{"id":1406,"date":"2024-05-09T21:45:55","date_gmt":"2024-05-09T21:45:55","guid":{"rendered":"https:\/\/www.nicekj.com\/?p=1406"},"modified":"2024-05-09T21:47:35","modified_gmt":"2024-05-09T21:47:35","slug":"yizhongjiyuhuadongchuangkoukuozhanshangxiawenderagjiansuozengqiangshengchengyouhua","status":"publish","type":"post","link":"https:\/\/www.nicekj.com\/yizhongjiyuhuadongchuangkoukuozhanshangxiawenderagjiansuozengqiangshengchengyouhua.html","title":{"rendered":"\u4e00\u79cd\u57fa\u4e8e\u6ed1\u52a8\u7a97\u53e3\u6269\u5c55\u4e0a\u4e0b\u6587\u7684RAG\uff08\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff09\u4f18\u5316\u5b9e\u73b0\u65b9\u6848\u5b9e\u8df5"},"content":{"rendered":"

\u672c\u6587\u9996\u53d1\u4e8e\u535a\u5ba2LLM \u5e94\u7528\u5f00\u53d1\u5b9e\u8df5<\/a><\/p>\n

\n

RAG\uff08\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff09\u662f\u4e00\u79cd\u7ed3\u5408\u4e86\u68c0\u7d22\uff08\u901a\u5e38\u662f\u77e5\u8bc6\u5e93\u6216\u6570\u636e\u5e93\uff09\u548c\u751f\u6210\u6a21\u578b\uff08\u5927\u8bed\u8a00\u6a21\u578b\uff09\u7684\u6280\u672f\uff0c\u76ee\u7684\u662f\u5728\u751f\u6210\u6587\u672c\u7684\u65f6\u5019\u80fd\u591f\u53c2\u8003\u76f8\u5173\u7684\u5916\u90e8\u77e5\u8bc6\u3002\u8fd9\u6837\uff0c\u5373\u4f7f\u751f\u6210\u6a21\u578b\u5728\u8bad\u7ec3\u65f6\u6ca1\u6709\u770b\u5230\u67d0\u4e9b\u4fe1\u606f\uff0c\u5b83\u4e5f\u80fd\u5728\u751f\u6210\u65f6\u901a\u8fc7\u68c0\u7d22\u5230\u7684\u77e5\u8bc6\u6765\u751f\u6210\u66f4\u52a0\u51c6\u786e\u548c\u4e30\u5bcc\u7684\u56de\u7b54\uff0c\u8fd9\u7bc7\u6587\u7ae0\u5b9e\u73b0\u4e00\u79cd\u57fa\u4e8e\u52a8\u6001\u4e0a\u4e0b\u6587\u7a97\u53e3\u7684\u65b9\u6848\uff0c\u80fd\u591f\u5904\u7406\u5927\u89c4\u6a21\u6587\u6863\uff0c\u4fdd\u7559\u91cd\u8981\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u63d0\u5347\u68c0\u7d22\u6548\u7387\uff0c\u540c\u65f6\u4fdd\u6301\u7075\u6d3b\u6027\u548c\u53ef\u914d\u7f6e\u6027\u3002<\/p>\n<\/blockquote>\n

\u5148\u770b\u6548\u679c<\/h3>\n

\n

\"\u6f14\u793a\u6548\u679c\"<\/figure>\n<\/p>\n

\u6f14\u793a\u6548\u679c<\/p>\n

\u6574\u4f53\u65b9\u6848<\/h3>\n

\u6574\u4f53\u65b9\u6848\u5728\u4e8e\u6587\u6863\u9884\u5904\u7406\u9636\u6bb5\u5b9e\u73b0\u6ee1\u8db3\u4e0a\u4e0b\u6587\u7a97\u53e3\u7684\u539f\u59cb\u6587\u672c\u5206\u5757\uff0c\u6587\u6863\u68c0\u7d22\u9636\u6bb5\u5b9e\u73b0\u6587\u672c\u7684\u4e09\u6b21\u68c0\u7d22\uff0c\u4e0b\u9762\u9010\u4e00\u8fdb\u884c\u8bf4\u660e\u3002\u6d4b\u8bd5\u6587\u7ae0\u6765\u81ea \u5927\u8bed\u8a00\u6a21\u578b\u7684\u5b89\u5168\u95ee\u9898\u63a2\u7a76-\u63d0\u793a\u8bcd\u653b\u51fb<\/a><\/p>\n

\u6587\u6863\u9884\u5904\u7406\u8fc7\u7a0b<\/h3>\n

\u5c0f\u6587\u672c\u5757\u62c6\u5206<\/h4>\n

\u4ee5 50 token<\/code>\u5927\u5c0f\uff08\u53ef\u6839\u636e\u81ea\u8eab\u6587\u6863\u4e4b\u95f4\u7684\u7ec4\u7ec7\u89c4\u5f8b\u52a8\u6001\u8c03\u6574\u7c92\u5ea6\uff09\u5bf9\u6587\u672c\u505a\u9996\u6b21\u5206\u5272<\/p>\n

<\/div>
ini<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div># \u5c0f\u6587\u6863\u5757\u5927\u5c0f<\/span><\/span>\nBASE_CHUNK_SIZE<\/span> = 50<\/span><\/span>\n# \u5c0f\u5757\u7684\u91cd\u53e0\u90e8\u5206\u5927\u5c0f<\/span><\/span>\nCHUNK_OVERLAP<\/span> = 0<\/span><\/span>\ndef split_doc(<\/span>\n    doc: List[Document]<\/span>, chunk_size<\/span>=BASE_CHUNK_SIZE, chunk_overlap=CHUNK_OVERLAP, chunk_idx_name: str<\/span>\n):<\/span>\n    data_splitter<\/span> = RecursiveCharacterTextSplitter(<\/span>\n        chunk_size<\/span>=chunk_size,<\/span>\n        chunk_overlap<\/span>=chunk_overlap,<\/span>\n        # \u4f7f\u7528\u4e86 tiktoken \u6765\u786e\u4fdd\u5206\u5272\u4e0d\u4f1a\u5728\u4e00\u4e2a token \u7684\u4e2d\u95f4\u53d1\u751f<\/span><\/span>\n        length_function<\/span>=tiktoken_len,<\/span>\n    )<\/span>\n    doc_split<\/span> = data_splitter.split_documents(doc)<\/span>\n    chunk_idx<\/span> = 0<\/span><\/span>\n    for d_split in doc_split:<\/span>\n        d_split.metadata[chunk_idx_name]<\/span> = chunk_idx<\/span>\n        chunk_idx += 1<\/span>\n    return doc_split<\/span>\n<\/code><\/pre>\n
