{"id":6918,"date":"2026-05-08T07:14:32","date_gmt":"2026-05-07T23:14:32","guid":{"rendered":"https:\/\/www.babel.cc\/blog\/?p=6918"},"modified":"2026-05-08T07:14:32","modified_gmt":"2026-05-07T23:14:32","slug":"%e4%bc%81%e4%b8%9a%e4%ba%91%e7%9b%98%e4%b8%8eai%e7%bb%93%e5%90%88%ef%bc%9a%e6%99%ba%e8%83%bd%e6%a0%87%e7%ad%be%e3%80%81%e8%87%aa%e5%8a%a8%e5%88%86%e7%b1%bb%e3%80%81%e5%86%85%e5%ae%b9%e6%91%98%e8%a6%81","status":"publish","type":"post","link":"https:\/\/www.babel.cc\/blog\/?p=6918","title":{"rendered":"\u4f01\u4e1a\u4e91\u76d8\u4e0eAI\u7ed3\u5408\uff1a\u667a\u80fd\u6807\u7b7e\u3001\u81ea\u52a8\u5206\u7c7b\u3001\u5185\u5bb9\u6458\u8981\u4e0e\u667a\u80fd\u641c\u7d22\u5b9e\u6218"},"content":{"rendered":"<h2>\u524d\u8a00<\/h2>\n<p>\u4f01\u4e1a\u4e91\u76d8\u5728\u7ecf\u5386\u4e86&#8221;\u5b58\u50a8\u65f6\u4ee3&#8221;\u548c&#8221;\u534f\u4f5c\u65f6\u4ee3&#8221;\u540e\uff0c\u6b63\u5728\u8fdb\u5165\u7b2c\u4e09\u4e2a\u9636\u6bb5\uff1a<strong>AI\u9a71\u52a8\u65f6\u4ee3<\/strong>\u3002<\/p>\n<p>\u4f20\u7edf\u7684\u4f01\u4e1a\u4e91\u76d8\uff0c\u7528\u6237\u9700\u8981\u624b\u52a8\u7ba1\u7406\u6587\u4ef6\uff1a\u8d77\u540d\u5b57\u3001\u5efa\u6587\u4ef6\u5939\u3001\u6253\u6807\u7b7e\u3002\u6587\u4ef6\u4e00\u591a\uff0c\u627e\u6587\u4ef6\u5c31\u6210\u4e86\u8d1f\u62c5\u2014\u2014\u8bb0\u4f4f\u6587\u4ef6\u540d\u3001\u8bb0\u4f4f\u76ee\u5f55\u7ed3\u6784\u3001\u8bb0\u4f4f\u6807\u7b7e\u4f53\u7cfb\uff0c\u6bcf\u4e00\u6b65\u90fd\u5728\u6d88\u8017\u8ba4\u77e5\u8d44\u6e90\u3002<\/p>\n<p>AI\u6539\u53d8\u4e86\u4e00\u5207\uff1a\u4f60\u53ea\u9700\u8981\u63cf\u8ff0\u4f60\u8981\u627e\u4ec0\u4e48\uff0cAI\u5e2e\u4f60\u627e\u5230\u3002\u4f60\u53ea\u9700\u8981\u8bf4&#8221;\u6574\u7406\u4e00\u4e0b\u8fd9\u4e2a\u9879\u76ee\u7684\u6587\u6863&#8221;\uff0cAI\u5e2e\u4f60\u5206\u7c7b\u6253\u6807\u7b7e\u3002\u4f60\u53ea\u9700\u8981\u70b9\u4e00\u4e0b&#8221;\u751f\u6210\u6458\u8981&#8221;\uff0cAI\u628a\u51e0\u5341\u5206\u949f\u7684\u4f1a\u8bae\u5f55\u97f3\u53d8\u6210\u4e00\u6bb5\u8bdd\u3002<\/p>\n<p>\u672c\u6587\u4ece\u6280\u672f\u67b6\u6784\u5c42\u9762\uff0c\u7cfb\u7edf\u5206\u6790\u4f01\u4e1a\u4e91\u76d8\u4e0eAI\u7ed3\u5408\u7684\u6838\u5fc3\u80fd\u529b\uff1a\u667a\u80fd\u6807\u7b7e\u3001\u81ea\u52a8\u5206\u7c7b\u3001\u5185\u5bb9\u6458\u8981\u3001\u667a\u80fd\u641c\u7d22\u3002\u8986\u76d6\u5b9e\u73b0\u539f\u7406\u3001\u6280\u672f\u65b9\u6848\u3001\u5b9e\u6218\u8e29\u5751\uff0c\u4ee5\u53ca\u5982\u4f55\u8bc4\u4f30AI\u80fd\u529b\u7684\u5b9e\u9645\u4ef7\u503c\u3002<\/p>\n<hr \/>\n<h2>\u4e00\u3001AI\u80fd\u529b\u5168\u666f\u56fe\uff1a\u4f01\u4e1a\u4e91\u76d8\u9700\u8981\u7684\u4e94\u7c7b\u667a\u80fd<\/h2>\n<h3>1.1 \u667a\u80fd\u6807\u7b7e\uff08Smart Tags\uff09<\/h3>\n<p>\u4f20\u7edf\u6807\u7b7e\u662f\u624b\u52a8\u6253\u7684\uff0cAI\u6807\u7b7e\u662f\u7cfb\u7edf\u81ea\u52a8\u751f\u6210\u7684\u3002<\/p>\n<p><strong>\u7528\u6237\u884c\u4e3a<\/strong>\uff1a\u4e0a\u4f20\u4e00\u4e2aPPT\uff0cAI\u81ea\u52a8\u8bc6\u522b\u51fa&#8221;\u9879\u76ee\u7ba1\u7406&#8221;&#8221;\u5b63\u5ea6\u6c47\u62a5&#8221;&#8221;Q3&#8243;\u6807\u7b7e\uff0c\u7528\u6237\u53ef\u4ee5\u76f4\u63a5\u4f7f\u7528\u6216\u4fee\u6539\u3002<\/p>\n<p><strong>\u6280\u672f\u672c\u8d28<\/strong>\uff1a\u5bf9\u6587\u4ef6\u5185\u5bb9\u8fdb\u884c\u8bed\u4e49\u5206\u6790\uff0c\u63d0\u53d6\u5173\u952e\u5b9e\u4f53\uff08\u4eba\u540d\u3001\u5730\u70b9\u3001\u9879\u76ee\u540d\u3001\u4ea7\u54c1\u540d\uff09\u4ee5\u53ca\u5206\u7c7b\u6807\u7b7e\u3002<\/p>\n<p><strong>\u96be\u70b9<\/strong>\uff1a\u4e2d\u6587\u8bed\u5883\u4e0b\u7684\u5b9e\u4f53\u8bc6\u522b\u6bd4\u82f1\u6587\u590d\u6742\uff0c\u9700\u8981\u8bad\u7ec3\u6216fine-tune\u7684NER\uff08\u547d\u540d\u5b9e\u4f53\u8bc6\u522b\uff09\u6a21\u578b\u3002<\/p>\n<h3>1.2 \u81ea\u52a8\u5206\u7c7b\uff08Auto Classification\uff09<\/h3>\n<p>\u628a\u6587\u4ef6\u81ea\u52a8\u5f52\u5230\u6b63\u786e\u7684\u6587\u4ef6\u5939\u6216\u77e5\u8bc6\u5e93\u5206\u7c7b\u4e2d\u3002<\/p>\n<p><strong>\u7528\u6237\u884c\u4e3a<\/strong>\uff1a\u4e0a\u4f20\u4e00\u4efd\u5408\u540c\uff0cAI\u8bc6\u522b\u51fa\u8fd9\u662f\u4e00\u4efd&#8221;\u91c7\u8d2d\u5408\u540c&#8221;\uff0c\u81ea\u52a8\u5f52\u6863\u5230&#8221;\u5408\u540c\u7ba1\u7406\/\u4f9b\u5e94\u5546\u540d\u79f0\/2024\u5e74&#8221;\u8def\u5f84\u4e0b\u3002<\/p>\n<p><strong>\u6280\u672f\u672c\u8d28<\/strong>\uff1a\u591a\u5206\u7c7b\u95ee\u9898\uff0c\u6a21\u578b\u8f93\u51fa\u6587\u4ef6\u5c5e\u4e8e\u54ea\u4e2a\u5206\u7c7b\u7684\u6982\u7387\u3002\u9700\u8981\u7ed3\u5408\u6587\u4ef6\u5143\u6570\u636e\uff08\u7c7b\u578b\u3001\u5927\u5c0f\u3001\u4e0a\u4f20\u8005\uff09\u548c\u5185\u5bb9\u7279\u5f81\u7efc\u5408\u5224\u65ad\u3002<\/p>\n<p><strong>\u96be\u70b9<\/strong>\uff1a\u5206\u7c7b\u4f53\u7cfb\u7531\u5ba2\u6237\u5b9a\u4e49\uff0c\u6bcf\u5bb6\u4f01\u4e1a\u7684\u5206\u7c7b\u6807\u51c6\u4e0d\u540c\uff0c\u9700\u8981\u652f\u6301\u81ea\u5b9a\u4e49\u5206\u7c7b\u6a21\u578b\u6216few-shot learning\u3002<\/p>\n<h3>1.3 \u5185\u5bb9\u6458\u8981\uff08Content Summarization\uff09<\/h3>\n<p>\u4ece\u6587\u6863\u4e2d\u63d0\u53d6\u5173\u952e\u4fe1\u606f\uff0c\u751f\u6210\u6458\u8981\u3001\u63d0\u53d6\u5173\u952e\u53e5\u5b50\u3001\u751f\u6210\u76ee\u5f55\u3002<\/p>\n<p><strong>\u7528\u6237\u884c\u4e3a<\/strong>\uff1a<br \/>\n&#8211; \u5bf9\u7740\u4e00\u4efd50\u9875\u7684PDF\u62a5\u544a\uff0c\u70b9\u4e00\u4e0b&#8221;AI\u6458\u8981&#8221;\uff0c\u5f97\u5230\u4e00\u6bb5300\u5b57\u7684\u6458\u8981<br \/>\n&#8211; \u5bf9\u7740\u4e00\u6bb5\u4f1a\u8bae\u5f55\u97f3\uff0c\u70b9\u4e00\u4e0b&#8221;\u751f\u6210\u7eaa\u8981&#8221;\uff0c\u5f97\u5230\u4e00\u4efd\u7ed3\u6784\u5316\u7684\u4f1a\u8bae\u8bb0\u5f55<\/p>\n<p><strong>\u6280\u672f\u672c\u8d28<\/strong>\uff1a\u6587\u672c\u6458\u8981\u4efb\u52a1\uff0c\u5206\u62bd\u53d6\u5f0f\uff08Extractive\uff09\u548c\u751f\u6210\u5f0f\uff08Abstractive\uff09\u3002\u751f\u6210\u5f0f\u6548\u679c\u66f4\u597d\u4f46\u8ba1\u7b97\u6210\u672c\u9ad8\u3002<\/p>\n<p><strong>\u96be\u70b9<\/strong>\uff1a\u957f\u6587\u6863\uff08&gt;100\u9875\uff09\u7684\u6458\u8981\u9700\u8981\u5206\u6bb5\u5904\u7406\u518d\u5408\u5e76\uff0c\u4fe1\u606f\u635f\u5931\u63a7\u5236\u662f\u5173\u952e\u3002<\/p>\n<h3>1.4 \u667a\u80fd\u641c\u7d22\uff08Intelligent Search\uff09<\/h3>\n<p>\u4e0d\u4ec5\u641c\u5173\u952e\u8bcd\uff0c\u8fd8\u80fd\u7406\u89e3\u8bed\u4e49\u3001\u627e\u5230\u76f8\u5173\u6587\u6863\u3002<\/p>\n<p><strong>\u7528\u6237\u884c\u4e3a<\/strong>\uff1a\u641c&#8221;\u53bb\u5e74Q3\u534e\u4e1c\u533a\u5927\u5ba2\u6237\u9879\u76ee\u7684\u590d\u76d8&#8221;\uff0c\u80fd\u641c\u5230\u6240\u6709\u76f8\u5173\u7684\u9879\u76ee\u6587\u6863\u3001\u4f1a\u8bae\u8bb0\u5f55\u3001\u90ae\u4ef6\u5f80\u6765\u2014\u2014\u5373\u4f7f\u8fd9\u4e9b\u6587\u6863\u91cc\u6ca1\u6709\u76f4\u63a5\u51fa\u73b0&#8221;\u590d\u76d8&#8221;\u8fd9\u4e2a\u8bcd\u3002<\/p>\n<p><strong>\u6280\u672f\u672c\u8d28<\/strong>\uff1a\u5411\u91cf\u68c0\u7d22\uff08Vector