\/knoka.vip\/wp-content\/uploads\/2025\/05\/R-C.png","contentUrl":"http:\/\/knoka.vip\/wp-content\/uploads\/2025\/05\/R-C.png","width":900,"height":506},{"@type":"BreadcrumbList","@id":"http:\/\/knoka.vip\/index.php\/2025\/05\/04\/%e4%b8%80%e6%96%87%e6%95%99%e4%bd%a0%e5%ad%a6%e4%bc%9a%e5%9f%ba%e4%ba%8emilvus%e6%9e%84%e5%bb%ba%e5%a4%9a%e6%a8%a1%e6%80%81rag%e9%97%ae%e7%ad%94%e7%b3%bb%e7%bb%9f\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/knoka.vip\/"},{"@type":"ListItem","position":2,"name":"\u4e00\u6587\u5feb\u901f\u5b66\u4f1a\u57fa\u4e8eMilvus\u6784\u5efa\u591a\u6a21\u6001RAG\u95ee\u7b54\u7cfb\u7edf"}]},{"@type":"WebSite","@id":"https:\/\/knoka.vip\/#website","url":"https:\/\/knoka.vip\/","name":"\u4eca\u5929\u5f00\u59cb\u5b66AI","description":"\u77e5\u8bc6\u5206\u4eab","publisher":{"@id":"https:\/\/knoka.vip\/#\/schema\/person\/59bf442e4e4c21f6969b984b8c5bfc84"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/knoka.vip\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":["Person","Organization"],"@id":"https:\/\/knoka.vip\/#\/schema\/person\/59bf442e4e4c21f6969b984b8c5bfc84","name":"Knoka","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/knoka.vip\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.pho.ink\/avatar\/a81dd1238e551ec6c470995487946913852d59837b557d5166ae60169466da05?s=96&r=g&d=404","contentUrl":"https:\/\/gravatar.pho.ink\/avatar\/a81dd1238e551ec6c470995487946913852d59837b557d5166ae60169466da05?s=96&r=g&d=404","caption":"Knoka"},"logo":{"@id":"https:\/\/knoka.vip\/#\/schema\/person\/image\/"},"sameAs":["http:\/\/118.178.172.39"],"url":"http:\/\/knoka.vip\/index.php\/author\/admin\/"}]}},"_links":{"self":[{"href":"http:\/\/knoka.vip\/index.php\/wp-json\/wp\/v2\/posts\/230","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/knoka.vip\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/knoka.vip\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/knoka.vip\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/knoka.vip\/index.php\/wp-json\/wp\/v2\/comments?post=230"}],"version-history":[{"count":0,"href":"http:\/\/knoka.vip\/index.php\/wp-json\/wp\/v2\/posts\/230\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/knoka.vip\/index.php\/wp-json\/wp\/v2\/media\/231"}],"wp:attachment":[{"href":"http:\/\/knoka.vip\/index.php\/wp-json\/wp\/v2\/media?parent=230"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/knoka.vip\/index.php\/wp-json\/wp\/v2\/categories?post=230"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/knoka.vip\/index.php\/wp-json\/wp\/v2\/tags?post=230"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}