\u4e0b\u9762\u793a\u4f8b\u663e\u793a\u4e86\u524d 7 \u4e2a\u5206\u5757\u4fe1\u606f\uff0c\u7ed3\u679c\u5982\u4e0b\uff1a<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>[Document(page_content='LLM \u5b89\u5168\u4e13\u9898 \u63d0\u793a\u8bcd'<\/span>, metadata={'source'<\/span>: '.\/data\/\u4e00\u6587\u5e26\u4f60\u4e86\u89e3\u63d0\u793a\u653b\u51fb.pdf'<\/span>, 'page'<\/span>: 0, 'small_chunk_idx'<\/span>: 0}),Document(page_content='\u662f\u6307\u5728\u8bad\u7ec3\u6216\u4e0e\u2f24\u578b\u8bed\u2f94\u6a21\u578b\uff08Claude\uff0cChatGPT\u7b49\uff09\u8fdb\u2f8f\u4ea4\u4e92\u65f6\uff0c\u63d0\u4f9b\u7ed9\u6a21'<\/span>, metadata={'source'<\/span>: '.\/data\/\u4e00\u6587\u5e26\u4f60\u4e86\u89e3\u63d0\u793a\u653b\u51fb.pdf'<\/span>, 'page'<\/span>: 0, 'small_chunk_idx'<\/span>: 1}),Document(page_content='\u578b\u7684\u8f93\u2f0a\u2f42\u672c\u3002\u901a\u8fc7\u7ed9\u5b9a\u7279\u5b9a\u7684'<\/span>, metadata={'source'<\/span>: '.\/data\/\u4e00\u6587\u5e26\u4f60\u4e86\u89e3\u63d0\u793a\u653b\u51fb.pdf'<\/span>, 'page'<\/span>: 0, 'small_chunk_idx'<\/span>: 2}),Document(page_content='\u63d0\u793a\u8bcd\uff0c\u53ef\u4ee5\u5f15\u5bfc\u6a21\u578b\u2f63\u6210\u7279\u5b9a\u4e3b\u9898\u6216\u7c7b\u578b\u7684\u2f42\u672c\u3002\u5728\u2f83\u7136\u8bed\u2f94\u5904\u7406\uff08NLP\uff09\u4efb\u52a1\u4e2d\uff0c\u63d0'<\/span>, metadata={'source'<\/span>: '.\/data\/\u4e00\u6587\u5e26\u4f60\u4e86\u89e3\u63d0\u793a\u653b\u51fb.pdf'<\/span>, 'page'<\/span>: 0, 'small_chunk_idx'<\/span>: 3}),Document(page_content='\u793a\u8bcd\u5145\u5f53\u4e86\u95ee\u9898\u6216\u8f93\u2f0a\u7684\u2ec6\u2f8a\uff0c\u2f7d\u6a21\u578b\u7684\u8f93\u51fa\u662f\u5bf9\u8fd9\u4e2a\u95ee\u9898\u7684\u56de\u7b54\u6216\u5b8c\u6210\u7684\u4efb\u52a1\u3002\u5173\u4e8e'<\/span>, metadata={'source'<\/span>: '.\/data\/\u4e00\u6587\u5e26\u4f60\u4e86\u89e3\u63d0\u793a\u653b\u51fb.pdf'<\/span>, 'page'<\/span>: 0, 'small_chunk_idx'<\/span>: 4}),Document(page_content='\u600e\u6837\u8bbe\u8ba1\u597d\u7684'<\/span>, metadata={'source'<\/span>: '.\/data\/\u4e00\u6587\u5e26\u4f60\u4e86\u89e3\u63d0\u793a\u653b\u51fb.pdf'<\/span>, 'page'<\/span>: 0, 'small_chunk_idx'<\/span>: 5}),Document(page_content='Prompt\uff0c\u67e5\u770bPrompt\u4e13\u9898\u7ae0\u8282\u5185\u5bb9\u5c31\u53ef\u4ee5\u4e86\uff0c\u6211\u4e0d\u5728\u8fd9\u2fa5\u8fc7\u591a\u9610\u8ff0\uff0c\u4e2a\u2f08\u2f50\u8f83\u611f\u5174\u8da3\u9488\u5bf9'<\/span>, metadata={'source'<\/span>: '.\/data\/\u4e00\u6587\u5e26\u4f60\u4e86\u89e3\u63d0\u793a\u653b\u51fb.pdf'<\/span>, 'page'<\/span>: 0, 'small_chunk_idx'<\/span>: 6}),...]<\/span><\/span>\n<\/code><\/pre>\n
\u6dfb\u52a0\u7a97\u53e3<\/h4>\n
\u4ee5\u6b65\u957f\u4e3a 3\uff0c\u7a97\u53e3\u5927\u5c0f\u4e3a 6\uff0c\u5c06\u4e0a\u8ff0\u6b65\u9aa4\u7684\u5c0f\u5757\u5339\u914d\u5230\u4e0d\u540c\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u3002<\/p>\n
<\/div>
ini<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div># \u6b65\u957f\u5b9a\u4e49\u4e86\u7a97\u53e3\u79fb\u52a8\u7684\u901f\u5ea6\uff0c\u5177\u4f53\u6765\u8bf4\uff0c\u5b83\u662f\u4e0a\u4e00\u4e2a\u7a97\u53e3\u4e2d\u7b2c\u4e00\u4e2a\u5757\u548c\u4e0b\u4e00\u4e2a\u7a97\u53e3\u4e2d\u7b2c\u4e00\u4e2a\u5757\u4e4b\u95f4\u7684\u8ddd\u79bb<\/span><\/span>\nWINDOW_STEPS<\/span> = 3<\/span><\/span>\n# \u7a97\u53e3\u5927\u5c0f\u76f4\u63a5\u5f71\u54cd\u5230\u6bcf\u4e2a\u7a97\u53e3\u4e2d\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\u91cf\uff0c\u7a97\u53e3\u5927\u5c0f= BASE_CHUNK_SIZE * WINDOW_SCALE<\/span><\/span>\nWINDOW_SCALE<\/span> = 6<\/span><\/span>\ndef add_window(<\/span>\n    doc: Document, window_steps<\/span>=WINDOW_STEPS, window_size=WINDOW_SCALE, window_idx_name: str<\/span>\n):<\/span>\n    window_id<\/span> = 0<\/span><\/span>\n    window_deque<\/span> = deque()<\/span>\n<\/span>\n    for idx, item in enumerate(doc):<\/span>\n        if idx % window_steps<\/span> == 0<\/span> and idx != 0<\/span> and idx < len(doc) - window_size:<\/span>\n            window_id += 1<\/span>\n        window_deque.append(window_id)<\/span>\n<\/span>\n        if len(window_deque) > window_size:<\/span>\n            for _ in range(window_steps):<\/span>\n                window_deque.popleft()<\/span>\n<\/span>\n        window<\/span> = set(window_deque)<\/span>\n        item.metadata[f\"{window_idx_name}_lower_bound\"]<\/span> = min(window)<\/span>\n        item.metadata[f\"{window_idx_name}_upper_bound\"]<\/span> = max(window)<\/span>\n<\/code><\/pre>\n