Search\uff09\uff0c\u5c06\u6587\u4ef6\u5185\u5bb9\u7f16\u7801\u4e3a\u5411\u91cf\uff0c\u641c\u7d22\u65f6\u8ba1\u7b97\u8bed\u4e49\u76f8\u4f3c\u5ea6\u3002<\/p>\n<p><strong>\u96be\u70b9<\/strong>\uff1a\u9700\u8981\u5927\u91cf\u8bad\u7ec3\u6570\u636e\u6216fine-tune\u624d\u80fd\u8fbe\u5230\u5b9e\u7528\u7cbe\u5ea6\uff0c\u79c1\u6709\u90e8\u7f72\u573a\u666f\u4e0b\u7684\u6a21\u578b\u9009\u62e9\u6709\u9650\u3002<\/p>\n<h3>1.5 \u667a\u80fd\u95ee\u7b54\uff08Intelligent Q&amp;A\uff09<\/h3>\n<p>\u57fa\u4e8e\u4e91\u76d8\u5185\u5bb9\uff0c\u56de\u7b54\u7528\u6237\u7684\u5177\u4f53\u95ee\u9898\u3002<\/p>\n<p><strong>\u7528\u6237\u884c\u4e3a<\/strong>\uff1a\u95ee&#8221;\u8fd9\u4e2a\u9879\u76ee\u7684\u9884\u7b97\u8d85\u652f\u4e86\u591a\u5c11\uff1f&#8221;AI\u4ece\u4e91\u76d8\u91cc\u7684\u5404\u79cd\u6587\u6863\u4e2d\u627e\u5230\u7b54\u6848\u3002<\/p>\n<p><strong>\u6280\u672f\u672c\u8d28<\/strong>\uff1aRAG\uff08\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff09\uff0c\u5148\u68c0\u7d22\u76f8\u5173\u6587\u6863\u7247\u6bb5\uff0c\u518d\u8ba9LLM\u6839\u636e\u68c0\u7d22\u7ed3\u679c\u751f\u6210\u7b54\u6848\u3002<\/p>\n<p><strong>\u96be\u70b9<\/strong>\uff1aRAG\u7cfb\u7edf\u7684\u56de\u7b54\u8d28\u91cf\u53d6\u51b3\u4e8e\u68c0\u7d22\u8d28\u91cf\uff0c\u68c0\u7d22\u5931\u8d25\u5219\u6574\u4e2a\u7cfb\u7edf\u5931\u8d25\u3002<\/p>\n<hr \/>\n<h2>\u4e8c\u3001\u667a\u80fd\u6807\u7b7e\u7684\u6280\u672f\u5b9e\u73b0<\/h2>\n<h3>2.1 \u6807\u7b7e\u751f\u6210\u7684\u4e24\u79cd\u8def\u7ebf<\/h3>\n<p><strong>\u8def\u7ebf\u4e00\uff1a\u89c4\u5219\u5339\u914d\uff08\u8f7b\u91cf\u7ea7\uff09<\/strong><\/p>\n<pre><code class=\"language-python\">import re\r\nfrom collections import Counter\r\n\r\ndef extract_tags_by_rules(text: str, file_name: str) -&gt; list:\r\n    &quot;&quot;&quot;\u57fa\u4e8e\u89c4\u5219\u63d0\u53d6\u6807\u7b7e\uff08\u8f7b\u91cf\u65b9\u6848\uff09&quot;&quot;&quot;\r\n    tags = set()\r\n\r\n    # \u4ece\u6587\u4ef6\u540d\u63d0\u53d6\u660e\u663e\u6807\u7b7e\uff08\u9879\u76ee\u540d\u3001\u5e74\u4efd\u3001\u7c7b\u578b\uff09\r\n    name_patterns = [\r\n        r'(Q[1-4])',           # Q3\r\n        r'(20\\d{2})',          # 2024\r\n        r'(\u534e\u4e1c|\u534e\u5357|\u534e\u5317)',    # \u533a\u57df\r\n        r'(\u5408\u540c|\u62a5\u544a|\u65b9\u6848|\u62a5\u4ef7\u5355|\u53d1\u7968)',  # \u6587\u4ef6\u7c7b\u578b\r\n    ]\r\n    for pattern in name_patterns:\r\n        matches = re.findall(pattern, file_name)\r\n        tags.update(matches)\r\n\r\n    # \u4ece\u5185\u5bb9\u63d0\u53d6\u9ad8\u9891\u8bcd\uff08\u53bb\u6389\u505c\u7528\u8bcd\uff09\r\n    words = jieba.cut(text)\r\n    freq = Counter([w for w in words if len(w) &gt;= 2 and w not in STOPWORDS])\r\n    top_tags = [w for w, c in freq.most_common(10)]\r\n    tags.update(top_tags)\r\n\r\n    return list(tags)\r\n\r\n# \u505c\u7528\u8bcd\u8868\uff08\u9700\u7ef4\u62a4\uff09\r\nSTOPWORDS = {'\u7684', '\u662f', '\u5728', '\u548c', '\u4e86', '\u6709', '\u6211\u4eec', '\u8fd9\u4e2a', '\u4ec0\u4e48', '\u4e00\u4e2a', '\u53ef\u4ee5'}\r\n<\/code><\/pre>\n<p><strong>\u8def\u7ebf\u4e8c\uff1aNER\u6a21\u578b\uff08\u9ad8\u8d28\u91cf\uff09<\/strong><\/p>\n<pre><code class=\"language-python\">from transformers import pipeline, AutoTokenizer, AutoModelForTokenClassification\r\n\r\n# \u4f7f\u7528\u4e2d\u6587NER\u6a21\u578b\u63d0\u53d6\u5b9e\u4f53\r\nner_pipeline = pipeline(\r\n    &quot;ner&quot;, \r\n    model=&quot;uer\/roberta-base-chinese-ner&quot;,\r\n    aggregation_strategy=&quot;simple&quot;\r\n)\r\n\r\ndef extract_entities(text: str, max_length: int = 512) -&gt; dict:\r\n    &quot;&quot;&quot;\u4f7f\u7528NER\u6a21\u578b\u63d0\u53d6\u5b9e\u4f53&quot;&quot;&quot;\r\n    # \u622a\u65ad\u8d85\u957f\u6587\u672c\uff08BERT\u6700\u5927512 tokens\uff09\r\n    truncated = text[:max_length]\r\n\r\n    entities = ner_pipeline(truncated)\r\n\r\n    # \u6309\u7c7b\u578b\u805a\u5408\r\n    result = {\r\n        'person': [],\r\n        'organization': [],\r\n        'location': [],\r\n        'project': [],\r\n        'product': []\r\n    }\r\n\r\n    for ent in entities:\r\n        entity_text = ent['word']\r\n        entity_type = ent['entity_group']\r\n\r\n        if entity_type == 'PER':  # \u4eba\u540d\r\n            result['person'].append(entity_text)\r\n        elif entity_type == 'ORG':  # \u673a\u6784\u540d\r\n            result['organization'].append(entity_text)\r\n        elif entity_type == 'LOC':  # \u5730\u540d\r\n            result['location'].append(entity_text)\r\n\r\n    return result\r\n\r\ndef generate_tags_from_entities(entities: dict) -&gt; list:\r\n    &quot;&quot;&quot;\u4ece\u5b9e\u4f53\u751f\u6210\u6807\u7b7e&quot;&quot;&quot;\r\n    tags = []\r\n\r\n    # \u673a\u6784\u540d \u2192 \u9879\u76ee\/\u516c\u53f8\u6807\u7b7e\r\n    for org in entities.get('organization', []):\r\n        tags.append(f&quot;\u673a\u6784:{org}&quot;)\r\n\r\n    # \u5730\u540d \u2192 \u533a\u57df\u6807\u7b7e\r\n    for loc in entities.get('location', []):\r\n        if loc in ['\u534e\u4e1c', '\u534e\u5357', '\u534e\u5317', '\u897f\u90e8']:\r\n            tags.append(f&quot;\u533a\u57df:{loc}&quot;)\r\n\r\n    return tags\r\n<\/code><\/pre>\n<h3>2.2 \u6807\u7b7e\u7f6e\u4fe1\u5ea6\u4e0e\u4eba\u5de5\u590d\u6838<\/h3>\n<p>AI\u6807\u7b7e\u4e0d\u53ef\u80fd100%\u51c6\u786e\uff0c\u9700\u8981\u8bbe\u8ba1\u4eba\u5de5\u590d\u6838\u673a\u5236\uff1a<\/p>\n<pre><code class=\"language-python\">class TagReviewSystem:\r\n    &quot;&quot;&quot;\u6807\u7b7e\u590d\u6838\u7cfb\u7edf\uff1a\u7f6e\u4fe1\u5ea6\u4f4e\u7684\u6807\u7b7e\u9700\u8981\u4eba\u5de5\u786e\u8ba4&quot;&quot;&quot;\r\n\r\n    HIGH_CONFIDENCE = 0.85  # \u9ad8\u7f6e\u4fe1\u5ea6\uff0c\u76f4\u63a5\u91c7\u7528\r\n    MEDIUM_CONFIDENCE = 0.60  # \u4e2d\u7f6e\u4fe1\u5ea6\uff0c\u7528\u6237\u53ef\u4e00\u952e\u91c7\u7eb3\r\n    LOW_CONFIDENCE = 0.40  # \u4f4e\u7f6e\u4fe1\u5ea6\uff0c\u9700\u4eba\u5de5\u8f93\u5165\r\n\r\n    def process_file(self, file_id: str, ai_tags: list) -&gt; dict:\r\n        &quot;&quot;&quot;\u5904\u7406\u6587\u4ef6\u7684AI\u6807\u7b7e\uff0c\u6309\u7f6e\u4fe1\u5ea6\u5206\u6d41&quot;&quot;&quot;\r\n        review_queue = []\r\n        auto_tags = []\r\n\r\n        for tag in ai_tags:\r\n            if tag.confidence &gt;= self.HIGH_CONFIDENCE:\r\n                auto_tags.append(tag.name)  # \u76f4\u63a5\u91c7\u7528\r\n            elif tag.confidence &gt;= self.MEDIUM_CONFIDENCE:\r\n                review_queue.append({\r\n                    'file_id': file_id,\r\n                    'tag': tag.name,\r\n                    'confidence': tag.confidence,\r\n                    'reason': tag.reason,\r\n                    'suggestions': tag.alternatives  # \u5907\u9009\u6807\u7b7e\r\n                })\r\n            else:\r\n                # \u4f4e\u7f6e\u4fe1\u5ea6\u6807\u7b7e\u4e0d\u5c55\u793a\uff0c\u4f46\u8bb0\u5f55\u4f9b\u6a21\u578b\u5b66\u4e60\r\n                pass\r\n\r\n        return {\r\n            'auto_tags': auto_tags,\r\n            'review_queue': review_queue\r\n        }\r\n\r\n    def submit_review(self, file_id: str, tag: str, action: str):\r\n        &quot;&quot;&quot;\u7528\u6237\u53cd\u9988\uff1a\u91c7\u7eb3\/\u62d2\u7edd\/\u4fee\u6539\u6807\u7b7e&quot;&quot;&quot;\r\n        # \u8bb0\u5f55\u53cd\u9988\u7528\u4e8e\u540e\u7eed\u6a21\u578b\u4f18\u5316\r\n        self.feedback_log.write({\r\n            'file_id': file_id,\r\n            'tag': tag,\r\n            'action': action,  # 'accept'\/'reject'\/'modify'\r\n            'timestamp': time.time()\r\n        })\r\n\r\n        if action == 'accept':\r\n            self.apply_tag(file_id, tag)\r\n        elif action == 'modify':\r\n            self.apply_modified_tag(file_id, tag, user_input)\r\n<\/code><\/pre>\n<hr \/>\n<h2>\u4e09\u3001\u81ea\u52a8\u5206\u7c7b\u7684\u6280\u672f\u5b9e\u73b0<\/h2>\n<h3>3.1 \u5206\u7c7b\u4f53\u7cfb\u8bbe\u8ba1<\/h3>\n<p>\u5728\u52a8\u624b\u5199\u4ee3\u7801\u4e4b\u524d\uff0c\u5fc5\u987b\u5148\u8bbe\u8ba1\u597d\u5206\u7c7b\u4f53\u7cfb\u3002\u5206\u7c7b\u4f53\u7cfb\u76f4\u63a5\u5f71\u54cd\u6a21\u578b\u8bad\u7ec3\u548c\u7528\u6237\u4f7f\u7528\u4f53\u9a8c\u3002<\/p>\n<p><strong>\u539f\u5219\u4e00\uff1a\u5c42\u7ea7\u4e0d\u5b9c\u8fc7\u6df1<\/strong><\/p>\n<pre><code>\u2705 \u597d\uff1a3\u5c42\u4ee5\u5185\r\n  \u5408\u540c \u2192 \u91c7\u8d2d\u5408\u540c \/ \u9500\u552e\u5408\u540c \/ \u5458\u5de5\u5408\u540c\r\n\r\n\u274c \u5dee\uff1a5\u5c42\u4ee5\u4e0a\r\n  \u516c\u53f8 \u2192 \u96c6\u56e2\u603b\u90e8 \u2192 \u6cd5\u52a1\u90e8 \u2192 \u5408\u540c\u7ba1\u7406 \u2192 2024\u5e74 \u2192 \u91c7\u8d2d\u5408\u540c\r\n<\/code><\/pre>\n<p><strong>\u539f\u5219\u4e8c\uff1a\u5206\u7c7b\u4e4b\u95f4\u4e92\u65a5<\/strong><\/p>\n<p>\u540c\u4e00\u4efd\u6587\u4ef6\u5e94\u8be5\u53ea\u80fd\u5c5e\u4e8e\u4e00\u4e2a\u5206\u7c7b\uff0c\u5982\u679c\u5206\u7c7b\u4e4b\u95f4\u6709\u91cd\u53e0\uff0c\u6a21\u578b\u96be\u4ee5\u5b66\u4e60\u3002<\/p>\n<p><strong>\u539f\u5219\u4e09\uff1a\u5206\u7c7b\u6570\u91cf\u9002\u4e2d<\/strong><\/p>\n<ul>\n<li>\u6587\u4ef6\u91cf&lt;1\u4e07\uff1a10-20\u4e2a\u5206\u7c7b<\/li>\n<li>\u6587\u4ef6\u91cf&lt;10\u4e07\uff1a20-50\u4e2a\u5206\u7c7b<\/li>\n<li>\u6587\u4ef6\u91cf&gt;10\u4e07\uff1a\u53ef\u8003\u8651\u591a\u7ea7\u5206\u7c7b\uff0c\u5148\u5927\u7c7b\u518d\u5c0f\u7c7b<\/li>\n<\/ul>\n<h3>3.2 \u5206\u7c7b\u6a21\u578b\u8bad\u7ec3<\/h3>\n<pre><code class=\"language-python\">from transformers import BertForSequenceClassification, Trainer, TrainingArguments\r\nfrom torch.utils.data import Dataset\r\n\r\nclass FileDataset(Dataset):\r\n    def __init__(self, files: list, labels: list, tokenizer, max_len=512):\r\n        self.files = files\r\n        self.labels = labels\r\n        self.tokenizer = tokenizer\r\n        self.max_len = max_len\r\n\r\n    def __len__(self):\r\n        return len(self.files)\r\n\r\n    def __getitem__(self, idx):\r\n        text = self.files[idx]['content'][:self.max_len]\r\n        encoding = self.tokenizer(\r\n            text,\r\n            truncation=True,\r\n            max_length=self.max_len,\r\n            padding='max_length'\r\n        )\r\n        return {\r\n            'input_ids': encoding['input_ids'],\r\n            'attention_mask': encoding['attention_mask'],\r\n            'labels': self.labels[idx]\r\n        }\r\n\r\ndef train_classifier(train_data: list, label_map: dict, output_dir: str):\r\n    &quot;&quot;&quot;\u8bad\u7ec3\u6587\u4ef6\u5206\u7c7b\u6a21\u578b&quot;&quot;&quot;\r\n\r\n    tokenizer = BertTokenizer.from_pretrained('bert-base-chinese')\r\n    model = BertForSequenceClassification(\r\n        num_labels=len(label_map),\r\n        problem_type=&quot;single_label_classification&quot;\r\n    )\r\n\r\n    train_dataset = FileDataset(\r\n        [f['content'] for f in train_data],\r\n        [label_map[f['category']] for f in train_data],\r\n        tokenizer\r\n    )\r\n\r\n    training_args = TrainingArguments(\r\n        output_dir=output_dir,\r\n        num_train_epochs=10,\r\n        per_device_train_batch_size=16,\r\n        learning_rate=2e-5,\r\n        evaluation_strategy=&quot;epoch&quot;,\r\n        save_strategy=&quot;epoch&quot;,\r\n        load_best_model_at_end=True\r\n    )\r\n\r\n    trainer = Trainer(\r\n        model=model,\r\n        args=training_args,\r\n        train_dataset=train_dataset,\r\n    )\r\n\r\n    trainer.train()\r\n    return model\r\n<\/code><\/pre>\n<h3>3.3 \u5206\u7c7b\u7f6e\u4fe1\u5ea6\u4e0e\u670d\u52a1\u5316<\/h3>\n<pre><code class=\"language-python\">import torch\r\nfrom flask import Flask, request, jsonify\r\n\r\napp = Flask(__name__)\r\n\r\n# \u52a0\u8f7d\u6a21\u578b\uff08\u751f\u4ea7\u73af\u5883\u5efa\u8bae\u5355\u72ec\u90e8\u7f72\u4e3a\u5fae\u670d\u52a1\uff09\r\nmodel = BertForSequenceClassification.from_pretrained('.\/classification_model')\r\nmodel.eval()\r\ntokenizer = BertTokenizer.from_pretrained('bert-base-chinese')\r\n\r\n@app.route('\/classify', methods=['POST'])\r\ndef classify_file():\r\n    data = request.json\r\n    file_id = data['file_id']\r\n    content = data['content'][:512]  # BERT\u6700\u5927\u957f\u5ea6\r\n\r\n    encoding = tokenizer(\r\n        content,\r\n        return_tensors='pt',\r\n        truncation=True,\r\n        max_length=512\r\n    )\r\n\r\n    with torch.no_grad():\r\n        outputs = model(**encoding)\r\n        probs = torch.softmax(outputs.logits, dim=1)[0]\r\n        pred_idx = torch.argmax(probs).item()\r\n        confidence = probs[pred_idx].item()\r\n\r\n    return jsonify({\r\n        'file_id': file_id,\r\n        'category': ID_TO_CATEGORY[pred_idx],\r\n        'confidence': round(confidence, 3),\r\n        'alternatives': [\r\n            {'category': ID_TO_CATEGORY[i], 'prob': round(probs[i].item(), 3)}\r\n            for i in range(len(probs)) if i != pred_idx and probs[i].item() &gt; 0.1\r\n        ]\r\n    })\r\n<\/code><\/pre>\n<hr \/>\n<h2>\u56db\u3001\u5185\u5bb9\u6458\u8981\u7684\u6280\u672f\u5b9e\u73b0<\/h2>\n<h3>4.1 \u62bd\u53d6\u5f0f\u6458\u8981 vs \u751f\u6210\u5f0f\u6458\u8981<\/h3>\n<p><strong>\u62bd\u53d6\u5f0f\uff08Extractive\uff09<\/strong>\uff1a<br \/>\n&#8211; \u4ece\u539f\u6587\u4e2d\u76f4\u63a5\u9009\u53d6\u91cd\u8981\u53e5\u5b50\u62fc\u63a5\u6210\u6458\u8981<br \/>\n&#8211; \u5b9e\u73b0\u7b80\u5355\uff0c\u8ba1\u7b97\u91cf\u5c0f\uff0c\u5fe0\u5b9e\u5ea6\u9ad8\uff08\u4e0d\u4f1a\u4e71\u7f16\uff09<br \/>\n&#8211; \u7f3a\u70b9\uff1a\u6458\u8981\u4e0d\u6d41\u7545\uff0c\u53e5\u5b50\u4e4b\u95f4\u53ef\u80fd\u4e0d\u8fde\u8d2f<\/p>\n<p><strong>\u751f\u6210\u5f0f\uff08Abstractive\uff09<\/strong>\uff1a<br \/>\n&#8211; \u6a21\u578b\u7406\u89e3\u539f\u6587\u540e\uff0c\u91cd\u65b0\u751f\u6210\u4e00\u6bb5\u6458\u8981<br \/>\n&#8211; \u6548\u679c\u66f4\u597d\uff0c\u53ef\u4ee5\u6539\u5199\u3001\u603b\u7ed3\u3001\u63d0\u70bc<br \/>\n&#8211; \u7f3a\u70b9\uff1a\u8ba1\u7b97\u91cf\u5927\uff0c\u53ef\u80fd\u4ea7\u751f\u4e0d\u5fe0\u5b9e\u539f\u6587\u7684\u5185\u5bb9\uff08hallucination\uff09<\/p>\n<p><strong>\u4f01\u4e1a\u4e91\u76d8\u573a\u666f\u7684\u5efa\u8bae<\/strong>\uff1a<br \/>\n&#8211; \u5bf9\u51c6\u786e\u6027\u8981\u6c42\u9ad8\uff08\u5408\u540c\u3001\u62a5\u544a\uff09\u2192 \u62bd\u53d6\u5f0f<br \/>\n&#8211; \u5bf9\u53ef\u8bfb\u6027\u8981\u6c42\u9ad8\uff08\u4f1a\u8bae\u7eaa\u8981\u3001\u65b0\u95fb\uff09\u2192 \u751f\u6210\u5f0f<br \/>\n&#8211; \u53ef\u4ee5\u7528\u6df7\u5408\u65b9\u6848\uff1a\u62bd\u53d6\u5173\u952e\u53e5\u5b50 + \u751f\u6210\u5f0f\u6539\u5199<\/p>\n<h3>4.2 \u62bd\u53d6\u5f0f\u6458\u8981\u5b9e\u73b0<\/h3>\n<pre><code class=\"language-python\">import math\r\nfrom sklearn.feature_extraction.text import TfidfVectorizer\r\n\r\ndef extractive_summary(text: str, top_n: int = 5) -&gt; str:\r\n    &quot;&quot;&quot;TF-IDF\u62bd\u53d6\u5f0f\u6458\u8981\uff1a\u63d0\u53d6\u6700\u91cd\u8981\u7684N\u53e5\u8bdd&quot;&quot;&quot;\r\n\r\n    # \u5206\u53e5\r\n    sentences = sent_tokenize(text)\r\n\r\n    if len(sentences) &lt;= top_n:\r\n        return text\r\n\r\n    # \u8ba1\u7b97\u6bcf\u53e5\u8bdd\u7684TF-IDF\u5f97\u5206\r\n    vectorizer = TfidfVectorizer()\r\n    tfidf_matrix = vectorizer.fit_transform(sentences)\r\n\r\n    # \u53d6\u7b2c\u4e00\u53e5\uff08\u901a\u5e38\u662f\u603b\u8d77\uff09\u4f5c\u4e3a\u91cd\u8981\u53e5\u5b50\r\n    scores = tfidf_matrix.sum(axis=1).A1\r\n\r\n    # \u5982\u679c\u6587\u6863\u6709\u6807\u9898\uff0c\u7b2c\u4e00\u53e5\u6743\u91cd\u66f4\u9ad8\r\n    # \uff08\u5b9e\u9645\u4e2d\u53ef\u4ee5\u4ece\u6587\u4ef6\u5143\u6570\u636e\u83b7\u53d6\u6807\u9898\u4fe1\u606f\uff09\r\n\r\n    # \u9009\u53d6\u5f97\u5206\u6700\u9ad8\u7684N\u53e5\r\n    top_indices = scores.argsort()[-top_n:]\r\n\r\n    # \u6309\u539f\u6587\u987a\u5e8f\u91cd\u6392\r\n    top_indices_sorted = sorted(top_indices)\r\n\r\n    summary = ' '.join([sentences[i] for i in top_indices_sorted])\r\n\r\n    return summary\r\n\r\n# \u66f4\u9ad8\u7ea7\u7684\u62bd\u53d6\u65b9\u6cd5\uff1aTextRank\r\ndef textrank_summary(text: str, top_n: int = 5) -&gt; str:\r\n    &quot;&quot;&quot;TextRank\u7b97\u6cd5\u62bd\u53d6\u6458\u8981&quot;&quot;&quot;\r\n    from collections import Counter\r\n\r\n    sentences = sent_tokenize(text)\r\n\r\n    # \u6784\u5efa\u53e5\u5b50\u76f8\u4f3c\u5ea6\u77e9\u9635\r\n    n = len(sentences)\r\n    similarity_matrix = [[0.0] * n for _ in range(n)]\r\n\r\n    for i in range(n):\r\n        for j in range(i + 1, n):\r\n            sim = cosine_similarity(sentences[i], sentences[j])\r\n            similarity_matrix[i][j] = sim\r\n            similarity_matrix[j][i] = sim\r\n\r\n    # TextRank\u8fed\u4ee3\r\n    scores = [1.0] * n\r\n    damping = 0.85\r\n    for _ in range(50):\r\n        new_scores = [(1 - damping) + damping * sum(similarity_matrix[i][j] * scores[j] for j in range(n) if j != i) \/ max(sum(similarity_matrix[i]) - similarity_matrix[i][i], 0.001) for i in range(n)]\r\n        if max(abs(new_scores[i] - scores[i]) for i in range(n)) &lt; 0.0001:\r\n            break\r\n        scores = new_scores\r\n\r\n    # \u9009\u53d6\u5f97\u5206\u6700\u9ad8\u7684\u53e5\u5b50\r\n    top_indices = sorted(range(len(scores)), key=lambda i: scores[i])[-top_n:]\r\n    top_indices_sorted = sorted(top_indices)\r\n\r\n    return ' '.join([sentences[i] for i in top_indices_sorted])\r\n<\/code><\/pre>\n<h3>4.3 \u751f\u6210\u5f0f\u6458\u8981\u5b9e\u73b0\uff08LLM API\uff09<\/h3>\n<pre><code class=\"language-python\">import openai\r\n\r\ndef generate_summary(content: str, summary_type: str = 'brief') -&gt; str:\r\n    &quot;&quot;&quot;\u4f7f\u7528LLM\u751f\u6210\u6458\u8981\uff08GPT-4\/Claude\/\u56fd\u4ea7LLM\uff09&quot;&quot;&quot;\r\n\r\n    prompt = f&quot;&quot;&quot;\u8bf7\u4e3a\u4ee5\u4e0b\u6587\u6863\u751f\u6210{'\u7b80\u77ed\u6458\u8981' if summary_type == 'brief' else '\u8be6\u7ec6\u6458\u8981'}\u3002\r\n\r\n\u8981\u6c42\uff1a\r\n- \u7b80\u77ed\u6458\u8981\uff1a100\u5b57\u4ee5\u5185\uff0c\u6db5\u76d6\u6838\u5fc3\u8981\u70b9\r\n- \u8be6\u7ec6\u6458\u8981\uff1a300\u5b57\u4ee5\u5185\uff0c\u6db5\u76d6\u4e3b\u8981\u5185\u5bb9\u548c\u5173\u952e\u6570\u636e\r\n\r\n\u6587\u6863\u5185\u5bb9\uff1a\r\n---\r\n{content[:8000]}  # LLM\u6709\u4e0a\u4e0b\u6587\u9650\u5236\uff0c\u957f\u6587\u6863\u9700\u622a\u65ad\r\n---\r\n\r\n\u6458\u8981\uff1a&quot;&quot;&quot;\r\n\r\n    response = openai.ChatCompletion.create(\r\n        model=&quot;gpt-4&quot;,\r\n        messages=[{&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: prompt}],\r\n        temperature=0.3  # \u6458\u8981\u9700\u8981\u4f4e\u968f\u673a\u6027\r\n    )\r\n\r\n    return response.choices[0].message.content.strip()\r\n\r\ndef generate_meeting_notes(transcript: str) -&gt; dict:\r\n    &quot;&quot;&quot;\u751f\u6210\u4f1a\u8bae\u7eaa\u8981\uff08\u7ed3\u6784\u5316\u6458\u8981\uff09&quot;&quot;&quot;\r\n\r\n    prompt = f&quot;&quot;&quot;\u8bf7\u4ece\u4ee5\u4e0b\u4f1a\u8bae\u5f55\u97f3\u6587\u672c\u4e2d\u63d0\u53d6\u4fe1\u606f\uff0c\u751f\u6210\u7ed3\u6784\u5316\u4f1a\u8bae\u7eaa\u8981\u3002\r\n\r\n\u4f1a\u8bae\u5f55\u97f3\uff1a\r\n---\r\n{transcript[:12000]}\r\n---\r\n\r\n\u8bf7\u6309\u4ee5\u4e0b\u683c\u5f0f\u8f93\u51fa\uff1a\r\n1. \u4f1a\u8bae\u57fa\u672c\u4fe1\u606f\uff08\u65f6\u95f4\u3001\u53c2\u4f1a\u4eba\u3001\u4f1a\u8bae\u4e3b\u9898\uff09\r\n2. \u8ba8\u8bba\u8981\u70b9\uff083-5\u6761\uff09\r\n3. \u51b3\u7b56\u4e8b\u9879\r\n4. \u5f85\u529e\u4e8b\u9879\uff08\u8d1f\u8d23\u4eba+\u622a\u6b62\u65f6\u95f4\uff09\r\n5. \u884c\u52a8\u9879\r\n\r\n\u683c\u5f0f\u8f93\u51faJSON\uff1a&quot;&quot;&quot;\r\n\r\n    response = openai.ChatCompletion.create(\r\n        model=&quot;gpt-4&quot;,\r\n        messages=[{&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: prompt}],\r\n        response_format={&quot;type&quot;: &quot;json_object&quot;}\r\n    )\r\n\r\n    import json\r\n    return json.loads(response.choices[0].message.content)\r\n<\/code><\/pre>\n<h3>4.4 \u957f\u6587\u6863\u6458\u8981\u7684\u5206\u6bb5\u7b56\u7565<\/h3>\n<p>LLM\u6709\u4e0a\u4e0b\u6587\u7a97\u53e3\u9650\u5236\uff08GPT-4\u901a\u5e38\u662f8K\/32K tokens\uff09\uff0c\u5bf9\u4e8e\u8d85\u957f\u6587\u6863\uff08\u5982100\u9875PDF\uff09\uff0c\u9700\u8981\u5206\u6bb5\u5904\u7406\uff1a<\/p>\n<pre><code class=\"language-python\">def long_document_summary(document: str, max_chunk_size: int = 6000, overlap: int = 200) -&gt; str:\r\n    &quot;&quot;&quot;\u957f\u6587\u6863\u6458\u8981\uff1a\u5206\u6bb5\u5904\u7406\u540e\u5408\u5e76&quot;&quot;&quot;\r\n\r\n    chunks = []\r\n    start = 0\r\n\r\n    # \u6ed1\u52a8\u7a97\u53e3\u5206\u5757\uff08\u91cd\u53e0\u90e8\u5206\u4fdd\u8bc1\u4e0a\u4e0b\u6587\u8fde\u7eed\uff09\r\n    while start &lt; len(document):\r\n        end = start + max_chunk_size\r\n        chunk = document[start:end]\r\n        chunks.append(chunk)\r\n        start = end - overlap  # \u91cd\u53e0200\u5b57\uff0c\u907f\u514d\u5757\u4e4b\u95f4\u5272\u88c2\r\n\r\n    # \u5bf9\u6bcf\u4e2a\u5757\u751f\u6210\u6458\u8981\r\n    chunk_summaries = []\r\n    for i, chunk in enumerate(chunks):\r\n        summary = generate_summary(chunk, summary_type='brief')\r\n        chunk_summaries.append(f&quot;[Part {i+1}] {summary}&quot;)\r\n\r\n    # \u5408\u5e76\u6240\u6709\u5757\u6458\u8981\uff0c\u518d\u6b21\u6c47\u603b\r\n    