\u4e0b\u9762\u793a\u4f8b\u663e\u793a\u4e86\u524d 7 \u4e2a\u589e\u52a0\u7a97\u53e3\u4fe1\u606f\u540e\u7684\u5206\u5757\u5185\u5bb9\uff0c\u7ed3\u679c\u5982\u4e0b<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>[Document(page_content='LLM \u5b89\u5168\u4e13\u9898 \u63d0\u793a\u8bcd'<\/span>, metadata={'source'<\/span>: '.\/data\/\u4e00\u6587\u5e26\u4f60\u4e86\u89e3\u63d0\u793a\u653b\u51fb.pdf'<\/span>, 'page'<\/span>: 0, 'small_chunk_idx'<\/span>: 0, 'large_chunks_idx_lower_bound'<\/span>: 0, 'large_chunks_idx_upper_bound'<\/span>: 0}),Document(page_content='\u662f\u6307\u5728\u8bad\u7ec3\u6216\u4e0e\u2f24\u578b\u8bed\u2f94\u6a21\u578b\uff08Claude\uff0cChatGPT\u7b49\uff09\u8fdb\u2f8f\u4ea4\u4e92\u65f6\uff0c\u63d0\u4f9b\u7ed9\u6a21'<\/span>, metadata={'source'<\/span>: '.\/data\/\u4e00\u6587\u5e26\u4f60\u4e86\u89e3\u63d0\u793a\u653b\u51fb.pdf'<\/span>, 'page'<\/span>: 0, 'small_chunk_idx'<\/span>: 1, 'large_chunks_idx_lower_bound'<\/span>: 0, 'large_chunks_idx_upper_bound'<\/span>: 0}),Document(page_content='\u578b\u7684\u8f93\u2f0a\u2f42\u672c\u3002\u901a\u8fc7\u7ed9\u5b9a\u7279\u5b9a\u7684'<\/span>, metadata={'source'<\/span>: '.\/data\/\u4e00\u6587\u5e26\u4f60\u4e86\u89e3\u63d0\u793a\u653b\u51fb.pdf'<\/span>, 'page'<\/span>: 0, 'small_chunk_idx'<\/span>: 2, 'large_chunks_idx_lower_bound'<\/span>: 0, 'large_chunks_idx_upper_bound'<\/span>: 0}),Document(page_content='\u63d0\u793a\u8bcd\uff0c\u53ef\u4ee5\u5f15\u5bfc\u6a21\u578b\u2f63\u6210\u7279\u5b9a\u4e3b\u9898\u6216\u7c7b\u578b\u7684\u2f42\u672c\u3002\u5728\u2f83\u7136\u8bed\u2f94\u5904\u7406\uff08NLP\uff09\u4efb\u52a1\u4e2d\uff0c\u63d0'<\/span>, metadata={'source'<\/span>: '.\/data\/\u4e00\u6587\u5e26\u4f60\u4e86\u89e3\u63d0\u793a\u653b\u51fb.pdf'<\/span>, 'page'<\/span>: 0, 'small_chunk_idx'<\/span>: 3, 'large_chunks_idx_lower_bound'<\/span>: 0, 'large_chunks_idx_upper_bound'<\/span>: 1}),Document(page_content='\u793a\u8bcd\u5145\u5f53\u4e86\u95ee\u9898\u6216\u8f93\u2f0a\u7684\u2ec6\u2f8a\uff0c\u2f7d\u6a21\u578b\u7684\u8f93\u51fa\u662f\u5bf9\u8fd9\u4e2a\u95ee\u9898\u7684\u56de\u7b54\u6216\u5b8c\u6210\u7684\u4efb\u52a1\u3002\u5173\u4e8e'<\/span>, metadata={'source'<\/span>: '.\/data\/\u4e00\u6587\u5e26\u4f60\u4e86\u89e3\u63d0\u793a\u653b\u51fb.pdf'<\/span>, 'page'<\/span>: 0, 'small_chunk_idx'<\/span>: 4, 'large_chunks_idx_lower_bound'<\/span>: 0, 'large_chunks_idx_upper_bound'<\/span>: 1}),Document(page_content='\u600e\u6837\u8bbe\u8ba1\u597d\u7684'<\/span>, metadata={'source'<\/span>: '.\/data\/\u4e00\u6587\u5e26\u4f60\u4e86\u89e3\u63d0\u793a\u653b\u51fb.pdf'<\/span>, 'page'<\/span>: 0, 'small_chunk_idx'<\/span>: 5, 'large_chunks_idx_lower_bound'<\/span>: 0, 'large_chunks_idx_upper_bound'<\/span>: 1}),Document(page_content='Prompt\uff0c\u67e5\u770bPrompt\u4e13\u9898\u7ae0\u8282\u5185\u5bb9\u5c31\u53ef\u4ee5\u4e86\uff0c\u6211\u4e0d\u5728\u8fd9\u2fa5\u8fc7\u591a\u9610\u8ff0\uff0c\u4e2a\u2f08\u2f50\u8f83\u611f\u5174\u8da3\u9488\u5bf9'<\/span>, metadata={'source'<\/span>: '.\/data\/\u4e00\u6587\u5e26\u4f60\u4e86\u89e3\u63d0\u793a\u653b\u51fb.pdf'<\/span>, 'page'<\/span>: 0, 'small_chunk_idx'<\/span>: 6, 'large_chunks_idx_lower_bound'<\/span>: 1, 'large_chunks_idx_upper_bound'<\/span>: 2}),Document(page_content='Prompt\u7684\u653b\u51fb\uff0c\u968f\u7740\u2f24\u8bed\u2f94\u6a21\u578b\u7684\u2f34\u6cdb\u5e94\u2f64\uff0c\u5b89\u5168\u5fc5\u5b9a\u662f\u2f00\u4e2a\u2fae\u5e38\u503c'<\/span>, metadata={'source'<\/span>: '.\/data\/\u4e00\u6587\u5e26\u4f60\u4e86\u89e3\u63d0\u793a\u653b\u51fb.pdf'<\/span>, 'page'<\/span>: 0, 'small_chunk_idx'<\/span>: 7, 'large_chunks_idx_lower_bound'<\/span>: 1, 'large_chunks_idx_upper_bound'<\/span>: 2}),...]<\/span><\/span>\n<\/code><\/pre>\n
\u4e2d\u7b49\u6587\u672c\u5757<\/h3>\n
\u4ee5\u5c0f\u6587\u672c\u5757 3 \u500d\uff08\u53ef\u52a8\u6001\u914d\u7f6e\uff09\uff0c\u5373150 token<\/code>\u5927\u5c0f\u5bf9\u6587\u672c\u505a\u4e8c\u6b21\u5206\u5272\uff0c\u5f62\u6210\u4e2d\u7b49\u6587\u672c\u5757<\/p>\n
<\/div>
ini<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div># \u4e2d\u7b49\u5927\u5c0f\u7684\u6587\u6863\u5757\u5927\u5c0f=BASE_CHUNK_SIZE * CHUNK_SCALE<\/span><\/span>\nCHUNK_SCALE<\/span> = 3<\/span><\/span>\ndef merge_metadata(dicts_list: dict):<\/span>\n    merged_dict<\/span> = {}<\/span>\n    bounds_dict<\/span> = {}<\/span>\n    keys_to_remove<\/span> = set()<\/span>\n<\/span>\n    for dic in dicts_list:<\/span>\n        for key, value in dic.items():<\/span>\n            if key in merged_dict:<\/span>\n                if value not in merged_dict[key]<\/span>:<\/span>\n                    merged_dict[key]<\/span>.append(value)<\/span>\n            else:<\/span>\n                merged_dict[key]<\/span> = [value]<\/span><\/span>\n<\/span>\n    for key, values in merged_dict.items():<\/span>\n        if len(values) > 1 and all(isinstance(x, (int, float)) for x in values):<\/span>\n            bounds_dict[f\"{key}_lower_bound\"]<\/span> = min(values)<\/span>\n            bounds_dict[f\"{key}_upper_bound\"]<\/span> = max(values)<\/span>\n            keys_to_remove.add(key)<\/span>\n<\/span>\n    merged_dict.update(bounds_dict)<\/span>\n<\/span>\n    for key in keys_to_remove:<\/span>\n        del merged_dict[key]<\/span><\/span>\n<\/span>\n    return {<\/span>\n        k: v[0]<\/span> if isinstance(v, list) and len(v) == 1 else v<\/span>\n        for k, v in merged_dict.items()<\/span>\n    }<\/span>\n<\/span>\ndef merge_chunks(doc: Document, scale_factor<\/span>=CHUNK_SCALE, chunk_idx_name: str):<\/span>\n    merged_doc<\/span> = []<\/span>\n    page_content<\/span> = \"\"<\/span><\/span>\n    metadata_list<\/span> = []<\/span>\n    chunk_idx<\/span> = 0<\/span><\/span>\n<\/span>\n    for idx, item in enumerate(doc):<\/span>\n        page_content += item.page_content<\/span>\n        metadata_list.append(item.metadata)<\/span>\n        if (idx + 1) % scale_factor<\/span> == 0<\/span> or idx == len(doc) - 1<\/span>:<\/span>\n            metadata<\/span> = merge_metadata(metadata_list)<\/span>\n            metadata[chunk_idx_name]<\/span> = chunk_idx<\/span>\n            merged_doc.append(<\/span>\n                Document(<\/span>\n                    page_content<\/span>=page_content,<\/span>\n                    metadata<\/span>=metadata,<\/span>\n                )<\/span>\n            )<\/span>\n            chunk_idx += 1<\/span>\n            page_content<\/span> = \"\"<\/span><\/span>\n            metadata_list<\/span> = []<\/span>\n<\/span>\n    return merged_doc<\/span>\n<\/code><\/pre>\n
\u4e0b\u9762\u793a\u4f8b\u663e\u793a\u4e86\u524d 3 \u4e2a\u4e2d\u7b49\u5206\u5757\u4fe1\u606f\uff0c\u7ed3\u679c\u5982\u4e0b\uff1a<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>[Document(page_content='LLM \u5b89\u5168\u4e13\u9898 \u63d0\u793a\u8bcd\u662f\u6307\u5728\u8bad\u7ec3\u6216\u4e0e\u2f24\u578b\u8bed\u2f94\u6a21\u578b\uff08Claude\uff0cChatGPT\u7b49\uff09\u8fdb\u2f8f\u4ea4\u4e92\u65f6\uff0c\u63d0\u4f9b\u7ed9\u6a21\u578b\u7684\u8f93\u2f0a\u2f42\u672c\u3002\u901a\u8fc7\u7ed9\u5b9a\u7279\u5b9a\u7684'<\/span>, metadata={'source'<\/span>: '.