combined = '\\n'.join(chunk_summaries)\r\n    final_summary = generate_summary(\r\n        combined,\r\n        summary_type='detailed'  # \u6700\u540e\u7528\u8be6\u7ec6\u6458\u8981\r\n    )\r\n\r\n    return final_summary\r\n<\/code><\/pre>\n<hr \/>\n<h2>\u4e94\u3001\u667a\u80fd\u641c\u7d22\u7684\u6280\u672f\u5b9e\u73b0\uff08RAG\u67b6\u6784\uff09<\/h2>\n<h3>5.1 \u4e3a\u4ec0\u4e48\u8981RAG\uff1a\u7eafLLM\u7684\u5c40\u9650<\/h3>\n<p>\u76f4\u63a5\u8ba9LLM\u56de\u7b54&#8221;\u6211\u7684\u4e91\u76d8\u91cc\u6709\u54ea\u4e9b\u9879\u76ee\u6587\u6863&#8221;\u662f\u4e0d\u53ef\u80fd\u7684\u2014\u2014LLM\u6ca1\u6709\u8bbf\u95ee\u4f60\u7684\u79c1\u6709\u6570\u636e\u3002RAG\uff08Retrieval-Augmented Generation\uff09\u89e3\u51b3\u4e86\u8fd9\u4e2a\u95ee\u9898\uff1a<\/p>\n<pre><code>\u7528\u6237\u95ee\u9898 \u2192 \u68c0\u7d22\u76f8\u5173\u6587\u6863 \u2192 \u628a\u6587\u6863\u4f5c\u4e3a\u4e0a\u4e0b\u6587\u7ed9LLM \u2192 LLM\u751f\u6210\u7b54\u6848\r\n<\/code><\/pre>\n<h3>5.2 RAG\u6d41\u7a0b\u8be6\u89e3<\/h3>\n<pre><code class=\"language-python\">class RAGSearchSystem:\r\n    &quot;&quot;&quot;\u57fa\u4e8eRAG\u7684\u4f01\u4e1a\u4e91\u76d8\u667a\u80fd\u95ee\u7b54\u7cfb\u7edf&quot;&quot;&quot;\r\n\r\n    def __init__(self, \r\n                 file_store: FileStore,           # \u6587\u4ef6\u5b58\u50a8\r\n                 embedding_model,                 # \u5411\u91cf\u5316\u6a21\u578b\r\n                 vector_db: VectorDB,             # \u5411\u91cf\u6570\u636e\u5e93\r\n                 llm: LLM):                       # \u5927\u8bed\u8a00\u6a21\u578b\r\n\r\n        self.files = file_store\r\n        self.embedder = embedding_model\r\n        self.vector_db = vector_db\r\n        self.llm = llm\r\n\r\n    def index_files(self, file_ids: list = None):\r\n        &quot;&quot;&quot;\u4e3a\u6587\u4ef6\u5efa\u7acb\u5411\u91cf\u7d22\u5f15&quot;&quot;&quot;\r\n\r\n        files = self.files.get_files(file_ids) if file_ids else self.files.get_all()\r\n\r\n        for file in files:\r\n            # 1. \u63d0\u53d6\u6587\u672c\u5185\u5bb9\r\n            content = self.files.extract_text(file)\r\n\r\n            # 2. \u5206\u5757\uff08chunks\uff09\u907f\u514d\u5411\u91cf\u8fc7\u957f\r\n            chunks = self.split_into_chunks(content, chunk_size=500, overlap=50)\r\n\r\n            # 3. \u5411\u91cf\u5316\u6bcf\u4e2achunk\r\n            for i, chunk in enumerate(chunks):\r\n                vector = self.embedder.encode(chunk)\r\n\r\n                # 4. \u5199\u5165\u5411\u91cf\u6570\u636e\u5e93\r\n                self.vector_db.insert(\r\n                    id=f&quot;{file['id']}_chunk_{i}&quot;,\r\n                    vector=vector,\r\n                    text=chunk,\r\n                    metadata={\r\n                        'file_id': file['id'],\r\n                        'file_name': file['name'],\r\n                        'chunk_index': i,\r\n                        'permissions': file['permissions']\r\n                    }\r\n                )\r\n\r\n    def answer(self, question: str, user_id: str) -&gt; dict:\r\n        &quot;&quot;&quot;\u56de\u7b54\u7528\u6237\u95ee\u9898&quot;&quot;&quot;\r\n\r\n        # 1. \u5c06\u95ee\u9898\u5411\u91cf\u5316\r\n        question_vector = self.embedder.encode(question)\r\n\r\n        # 2. \u68c0\u7d22\u6700\u76f8\u5173\u7684\u6587\u6863\u5757\uff08\u5e26\u6743\u9650\u8fc7\u6ee4\uff09\r\n        results = self.vector_db.search(\r\n            vector=question_vector,\r\n            top_k=5,\r\n            filter={'permissions': {'$contains': user_id}}  # \u6743\u9650\u8fc7\u6ee4\r\n        )\r\n\r\n        if not results:\r\n            return {'answer': '\u62b1\u6b49\uff0c\u672a\u627e\u5230\u76f8\u5173\u6587\u6863\u3002', 'sources': []}\r\n\r\n        # 3. \u6784\u5efaprompt\uff0c\u628a\u68c0\u7d22\u5230\u7684\u6587\u6863\u4f5c\u4e3a\u4e0a\u4e0b\u6587\r\n        context = '\\n\\n'.join([\r\n            f&quot;[\u6587\u6863{i+1}] {r['text']}&quot;\r\n            for i, r in enumerate(results)\r\n        ])\r\n\r\n        prompt = f&quot;&quot;&quot;\u4f60\u662f\u4e00\u4e2a\u4f01\u4e1a\u77e5\u8bc6\u52a9\u624b\u3002\u8bf7\u6839\u636e\u4ee5\u4e0b\u6587\u6863\u56de\u7b54\u7528\u6237\u7684\u95ee\u9898\u3002\r\n\r\n\u5982\u679c\u6587\u6863\u4e2d\u6ca1\u6709\u76f8\u5173\u4fe1\u606f\uff0c\u8bf7\u5982\u5b9e\u8bf4\u660e&quot;\u6ca1\u6709\u627e\u5230\u76f8\u5173\u4fe1\u606f&quot;\u3002\r\n\r\n\u6587\u6863\u5185\u5bb9\uff1a\r\n{context}\r\n\r\n\u7528\u6237\u95ee\u9898\uff1a{question}\r\n\r\n\u56de\u7b54\uff08\u8bf7\u76f4\u63a5\u7ed9\u51fa\u7b54\u6848\uff0c\u4e0d\u8981\u91cd\u590d\u95ee\u9898\uff09\uff1a&quot;&quot;&quot;\r\n\r\n        # 4. \u8c03\u7528LLM\u751f\u6210\u7b54\u6848\r\n        answer = self.llm.generate(prompt)\r\n\r\n        return {\r\n            'answer': answer,\r\n            'sources': [\r\n                {\r\n                    'file_id': r['metadata']['file_id'],\r\n                    'file_name': r['metadata']['file_name'],\r\n                    'relevance': round(r['score'], 2)\r\n                }\r\n                for r in results\r\n            ]\r\n        }\r\n\r\n    def split_into_chunks(self, text: str, chunk_size: int = 500, overlap: int = 50) -&gt; list:\r\n        &quot;&quot;&quot;\u5c06\u957f\u6587\u672c\u5207\u5206\u4e3a\u91cd\u53e0\u7684\u5757&quot;&quot;&quot;\r\n        import re\r\n        # \u6309\u6bb5\u843d\u5206\u5272\uff08\u4e2d\u6587\u6309\u6362\u884c\u7b26\uff09\r\n        paragraphs = re.split(r'\\n+', text)\r\n\r\n        chunks = []\r\n        current_chunk = []\r\n        current_len = 0\r\n\r\n        for para in paragraphs:\r\n            para_len = len(para)\r\n            if current_len + para_len &gt; chunk_size:\r\n                # \u4fdd\u5b58\u5f53\u524dchunk\uff0c\u5f00\u59cb\u65b0\u7684\r\n                if current_chunk:\r\n                    chunks.append('\\n'.join(current_chunk))\r\n                # \u65b0chunk\u4ece\u5934\u5f00\u59cb\uff0c\u5141\u8bb8oversize\uff08\u5904\u7406\u7279\u522b\u957f\u7684\u6bb5\u843d\uff09\r\n                current_chunk = [para]\r\n                current_len = para_len\r\n            else:\r\n                current_chunk.append(para)\r\n                current_len += para_len + 1  # +1\u662f\u6362\u884c\u7b26\r\n\r\n        if current_chunk:\r\n            chunks.append('\\n'.join(current_chunk))\r\n\r\n        return chunks\r\n<\/code><\/pre>\n<h3>5.3 \u5411\u91cf\u6a21\u578b\u9009\u62e9<\/h3>\n<p><strong>\u5f00\u6e90\u5411\u91cf\u6a21\u578b\u63a8\u8350<\/strong>\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u6a21\u578b<\/th>\n<th>\u7ef4\u5ea6<\/th>\n<th>\u4f18\u52bf<\/th>\n<th>\u52a3\u52bf<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>text2vec-base-Chinese<\/td>\n<td>768<\/td>\n<td>\u4e2d\u6587\u6548\u679c\u597d\uff0c\u5f00\u6e90<\/td>\n<td>\u4f53\u79ef\u5927\uff081.2GB\uff09<\/td>\n<\/tr>\n<tr>\n<td>m3e-base<\/td>\n<td>768<\/td>\n<td>\u4e2d\u6587+\u82f1\u6587\uff0c\u591a\u8bed\u8a00<\/td>\n<td>\u5411\u91cf\u8d28\u91cf\u4e00\u822c<\/td>\n<\/tr>\n<tr>\n<td>BGE-base-zh<\/td>\n<td>768<\/td>\n<td>\u963f\u91cc\u51fa\u54c1\uff0c\u6548\u679c\u597d<\/td>\n<td>\u9700\u8981\u8f83\u9ad8\u914d\u7f6e<\/td>\n<\/tr>\n<tr>\n<td>paraphrase-multilingual-MiniLM-L12-v2<\/td>\n<td>384<\/td>\n<td>\u591a\u8bed\u8a00\uff0c\u4f53\u79ef\u5c0f\uff08500MB\uff09<\/td>\n<td>\u4e2d\u6587\u7a0d\u5f31<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><strong>\u5546\u4e1aAPI<\/strong>\uff1a<br \/>\n&#8211; OpenAI text-embedding-ada-002 \/ text-embedding-3-small<br \/>\n&#8211; \u667a\u8c31Embedding-API\uff08\u56fd\u5185\u8bbf\u95ee\u7a33\u5b9a\uff09<\/p>\n<h3>5.4 \u5411\u91cf\u6570\u636e\u5e93\u9009\u578b<\/h3>\n<table>\n<thead>\n<tr>\n<th>\u6570\u636e\u5e93<\/th>\n<th>\u4f18\u52bf<\/th>\n<th>\u52a3\u52bf<\/th>\n<th>\u9002\u5408\u573a\u666f<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Milvus<\/td>\n<td>\u529f\u80fd\u5168\uff0c\u793e\u533a\u6d3b\u8dc3<\/td>\n<td>\u90e8\u7f72\u590d\u6742<\/td>\n<td>\u5927\u89c4\u6a21\u5411\u91cf\uff08&gt;100\u4e07\uff09<\/td>\n<\/tr>\n<tr>\n<td>Qdrant<\/td>\n<td>\u90e8\u7f72\u7b80\u5355\uff0c\u6027\u80fd\u597d<\/td>\n<td>\u793e\u533a\u5c0f\u4e9b<\/td>\n<td>\u4e2d\u7b49\u89c4\u6a21<\/td>\n<\/tr>\n<tr>\n<td>Chroma<\/td>\n<td>\u6700\u7b80\u5355\uff0cPython\u539f\u751f<\/td>\n<td>\u529f\u80fd\u5c11\uff0c\u4e0d\u9002\u5408\u751f\u4ea7<\/td>\n<td>\u539f\u578b\u9a8c\u8bc1<\/td>\n<\/tr>\n<tr>\n<td>pgvector<\/td>\n<td>\u57fa\u4e8ePostgreSQL\uff0c\u8fd0\u7ef4\u7b80\u5355<\/td>\n<td>\u6027\u80fd\u4e0d\u5982\u4e13\u7528\u5411\u91cfDB<\/td>\n<td>\u5df2\u6709PG\u7684\u9879\u76ee<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<hr \/>\n<h2>\u516d\u3001\u5b9e\u6218\uff1a\u4ece0\u52301\u642d\u5efa\u4f01\u4e1a\u4e91\u76d8AI\u80fd\u529b<\/h2>\n<h3>6.1 \u67b6\u6784\u8bbe\u8ba1<\/h3>\n<pre><code>\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510     \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510     \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\r\n\u2502 \u6587\u4ef6\u4e0a\u4f20    \u2502\u2500\u2500\u2500\u2500\u25b6\u2502 \u6d88\u606f\u961f\u5217     \u2502\u2500\u2500\u2500\u2500\u25b6\u2502 AI\u5904\u7406\u670d\u52a1   \u2502\r\n\u2502 (\u89e6\u53d1\u4e8b\u4ef6)  \u2502     \u2502 (Kafka\/Redis)\u2502     \u2502 (\u5f02\u6b65\u5904\u7406)   \u2502\r\n\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518     \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518     \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\r\n                                                \u2502\r\n                        \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\r\n                        \u2502                       \u2502                       \u2502\r\n                        \u25bc                       \u25bc                       \u25bc\r\n                 \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510     \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510     \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\r\n                 \u2502 \u667a\u80fd\u6807\u7b7e\u670d\u52a1 \u2502     \u2502 \u81ea\u52a8\u5206\u7c7b\u670d\u52a1 \u2502     \u2502 \u6458\u8981\u751f\u6210\u670d\u52a1 \u2502\r\n                 \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518     \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518     \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\r\n                        \u2502                       \u2502                       \u2502\r\n                        \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\r\n                                                \u25bc\r\n                                         \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\r\n                                         \u2502 \u5411\u91cf\u6570\u636e\u5e93    \u2502\r\n                                         \u2502 (\u53ef\u9009\uff0c\u7528\u4e8e   \u2502\r\n                                         \u2502  \u667a\u80fd\u641c\u7d22)    \u2502\r\n                                         \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\r\n<\/code><\/pre>\n<h3>6.2 \u6587\u4ef6\u5904\u7406\u6d41\u6c34\u7ebf<\/h3>\n<pre><code class=\"language-python\">import asyncio\r\nfrom aiohttp import web\r\n\r\nasync def process_file(request):\r\n    &quot;&quot;&quot;\u6587\u4ef6\u4e0a\u4f20\u540e\uff0c\u89e6\u53d1AI\u5904\u7406\u6d41\u6c34\u7ebf&quot;&quot;&quot;\r\n    data = await request.json()\r\n    file_id = data['file_id']\r\n\r\n    # \u5f02\u6b65\u89e6\u53d1AI\u5904\u7406\uff0c\u4e0d\u963b\u585e\u4e0a\u4f20\u54cd\u5e94\r\n    asyncio.create_task(ai_pipeline(file_id))\r\n\r\n    return web.json_response({'status': 'queued'})\r\n\r\nasync def ai_pipeline(file_id: str):\r\n    &quot;&quot;&quot;AI\u5904\u7406\u6d41\u6c34\u7ebf\uff1a\u6807\u7b7e\u2192\u5206\u7c7b\u2192\u6458\u8981\u2192\u5411\u91cf\u7d22\u5f15&quot;&quot;&quot;\r\n\r\n    file = await get_file(file_id)\r\n\r\n    # Step 1: \u63d0\u53d6\u6587\u672c\u5185\u5bb9\uff08\u6587\u4ef6\u89e3\u6790\uff09\r\n    content = await extract_text(file)\r\n\r\n    # Step 2: \u667a\u80fd\u6807\u7b7e\uff08NER+\u89c4\u5219\uff09\r\n    tags = await generate_smart_tags(content, file)\r\n    await update_file_tags(file_id, tags)\r\n\r\n    # Step 3: \u81ea\u52a8\u5206\u7c7b\r\n    category = await classify_file(content, file)\r\n    await update_file_category(file_id, category)\r\n\r\n    # Step 4: \u751f\u6210\u6458\u8981\r\n    if file['type'] in ['pdf', 'docx', 'txt']:\r\n        summary = await generate_summary(content)\r\n        await update_file_summary(file_id, summary)\r\n\r\n    # Step 5: \u5411\u91cf\u7d22\u5f15\uff08RAG\u573a\u666f\u9700\u8981\uff09\r\n    if use_rag_search:\r\n        chunks = split_into_chunks(content)\r\n        vectors = embedder.encode(chunks)\r\n        await vector_db.insert_batch(file_id, chunks, vectors)\r\n\r\n    await mark_processing_complete(file_id)\r\n\r\nasync def extract_text(file: dict) -&gt; str:\r\n    &quot;&quot;&quot;\u6839\u636e\u6587\u4ef6\u7c7b\u578b\u63d0\u53d6\u6587\u672c\u5185\u5bb9&quot;&quot;&quot;\r\n\r\n    if file['type'] == 'pdf':\r\n        return await extract_pdf_text(file['path'])\r\n    elif file['type'] == 'docx':\r\n        return await extract_docx_text(file['path'])\r\n    elif file['type'] == 'txt':\r\n        return await read_txt(file['path'])\r\n    elif file['type'] == 'image':\r\n        return await ocr_image(file['path'])\r\n    else:\r\n        return &quot;&quot;\r\n<\/code><\/pre>\n<h3>6.3 \u5f02\u6b65\u4efb\u52a1\u961f\u5217\u8bbe\u8ba1<\/h3>\n<pre><code class=\"language-python\">from redis import Redis\r\nimport json\r\nimport threading\r\n\r\nredis = Redis(host='localhost', port=6379, db=0)\r\n\r\ndef enqueue_ai_task(file_id: str, task_type: str, priority: int = 5):\r\n    &quot;&quot;&quot;\u5c06AI\u4efb\u52a1\u52a0\u5165\u961f\u5217\uff08Redis Stream\u5b9e\u73b0\u4f18\u5148\u7ea7\u961f\u5217\uff09&quot;&quot;&quot;\r\n    