\/data\/\u4e00\u6587\u5e26\u4f60\u4e86\u89e3\u63d0\u793a\u653b\u51fb.pdf'<\/span>, 'page'<\/span>: 0, 'large_chunks_idx_lower_bound'<\/span>: 0, 'large_chunks_idx_upper_bound'<\/span>: 0, 'small_chunk_idx_lower_bound'<\/span>: 0, 'small_chunk_idx_upper_bound'<\/span>: 2, 'medium_chunk_idx'<\/span>: 0}),Document(page_content='\u63d0\u793a\u8bcd\uff0c\u53ef\u4ee5\u5f15\u5bfc\u6a21\u578b\u2f63\u6210\u7279\u5b9a\u4e3b\u9898\u6216\u7c7b\u578b\u7684\u2f42\u672c\u3002\u5728\u2f83\u7136\u8bed\u2f94\u5904\u7406\uff08NLP\uff09\u4efb\u52a1\u4e2d\uff0c\u63d0\u793a\u8bcd\u5145\u5f53\u4e86\u95ee\u9898\u6216\u8f93\u2f0a\u7684\u2ec6\u2f8a\uff0c\u2f7d\u6a21\u578b\u7684\u8f93\u51fa\u662f\u5bf9\u8fd9\u4e2a\u95ee\u9898\u7684\u56de\u7b54\u6216\u5b8c\u6210\u7684\u4efb\u52a1\u3002\u5173\u4e8e\u600e\u6837\u8bbe\u8ba1\u597d\u7684'<\/span>, metadata={'source'<\/span>: '.\/data\/\u4e00\u6587\u5e26\u4f60\u4e86\u89e3\u63d0\u793a\u653b\u51fb.pdf'<\/span>, 'page'<\/span>: 0, 'large_chunks_idx_lower_bound'<\/span>: 0, 'large_chunks_idx_upper_bound'<\/span>: 1, 'small_chunk_idx_lower_bound'<\/span>: 3, 'small_chunk_idx_upper_bound'<\/span>: 5, 'medium_chunk_idx'<\/span>: 1}),Document(page_content='Prompt\uff0c\u67e5\u770bPrompt\u4e13\u9898\u7ae0\u8282\u5185\u5bb9\u5c31\u53ef\u4ee5\u4e86\uff0c\u6211\u4e0d\u5728\u8fd9\u2fa5\u8fc7\u591a\u9610\u8ff0\uff0c\u4e2a\u2f08\u2f50\u8f83\u611f\u5174\u8da3\u9488\u5bf9Prompt\u7684\u653b\u51fb\uff0c\u968f\u7740\u2f24\u8bed\u2f94\u6a21\u578b\u7684\u2f34\u6cdb\u5e94\u2f64\uff0c\u5b89\u5168\u5fc5\u5b9a\u662f\u2f00\u4e2a\u2fae\u5e38\u503c\u5f97\u5173\u6ce8\u7684\u9886\u57df\u3002\u63d0\u793a\u653b\u51fb'<\/span>, metadata={'source'<\/span>: '.\/data\/\u4e00\u6587\u5e26\u4f60\u4e86\u89e3\u63d0\u793a\u653b\u51fb.pdf'<\/span>, 'page'<\/span>: 0, 'large_chunks_idx_lower_bound'<\/span>: 1, 'large_chunks_idx_upper_bound'<\/span>: 2, 'small_chunk_idx_lower_bound'<\/span>: 6, 'small_chunk_idx_upper_bound'<\/span>: 8, 'medium_chunk_idx'<\/span>: 2}),...]<\/span><\/span>\n<\/code><\/pre>\n
\u6587\u6863\u68c0\u7d22\u8fc7\u7a0b<\/h3>\n
\u68c0\u7d22\u5668\u58f0\u660e<\/h4>\n
\u9996\u5148\u58f0\u660e\u4e00\u4e2a\u68c0\u7d22\u5668\uff0c\u7528\u4e8e\u68c0\u7d22\u6587\u6863\uff0c\u8fd9\u91cc\u5c06 BM25 \u68c0\u7d22\u5668\u548c\u5d4c\u5165\u5f0f\u68c0\u7d22\u5668\u7ec4\u5408\u6210\u4e00\u4e2a\u96c6\u6210\u68c0\u7d22\u5668\uff0c\u7528\u4e8e\u68c0\u7d22\u548c\u8bc4\u4f30\u6587\u6863\u76f8\u4f3c\u5ea6\u3002\u4e0b\u9762\u662f\u4e00\u4e9b\u9700\u8981\u76f8\u5173\u77e5\u8bc6\uff1a<\/p>\n
    \n
  • \n
    BM25 \u662f\u4e00\u79cd\u57fa\u4e8e\u8bcd\u888b\u6a21\u578b\u7684\u68c0\u7d22\u65b9\u6cd5\uff0c\u5b83\u901a\u8fc7\u8003\u8651\u5355\u8bcd\u5728\u6587\u6863\u4e2d\u7684\u9891\u7387\u548c\u5728\u6574\u4e2a\u6587\u6863\u96c6\u5408\u4e2d\u7684\u9006\u6587\u6863\u9891\u7387\u6765\u8ba1\u7b97\u6587\u6863\u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6<\/p>\n<\/li>\n
  • \n
    \u5d4c\u5165\u5f0f\u68c0\u7d22\u5668\u901a\u5e38\u4f7f\u7528\u9884\u8bad\u7ec3\u7684\u5d4c\u5165\u6a21\u578b\uff08\u672c\u6848\u4f8b\u4f7f\u7528 OpenAI \u7684 text-embedding-ada-002 \u6a21\u578b\uff09\u5c06\u6587\u6863\u8f6c\u6362\u4e3a\u5bc6\u96c6\u5411\u91cf\uff0c\u7136\u540e\u901a\u8fc7\u8ba1\u7b97\u8fd9\u4e9b\u5411\u91cf\u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6\u6765\u8bc4\u4f30\u6587\u6863\u4e4b\u95f4\u7684\u76f8\u4f3c\u6027<\/p>\n<\/li>\n
  • \n
    emb_filter: \u7528\u4e8e\u5728\u5d4c\u5165\u5f0f\u68c0\u7d22\u8fc7\u7a0b\u4e2d\u8fc7\u6ee4\u7ed3\u679c\u3002\u4f8b\u5982\uff0c\u53ef\u4ee5\u6839\u636e\u67d0\u4e9b\u6807\u51c6\u6392\u9664\u4e0d\u76f8\u5173\u7684\u6587\u6863<\/p>\n<\/li>\n
  • \n
    k: \u8fd9\u662f\u4e00\u4e2a\u6574\u6570\uff0c\u8868\u793a\u8981\u8fd4\u56de\u7684\u6700\u5339\u914d\u7684\u524d\u51e0\u4e2a\u7ed3\u679c\u6570\u91cf<\/p>\n<\/li>\n
  • \n
    weights: \u5305\u542b\u4e24\u4e2a\u6743\u91cd\u503c\uff0c\u5206\u522b\u7528\u4e8e BM25 \u68c0\u7d22\u5668\u548c\u5d4c\u5165\u5f0f\u68c0\u7d22\u5668\u5728\u96c6\u6210\u68c0\u7d22\u4e2d\u7684\u6743\u91cd\u3002<\/p>\n
    <\/div>
    ini<\/span>
    \u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>def get_retriever(<\/span>\n    self,<\/span>\n    docs_chunks,<\/span>\n    emb_chunks,<\/span>\n    emb_filter<\/span>=None,<\/span>\n    k<\/span>=2<\/span>,<\/span>\n    weights<\/span>=(0.5<\/span>, 0.5<\/span>),<\/span>\n):<\/span>\n    bm25_retriever<\/span> = BM25Retriever.from_documents(docs_chunks)<\/span>\n    bm25_retriever.k<\/span> = k<\/span>\n<\/span>\n    emb_retriever<\/span> = emb_chunks.as_retriever(<\/span>\n        search_kwargs<\/span>={<\/span>\n            \"filter\": emb_filter,<\/span>\n            \"k\": k,<\/span>\n            \"search_type\": \"mmr\",<\/span>\n        }<\/span>\n    )<\/span>\n    return MyEnsembleRetriever(<\/span>\n        retrievers<\/span>={\"bm25\"<\/span>: bm25_retriever, \"chroma\"<\/span>: emb_retriever},<\/span>\n        weights<\/span>=weights,<\/span>\n    )<\/span>\n<\/code><\/pre>\n<\/li>\n<\/ul>\n
    \u68c0\u7d22\u76f8\u5173\u6587\u6863<\/h4>\n
    \u6587\u6863\u68c0\u7d22\u901a\u8fc7\u91c7\u7528\u591a\u9636\u6bb5\uff08\u4e09\u6b21\uff09\u7684\u65b9\u5f0f\u8fdb\u884c<\/p>\n
      \n
    • \n
      \u7b2c\u4e00\u9636\u6bb5\uff1a\u5c0f\u5206\u5757\u68c0\u7d22<\/strong> \u4f7f\u7528\u5c0f\u6587\u6863\u5757(docs_index_small)\u548c\u5c0f\u5d4c\u5165\u5757(embedding_chunks_small)\u521d\u59cb\u5316\u4e00\u4e2a\u68c0\u7d22\u5668(first_retriever)\uff0c\u4f7f\u7528\u8fd9\u4e2a\u68c0\u7d22\u5668\u68c0\u7d22\u4e0e\u67e5\u8be2\u76f8\u5173\u7684\u6587\u6863\uff0c\u5e76\u5c06\u7ed3\u679c\u5b58\u50a8\u5728 first \u53d8\u91cf\u4e2d\uff0c\u5bf9\u68c0\u7d22\u5230\u7684\u6587\u6863 ID \u8fdb\u884c\u6e05\u7406\u548c\u8fc7\u6ee4\uff0c\u786e\u4fdd\u5b83\u4eec\u662f\u76f8\u5173\u7684\uff0c\u5e76\u5b58\u50a8\u5728 ids_clean \u53d8\u91cf\u4e2d\u3002<\/p>\n<\/li>\n
    • \n
      \u7b2c\u4e8c\u9636\u6bb5\uff1a\u79fb\u52a8\u7a97\u53e3\u68c0\u7d22<\/strong> \u9488\u5bf9\u6bcf\u4e2a\u552f\u4e00\u7684\u6e90\u6587\u6863\uff0c\u4f7f\u7528\u5c0f\u6587\u6863\u5757\u68c0\u7d22\u4e0e\u8be5\u6e90\u6587\u6863\u76f8\u5173\u7684\u6240\u6709\u6587\u6863\u5757\u3002\u4f7f\u7528\u5305\u542b\u8fd9\u4e9b\u6587\u6863\u5757\u7684\u65b0\u68c0\u7d22\u5668(second_retriever)\uff0c\u518d\u6b21\u8fdb\u884c\u68c0\u7d22\uff0c\u4ee5\u8fdb\u4e00\u6b65\u7f29\u5c0f\u76f8\u5173\u6587\u6863\u7684\u8303\u56f4\uff0c\u5c06\u68c0\u7d22\u5230\u7684\u6587\u6863\u6dfb\u52a0\u5230 docs \u5217\u8868\u4e2d\u3002<\/p>\n<\/li>\n
    • \n
      \u7b2c\u4e09\u9636\u6bb5\uff1a\u4e2d\u7b49\u5206\u5757\u68c0\u7d22<\/strong> \u4f7f\u7528\u8fc7\u6ee4\u6761\u4ef6\u4ece\u4e2d\u7b49\u6587\u6863\u5757(docs_index_medium)\u68c0\u7d22\u76f8\u5173\u6587\u6863\uff0c\u4f7f\u7528\u5305\u542b\u8fd9\u4e9b\u6587\u6863\u5757\u7684\u65b0\u68c0\u7d22\u5668(third_retriever)\u8fdb\u884c\u68c0\u7d22\u3002\u4ece\u68c0\u7d22\u5230\u7684\u6587\u6863\u4e2d\u9009\u62e9\u524d third_num_k \u4e2a\u6587\u6863\uff0c\u5b58\u50a8\u5728 third \u53d8\u91cf\u4e2d\uff0c\u6e05\u7406\u6587\u6863\u7684\u5143\u6570\u636e\uff0c\u5220\u9664\u4e0d\u9700\u8981\u7684\u5185\u5bb9\uff0c\u5c06\u6700\u7ec8\u68c0\u7d22\u5230\u7684\u6587\u6863\u6309\u6587\u4ef6\u540d\u5206\u7c7b\uff0c\u5e76\u5b58\u50a8\u5728 qa_chunks \u5b57\u5178\u4e2d\u3002<\/p>\n
      <\/div>
      ini<\/span>
      \u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>def get_relevant_documents(<\/span>\n    self,<\/span>\n    query: str,<\/span>\n    num_query: int,<\/span>\n    *,<\/span>\n    run_manager: Optional[CallbackManagerForChainRun]<\/span> = None,<\/span>\n) -> List[Document]<\/span>:<\/span>\n    # \u7b2c\u4e00\u6b21\u68c0\u7d22\uff0c\u5c0f\u5206\u5757\u4fe1\u606f<\/span><\/span>\n    first_retriever<\/span> = self.get_retriever(<\/span>\n        docs_chunks<\/span>=self.docs_index_small.documents,<\/span>\n        emb_chunks<\/span>=self.embedding_chunks_small,<\/span>\n        emb_filter<\/span>=None,<\/span>\n        k<\/span>=self.first_retrieval_k,<\/span>\n        weights<\/span>=self.retriever_weights,<\/span>\n    )<\/span>\n    first<\/span> = first_retriever.get_relevant_documents(<\/span>\n        query, callbacks<\/span>=run_manager.get_child()<\/span>\n    )<\/span>\n    ids_clean<\/span> = self.get_relevant_doc_ids(first, query)<\/span>\n    qa_chunks<\/span> = {}<\/span>\n    if