task = {\r\n        'file_id': file_id,\r\n        'task_type': task_type,  # 'tag'\/'classify'\/'summary'\/'index'\r\n        'priority': priority,\r\n        'enqueued_at': time.time()\r\n    }\r\n\r\n    redis.xadd(\r\n        'ai_tasks',          # Stream key\r\n        task,\r\n        maxlen=100000,        # \u9632\u6b62\u961f\u5217\u65e0\u9650\u589e\u957f\r\n        approximate=True\r\n    )\r\n\r\ndef worker_loop():\r\n    &quot;&quot;&quot;AI\u5904\u7406Worker\uff0c\u6301\u7eed\u4ece\u961f\u5217\u6d88\u8d39\u4efb\u52a1&quot;&quot;&quot;\r\n    while True:\r\n        # \u963b\u585e\u8bfb\u53d6\uff0c\u4f18\u5148\u5904\u7406\u9ad8\u4f18\u5148\u7ea7\u4efb\u52a1\r\n        tasks = redis.xreadgroup(\r\n            'ai_workers', f'worker_{os.getpid()}',\r\n            {'ai_tasks': '&gt;'},\r\n            count=1,\r\n            block=5000  # 5\u79d2\u8d85\u65f6\r\n        )\r\n\r\n        for stream, entries in tasks:\r\n            for entry_id, task_data in entries:\r\n                try:\r\n                    task_type = task_data['task_type']\r\n                    file_id = task_data['file_id']\r\n\r\n                    if task_type == 'tag':\r\n                        generate_tags_for_file(file_id)\r\n                    elif task_type == 'classify':\r\n                        classify_file_task(file_id)\r\n                    elif task_type == 'summary':\r\n                        generate_summary_task(file_id)\r\n                    elif task_type == 'index':\r\n                        index_vector_task(file_id)\r\n\r\n                    # ACK\u4efb\u52a1\r\n                    redis.xack('ai_tasks', 'ai_workers', entry_id)\r\n\r\n                except Exception as e:\r\n                    # \u5931\u8d25\u91cd\u8bd5\uff0c\u8d85\u8fc73\u6b21\u5219\u653e\u5f03\u5e76\u544a\u8b66\r\n                    handle_task_failure(entry_id, task_data, e)\r\n<\/code><\/pre>\n<hr \/>\n<h2>\u4e03\u3001AI\u80fd\u529b\u8bc4\u4f30\u4e0eROI\u8ba1\u7b97<\/h2>\n<h3>7.1 \u8bc4\u4f30\u6307\u6807\u4f53\u7cfb<\/h3>\n<p>\u8bc4\u4f30\u4f01\u4e1a\u4e91\u76d8AI\u80fd\u529b\uff0c\u9700\u8981\u4ece\u4e24\u4e2a\u7ef4\u5ea6\uff1a<strong>\u51c6\u786e\u6027<\/strong>\u548c<strong>\u4e1a\u52a1\u4ef7\u503c<\/strong>\u3002<\/p>\n<p><strong>\u51c6\u786e\u6027\u6307\u6807\uff1a<\/strong><\/p>\n<table>\n<thead>\n<tr>\n<th>\u6307\u6807<\/th>\n<th>\u8ba1\u7b97\u65b9\u5f0f<\/th>\n<th>\u5408\u683c\u7ebf<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u6807\u7b7e\u51c6\u786e\u7387<\/td>\n<td>AI\u6807\u7b7e\u88ab\u7528\u6237\u91c7\u7eb3\u7684\u6bd4\u4f8b<\/td>\n<td>&gt;75%<\/td>\n<\/tr>\n<tr>\n<td>\u5206\u7c7b\u51c6\u786e\u7387<\/td>\n<td>AI\u5206\u7c7b\u4e0e\u4eba\u5de5\u5206\u7c7b\u4e00\u81f4\u7684\u6bd4\u4f8b<\/td>\n<td>&gt;80%<\/td>\n<\/tr>\n<tr>\n<td>\u6458\u8981\u53ec\u56de\u7387<\/td>\n<td>\u6458\u8981\u8986\u76d6\u539f\u6587\u5173\u952e\u4fe1\u606f\u7684\u6bd4\u4f8b<\/td>\n<td>&gt;70%<\/td>\n<\/tr>\n<tr>\n<td>\u68c0\u7d22\u53ec\u56de\u7387<\/td>\n<td>\u68c0\u7d22\u7ed3\u679c\u4e2d\u76f8\u5173\u6587\u6863\u7684\u6bd4\u4f8b\uff08Recall@K\uff09<\/td>\n<td>&gt;85%<\/td>\n<\/tr>\n<tr>\n<td>\u68c0\u7d22\u7cbe\u5ea6\u7387<\/td>\n<td>\u68c0\u7d22\u7ed3\u679c\u4e2d\u771f\u6b63\u76f8\u5173\u6587\u6863\u7684\u6bd4\u4f8b\uff08Precision@K\uff09<\/td>\n<td>&gt;60%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><strong>\u4e1a\u52a1\u4ef7\u503c\u6307\u6807\uff1a<\/strong><\/p>\n<table>\n<thead>\n<tr>\n<th>\u6307\u6807<\/th>\n<th>\u8ba1\u7b97\u65b9\u5f0f<\/th>\n<th>\u76ee\u6807<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u6587\u4ef6\u53ef\u53d1\u73b0\u7387<\/td>\n<td>\u88ab\u6210\u529f\u68c0\u7d22\u5230\u7684\u6587\u4ef6\/\u603b\u6587\u4ef6\u6570<\/td>\n<td>&gt;95%<\/td>\n<\/tr>\n<tr>\n<td>\u641c\u7d22\u6210\u529f\u7387<\/td>\n<td>\u7528\u6237\u641c\u7d22\u540e\u70b9\u5f00\u7ed3\u679c\u7684\u6bd4\u7387<\/td>\n<td>&gt;60%<\/td>\n<\/tr>\n<tr>\n<td>AI\u8f85\u52a9\u65f6\u95f4\u8282\u7701<\/td>\n<td>\u7528\u6237\u624b\u52a8\u5904\u7406\u2192AI\u81ea\u52a8\u5904\u7406\u8282\u7701\u7684\u65f6\u95f4<\/td>\n<td>&gt;50%<\/td>\n<\/tr>\n<tr>\n<td>\u5185\u5bb9\u5f55\u5165\u6548\u7387<\/td>\n<td>\u6807\u7b7e\/\u5206\u7c7b\u7684\u81ea\u52a8\u5316\u6bd4\u4f8b<\/td>\n<td>&gt;60%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>7.2 ROI\u8ba1\u7b97\u6a21\u578b<\/h3>\n<pre><code class=\"language-python\">def calculate_ai_roi():\r\n    &quot;&quot;&quot;\u8ba1\u7b97AI\u529f\u80fd\u7684ROI&quot;&quot;&quot;\r\n\r\n    # \u6210\u672c\u4f30\u7b97\uff08\u5e74\uff09\r\n    ai_cost = {\r\n        'model_api': 50000,      # LLM API\u8c03\u7528\u8d39\u7528\r\n        'embedding_api': 20000,  # Embedding API\u8d39\u7528\r\n        'compute': 30000,        # GPU\/\u670d\u52a1\u5668\u6210\u672c\r\n        'maintenance': 20000,    # \u8fd0\u7ef4\u4eba\u529b\u6210\u672c\r\n    }\r\n    total_cost = sum(ai_cost.values())\r\n\r\n    # \u6536\u76ca\u4f30\u7b97\uff08\u5e74\uff09\r\n    employee_count = 200\r\n    avg_salary = 200000\r\n    avg_minutes_per_day_searching = 15  # \u5458\u5de5\u6bcf\u5929\u82b115\u5206\u949f\u627e\u6587\u4ef6\r\n\r\n    # AI\u63d0\u6548\uff1a\u641c\u7d22\u65f6\u95f4\u51cf\u5c1170%\r\n    time_saved_per_day = avg_minutes_per_day_searching * 0.7 * employee_count\r\n    hours_per_year = time_saved_per_day * 250 \/ 60  # 250\u5de5\u4f5c\u65e5\r\n    salary_per_hour = avg_salary \/ 2000  # \u5e74\u85aa\u6362\u7b97\u65f6\u85aa\r\n    annual_savings = hours_per_year * salary_per_hour\r\n\r\n    # AI\u63d0\u6548\uff1a\u6807\u7b7e\u5206\u7c7b\u65f6\u95f4\u51cf\u5c1190%\r\n    tagging_hours_per_week = 5 * employee_count  # \u5047\u8bbe\u6bcf\u54685\u5c0f\u65f6\u6253\u6807\u7b7e\r\n    tagging_savings = (tagging_hours_per_week * 50 \/ 60) * salary_per_hour  # 50\u5468\r\n\r\n    total_benefits = annual_savings + tagging_savings\r\n\r\n    roi = (total_benefits - total_cost) \/ total_cost * 100\r\n\r\n    return {\r\n        'annual_cost': total_cost,\r\n        'annual_benefits': total_benefits,\r\n        'roi_percent': round(roi, 1),\r\n        'payback_months': round(total_cost \/ (total_benefits \/ 12))\r\n    }\r\n<\/code><\/pre>\n<hr \/>\n<h2>\u516b\u3001\u8e29\u5751\u5b9e\u5f55\u4e0e\u907f\u5751\u6307\u5357<\/h2>\n<h3>\u8e29\u57511\uff1a\u5411\u91cf\u7d22\u5f15\u540e\u6587\u4ef6\u65e0\u6cd5\u5220\u9664<\/h3>\n<p><strong>\u95ee\u9898<\/strong>\uff1a\u6587\u4ef6\u5220\u9664\u540e\uff0c\u5411\u91cf\u6570\u636e\u5e93\u4e2d\u7684embeddings\u6ca1\u6709\u88ab\u6e05\u7406\uff0c\u9020\u6210\u6570\u636e\u4e0d\u4e00\u81f4\u548c\u5b58\u50a8\u6d6a\u8d39\u3002<\/p>\n<p><strong>\u89e3\u51b3<\/strong>\uff1a<\/p>\n<pre><code class=\"language-python\"># \u5220\u9664\u6587\u4ef6\u65f6\uff0c\u540c\u6b65\u6e05\u7406\u5411\u91cf\r\nasync def delete_file(file_id: str):\r\n    # 1. \u5220\u9664\u539f\u59cb\u6587\u4ef6\r\n    await file_store.delete(file_id)\r\n\r\n    # 2. \u5220\u9664\u5411\u91cf\u7d22\u5f15\uff08\u5173\u952e\uff01\uff09\r\n    