ids_clean and isinstance(ids_clean, list):<\/span>\n        source_md5_dict<\/span> = {}<\/span>\n        for ids_c in ids_clean:<\/span>\n            if ids_c < len(first):<\/span>\n                if ids_c not in source_md5_dict:<\/span>\n                    source_md5_dict[first[ids_c]<\/span>.metadata[\"source_md5\"]]<\/span> = [<\/span><\/span>\n                        first[ids_c]<\/span>\n                    ]<\/span>\n        if len(source_md5_dict) == 0:<\/span>\n            source_md5_dict[first[0]<\/span>.metadata[\"source_md5\"]]<\/span> = [first[0]]<\/span><\/span>\n        num_docs<\/span> = len(source_md5_dict.keys())<\/span>\n        third_num_k<\/span> = max(<\/span>\n            1,<\/span>\n            (<\/span>\n                int(<\/span>\n                    (<\/span>\n                        MAX_LLM_CONTEXT<\/span>\n                        \/ (BASE_CHUNK_SIZE * CHUNK_SCALE)<\/span>\n                    )<\/span>\n                    \/\/ (num_docs * num_query)<\/span>\n                )<\/span>\n            ),<\/span>\n        )<\/span>\n<\/span>\n        for source_md5, docs in source_md5_dict.items():<\/span>\n            second_docs_chunks<\/span> = self.docs_index_small.retrieve_metadata(<\/span>\n                {<\/span>\n                    \"source_md5\": (IndexerOperator.EQ, source_md5),<\/span>\n                }<\/span>\n            )<\/span>\n            # \u7b2c\u4e8c\u6b21\u68c0\u7d22<\/span><\/span>\n            second_retriever<\/span> = self.get_retriever(<\/span>\n                docs_chunks<\/span>=second_docs_chunks,<\/span>\n                emb_chunks<\/span>=self.embedding_chunks_small,<\/span>\n                emb_filter<\/span>={\"source_md5\"<\/span>: source_md5},<\/span>\n                k<\/span>=self.second_retrieval_k,<\/span>\n                weights<\/span>=self.retriever_weights,<\/span>\n            )<\/span>\n            second<\/span> = second_retriever.get_relevant_documents(<\/span>\n                query, callbacks<\/span>=run_manager.get_child()<\/span>\n            )<\/span>\n            docs.extend(second)<\/span>\n            docindexer_filter, chroma_filter<\/span> = self.get_filter(<\/span>\n                self.num_windows, source_md5, docs<\/span>\n            )<\/span>\n            third_docs_chunks<\/span> = self.docs_index_medium.retrieve_metadata(<\/span>\n                docindexer_filter<\/span>\n            )<\/span>\n            # \u7b2c\u4e09\u6b21\u68c0\u7d22<\/span><\/span>\n            third_retriever<\/span> = self.get_retriever(<\/span>\n                docs_chunks<\/span>=third_docs_chunks,<\/span>\n                emb_chunks<\/span>=self.embedding_chunks_medium,<\/span>\n                emb_filter<\/span>=chroma_filter,<\/span>\n                k<\/span>=third_num_k,<\/span>\n                weights<\/span>=self.retriever_weights,<\/span>\n            )<\/span>\n            third_temp<\/span> = third_retriever.get_relevant_documents(<\/span>\n                query, callbacks<\/span>=run_manager.get_child()<\/span>\n            )<\/span>\n            third<\/span> = third_temp[:third_num_k]<\/span>\n            for doc in third:<\/span>\n                mtdata<\/span> = doc.metadata<\/span>\n                mtdata[\"page_content\"]<\/span> = None<\/span>\n            file_name<\/span> = third[0<\/span>].metadata[\"source\"<\/span>].split(\"\/\"<\/span>)[-1<\/span>]<\/span>\n            if file_name not in qa_chunks:<\/span>\n                qa_chunks[file_name]<\/span> = third<\/span>\n            else:<\/span>\n                qa_chunks[file_name]<\/span>.extend(third)<\/span>\n    return qa_chunks<\/span>\n<\/code><\/pre>\n<\/li>\n<\/ul>\n