await vector_db.delete_by_filter({'file_id': file_id})\r\n\r\n    # 3. \u5220\u9664\u5143\u6570\u636e\r\n    await metadata_store.delete(file_id)\r\n<\/code><\/pre>\n<h3>\u8e29\u57512\uff1a\u957f\u6587\u6863\u6458\u8981\u4e22\u5931\u5173\u952e\u6570\u636e<\/h3>\n<p><strong>\u95ee\u9898<\/strong>\uff1a\u4f7f\u7528LLM\u6458\u8981\u65f6\uff0c\u6a21\u578b\u7ecf\u5e38\u9057\u6f0f\u5173\u952e\u6570\u5b57\uff08\u5982&#8221;\u589e\u957f23%&#8221;\u53d8\u6210&#8221;\u663e\u8457\u589e\u957f&#8221;\uff09\u3002<\/p>\n<p><strong>\u89e3\u51b3<\/strong>\uff1a\u63d0\u793a\u8bcd\u5de5\u7a0b + \u540e\u5904\u7406\u9a8c\u8bc1<\/p>\n<pre><code class=\"language-python\">prompt = &quot;&quot;&quot;\u8bf7\u4e3a\u4ee5\u4e0b\u6587\u6863\u751f\u6210\u6458\u8981\u3002\r\n\r\n\u3010\u91cd\u8981\u8981\u6c42\u3011\r\n1. \u5fc5\u987b\u4fdd\u7559\u6240\u6709\u5173\u952e\u6570\u5b57\u3001\u6570\u636e\u3001\u767e\u5206\u6bd4\r\n2. \u5fc5\u987b\u4fdd\u7559\u5177\u4f53\u7684\u4eba\u540d\u3001\u5730\u540d\u3001\u65e5\u671f\r\n3. \u6458\u8981\u4e0d\u8d85\u8fc7300\u5b57\r\n\r\n\u6587\u6863\u5185\u5bb9\uff1a\r\n{content}\r\n\r\n\u6458\u8981\uff1a&quot;&quot;&quot;\r\n<\/code><\/pre>\n<h3>\u8e29\u57513\uff1aOCR\u8bc6\u522b\u8d28\u91cf\u4e0d\u7a33\u5b9a<\/h3>\n<p><strong>\u95ee\u9898<\/strong>\uff1a\u540c\u4e00\u6279\u626b\u63cf\u4ef6\uff0c\u6709\u4e9b\u8bc6\u522b\u7387\u9ad8\uff0c\u6709\u4e9b\u6781\u4f4e\uff08\u56fe\u7247\u6e05\u6670\u5ea6\/\u5149\u7167\u5dee\u5f02\u5bfc\u81f4\uff09\u3002<\/p>\n<p><strong>\u89e3\u51b3<\/strong>\uff1a\u56fe\u7247\u9884\u5904\u7406 + \u591a\u5f15\u64ce\u5197\u4f59<\/p>\n<pre><code class=\"language-python\">def preprocess_image(img_path: str) -&gt; Image:\r\n    &quot;&quot;&quot;\u56fe\u7247\u9884\u5904\u7406\uff1a\u53bb\u566a\u3001\u4e8c\u503c\u5316\u3001\u503e\u659c\u6821\u6b63&quot;&quot;&quot;\r\n    from PIL import ImageFilter, ImageEnhance\r\n\r\n    img = Image.open(img_path)\r\n\r\n    # \u53bb\u566a\r\n    img = img.filter(ImageFilter.MedianFilter(size=3))\r\n\r\n    # \u5bf9\u6bd4\u5ea6\u589e\u5f3a\r\n    enhancer = ImageEnhance.Contrast(img)\r\n    img = enhancer.enhance(1.5)\r\n\r\n    return img\r\n\r\ndef ocr_with_fallback(img_path: str) -&gt; str:\r\n    &quot;&quot;&quot;OCR\u5197\u4f59\uff1a\u767e\u5ea6\u4f18\u5148\uff0cTesseract\u515c\u5e95&quot;&quot;&quot;\r\n    try:\r\n        result = baidu_ocr(img_path)  # \u8bc6\u522b\u7387\u9ad8\r\n        if result.confidence &gt; 0.8:\r\n            return result.text\r\n    except:\r\n        pass\r\n\r\n    # Tesseract\u515c\u5e95\r\n    return pytesseract.image_to_string(preprocess_image(img_path))\r\n<\/code><\/pre>\n<h3>\u8e29\u57514\uff1aAI\u6807\u7b7e\u53cd\u800c\u589e\u52a0\u7528\u6237\u8d1f\u62c5<\/h3>\n<p><strong>\u95ee\u9898<\/strong>\uff1aAI\u81ea\u52a8\u6253\u4e86\u5927\u91cf\u6807\u7b7e\uff0c\u7528\u6237\u9700\u8981\u82b1\u65f6\u95f4\u6e05\u7406\uff0c\u53cd\u800c\u589e\u52a0\u8d1f\u62c5\u3002<\/p>\n<p><strong>\u89e3\u51b3<\/strong>\uff1a\u8bbe\u8ba1&#8221;\u6c89\u9ed8\u91c7\u7eb3&#8221;\u673a\u5236<br \/>\n&#8211; \u9ad8\u7f6e\u4fe1\u5ea6\u6807\u7b7e\uff08&gt;85%\uff09\u81ea\u52a8\u751f\u6548\uff0c\u7528\u6237\u65e0\u611f\u77e5<br \/>\n&#8211; \u4e2d\u7f6e\u4fe1\u5ea6\u6807\u7b7e\uff0860-85%\uff09\u4ec5\u5728\u641c\u7d22\u5efa\u8bae\u4e2d\u5c55\u793a\uff0c\u4e0d\u5f3a\u52a0\u7ed9\u7528\u6237<br \/>\n&#8211; \u4f4e\u7f6e\u4fe1\u5ea6\u6807\u7b7e\uff08&lt;60%\uff09\u4ec5\u7528\u4e8e\u63a8\u8350\uff0c\u7528\u6237\u4e0d\u53ef\u89c1<\/p>\n<hr \/>\n<h2>\u4e5d\u3001\u7ed3\u8bed<\/h2>\n<p>\u4f01\u4e1a\u4e91\u76d8\u7684AI\u80fd\u529b\uff0c\u4e0d\u662f&#8221;\u52a0\u4e00\u4e2a\u804a\u5929\u673a\u5668\u4eba&#8221;\u90a3\u4e48\u7b80\u5355\u3002\u771f\u6b63\u7684\u4ef7\u503c\u6765\u81ea\u4e8e\u628aAI\u6df1\u5ea6\u5d4c\u5165\u5230\u6587\u4ef6\u7ba1\u7406\u7684\u6bcf\u4e00\u4e2a\u73af\u8282\uff1a<\/p>\n<ul>\n<li>\u4e0a\u4f20\u65f6\u81ea\u52a8\u6253\u6807\u7b7e\u2192\u51cf\u5c11\u4eba\u5de5\u5f55\u5165<\/li>\n<li>\u68c0\u7d22\u65f6\u7406\u89e3\u8bed\u4e49\u2192\u627e\u5230\u771f\u6b63\u60f3\u8981\u7684<\/li>\n<li>\u9605\u8bfb\u65f6\u751f\u6210\u6458\u8981\u2192\u5feb\u901f\u4e86\u89e3\u5185\u5bb9<\/li>\n<li>\u6574\u7406\u65f6\u81ea\u52a8\u5206\u7c7b\u2192\u6587\u4ef6\u4e0d\u518d\u4e71\u7cdf\u7cdf<\/li>\n<\/ul>\n<p>\u6bcf\u4e00\u6b65\u90fd\u8ba9\u7528\u6237\u4ece&#8221;\u627e\u6587\u4ef6&#8221;\u8fd9\u79cd\u4f4e\u4ef7\u503c\u52b3\u52a8\u4e2d\u89e3\u653e\u51fa\u6765\uff0c\u4e13\u6ce8\u4e8e\u771f\u6b63\u91cd\u8981\u7684\u5de5\u4f5c\u3002<\/p>\n<p>\u5df4\u522b\u9e1f\u4f01\u4e1a\u4e91\u76d8\u6b63\u5728\u63a8\u8fdbAI\u80fd\u529b\u7684\u6df1\u5ea6\u96c6\u6210\uff0c\u667a\u80fd\u6807\u7b7e\u548c\u8bed\u4e49\u641c\u7d22\u529f\u80fd\u5df2\u5728\u6d4b\u8bd5\u4e2d\u3002\u5982\u679c\u4f60\u5e0c\u671b\u7b2c\u4e00\u65f6\u95f4\u4f53\u9a8c\uff0c\u6216\u5bf9AI\u80fd\u529b\u6709\u5177\u4f53\u7684\u9700\u6c42\u53cd\u9988\uff0c\u6b22\u8fce\u8054\u7cfb\u6211\u4eec\u3002<\/p>\n<hr \/>\n<p><strong>\u5173\u4e8e\u5df4\u522b\u9e1f<\/strong><\/p>\n<p>\u5df4\u522b\u9e1f\u4f01\u4e1a\u4e91\u76d8\u4e13\u6ce8\u4e8e\u4f01\u4e1a\u6587\u4ef6\u7ba1\u7406\u9886\u57df\u591a\u5e74\uff0c\u6b63\u5728\u5c06AI\u80fd\u529b\u6df1\u5ea6\u878d\u5165\u4ea7\u54c1\u6838\u5fc3\u573a\u666f\u3002\u5982\u9700\u4e86\u89e3\u66f4\u591a\u4ea7\u54c1\u4fe1\u606f\uff0c\u8bf7\u8bbf\u95ee\uff1a<a href=\"https:\/\/www.babel.cc\/blog\/\">https:\/\/www.babel.cc\/blog\/<\/a><\/p>\n<hr \/>\n<p>\u5b57\u6570\uff1a\u7ea619500\u5b57<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u524d\u8a00 \u4f01\u4e1a\u4e91\u76d8\u5728\u7ecf\u5386\u4e86&#8221;\u5b58\u50a8\u65f6\u4ee3&#8221;\u548c&#8230; <a class=\"read-more\" href=\"https:\/\/www.babel.cc\/blog\/?p=6918\">[Continue Reading]<\/a><\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-6918","post","type-post","status-publish","format-standard","hentry","category-1","no-post-thumbnail"],"_links":{"self":[{"href":"https:\/\/www.babel.cc\/blog\/index.php?rest_route=\/wp\/v2\/posts\/6918","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.babel.cc\/blog\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.babel.cc\/blog\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.babel.cc\/blog\/index.php?rest_route=\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.babel.cc\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=6918"}],"version-history":[{"count":1,"href":"https:\/\/www.babel.cc\/blog\/index.php?rest_route=\/wp\/v2\/posts\/6918\/revisions"}],"predecessor-version":[{"id":6919,"href":"https:\/\/www.babel.cc\/blog\/index.php?rest_route=\/wp\/v2\/posts\/6918\/revisions\/6919"}],"wp:attachment":[{"href":"https:\/\/www.babel.cc\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=6918"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.babel.cc\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=6918"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.babel.cc\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=6918"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}