      \u6574\u4e2a\u8fc7\u7a0b\u662f\u4e00\u4e2a\u5206\u5c42\u7684\u68c0\u7d22\u8fc7\u7a0b\uff0c\u9996\u5148\u5728\u5c0f\u6587\u6863\u5757\u4e2d\u8fdb\u884c\u7c97\u7565\u68c0\u7d22\uff0c\u7136\u540e\u5728\u7279\u5b9a\u7684\u6e90\u6587\u6863\u4e2d\u8fdb\u884c\u66f4\u7cbe\u786e\u7684\u68c0\u7d22\uff0c\u6700\u540e\u5728\u4e2d\u7b49\u6587\u6863\u5757\u4e2d\u8fdb\u884c\u6700\u7ec8\u7684\u68c0\u7d22\u3002\u8fd9\u79cd\u5206\u5c42\u7684\u65b9\u6cd5\u6709\u52a9\u4e8e\u63d0\u9ad8\u68c0\u7d22\u7684\u6548\u7387\u548c\u51c6\u786e\u6027\uff0c\u56e0\u4e3a\u5b83\u5141\u8bb8\u7cfb\u7edf\u5728\u66f4\u5c0f\u7684\u6587\u6863\u96c6\u4e0a\u8fdb\u884c\u66f4\u7cbe\u786e\u7684\u68c0\u7d22\uff0c\u4ece\u800c\u51cf\u5c11\u4e86\u5728\u5927\u6587\u6863\u96c6\u4e0a\u8fdb\u884c\u590d\u6742\u68c0\u7d22\u6240\u9700\u7684\u8ba1\u7b97\u91cf\u3002<\/p>\n
      \u8be5\u65b9\u6848\u7684\u4f18\u52bf<\/h3>\n
      \u5904\u7406\u5927\u89c4\u6a21\u6587\u6863<\/h4>\n
      \u7531\u4e8e\u77e5\u8bc6\u5e93\u901a\u5e38\u5305\u542b\u5927\u91cf\u7684\u6587\u6863\uff0c\u76f4\u63a5\u5728\u8fd9\u4e48\u5927\u7684\u6587\u6863\u96c6\u5408\u4e0a\u8fdb\u884c\u68c0\u7d22\u662f\u975e\u5e38\u8017\u65f6\u7684\u3002\u901a\u8fc7\u5c06\u6587\u6863\u5206\u5272\u6210\u66f4\u5c0f\u7684\u5757\uff08chunk_small<\/code>\uff09\uff0c\u5c0f\u5757\u66f4\u5bb9\u6613\u88ab\u7d22\u5f15\u548c\u68c0\u7d22\u3002<\/p>\n
      \u4fdd\u7559\u4e0a\u4e0b\u6587\u4fe1\u606f<\/h4>\n
      \u901a\u8fc7\u4e3a\u5c0f\u5757\u6dfb\u52a0\u7a97\u53e3\u4fe1\u606f\uff08add_window<\/code>\uff09\uff0c\u53ef\u4ee5\u786e\u4fdd\u5728\u68c0\u7d22\u65f6\u4e0d\u4f1a\u4e22\u5931\u91cd\u8981\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\u3002\u8fd9\u662f\u56e0\u4e3a\u6709\u4e9b\u4fe1\u606f\u53ef\u80fd\u5206\u5e03\u5728\u591a\u4e2a\u5c0f\u5757\u4e2d\uff0c\u5355\u72ec\u68c0\u7d22\u4e00\u4e2a\u5c0f\u5757\u53ef\u80fd\u4f1a\u9057\u6f0f\u8fd9\u4e9b\u4fe1\u606f\uff0c\u7a97\u53e3\u673a\u5236\u786e\u4fdd\u5728\u68c0\u7d22\u65f6\u8003\u8651\u5230\u8db3\u591f\u7684\u4e0a\u4e0b\u6587\uff0c\u4ece\u800c\u751f\u6210\u66f4\u51c6\u786e\u7684\u56de\u7b54\u3002<\/p>\n
      \u63d0\u5347\u68c0\u7d22\u6548\u7387<\/h4>\n
      \u901a\u8fc7\u5c06\u76f8\u90bb\u7684\u5c0f\u5757\u5408\u5e76\u6210\u4e2d\u7b49\u5927\u5c0f\u7684\u5757\uff08chunk_medium<\/code>\uff09\uff0c\u5728\u4fdd\u7559\u7ec6\u7c92\u5ea6\u7279\u6027\u7684\u540c\u65f6\u589e\u52a0\u4e86\u66f4\u5927\u8303\u56f4\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\u3002\u8fd9\u6709\u52a9\u4e8e\u63d0\u5347\u68c0\u7d22\u7684\u6548\u7387\u548c\u51c6\u786e\u6027\uff0c\u56e0\u4e3a\u4e2d\u7b49\u5927\u5c0f\u7684\u5757\u65e2\u4e0d\u4f1a\u50cf\u5927\u5757\u90a3\u6837\u5bfc\u81f4\u68c0\u7d22\u6548\u7387\u4e0b\u964d\uff0c\u4e5f\u4e0d\u4f1a\u50cf\u5c0f\u5757\u90a3\u6837\u7f3a\u4e4f\u8db3\u591f\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\u3002<\/p>\n
      \u7075\u6d3b\u6027\u548c\u53ef\u914d\u7f6e\u6027<\/h4>\n
      \u7531\u4e8e\u6574\u4e2a\u6d41\u7a0b\u662f\u6a21\u5757\u5316\u7684\uff0c\u53ef\u4ee5\u6839\u636e\u5177\u4f53\u5e94\u7528\u7684\u9700\u6c42\u7075\u6d3b\u5730\u914d\u7f6e\u6bcf\u4e2a\u6b65\u9aa4\u7684\u53c2\u6570\uff08\u5982\u5757\u7684\u5927\u5c0f\u3001\u7a97\u53e3\u7684\u5927\u5c0f\u548c\u6b65\u957f\u7b49\uff09\uff0c\u4ee5\u8fbe\u5230\u6700\u4f73\u7684\u6027\u80fd\u548c\u6548\u679c\u5e73\u8861\u3002<\/p>\n
      \u652f\u6301\u591a\u79cd\u68c0\u7d22\u7b56\u7565<\/h4>\n
      \u7531\u4e8e\u6709\u4e86\u4e0d\u540c\u5927\u5c0f\u548c\u5305\u542b\u7a97\u53e3\u4fe1\u606f\u7684\u6587\u6863\u5757\uff0c\u53ef\u4ee5\u6839\u636e\u67e5\u8be2\u7684\u9700\u8981\u9009\u62e9\u6700\u9002\u5408\u7684\u5757\u6765\u8fdb\u884c\u68c0\u7d22\u3002\u5bf9\u4e8e\u9700\u8981\u5e7f\u6cdb\u4e0a\u4e0b\u6587\u7684\u67e5\u8be2\uff0c\u53ef\u4ee5\u4f7f\u7528\u5305\u542b\u66f4\u591a\u4e0a\u4e0b\u6587\u4fe1\u606f\u7684\u4e2d\u7b49\u5927\u5c0f\u6216\u5927\u5757\u6765\u8fdb\u884c\u68c0\u7d22\uff0c\u5bf9\u4e8e\u9700\u8981\u5feb\u901f\u54cd\u5e94\u7684\u67e5\u8be2\uff0c\u53ef\u4ee5\u4f7f\u7528\u5c0f\u5757\u6765\u63d0\u9ad8\u68c0\u7d22\u901f\u5ea6\u3002<\/p>\n
      \u6574\u4f53\u65b9\u6848\u5728\u4fdd\u8bc1\u751f\u6210\u8d28\u91cf\u7684\u540c\u65f6\uff0c\u5b9e\u73b0\u9ad8\u6548\u5904\u7406\uff0c\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u6548\u679c\u660e\u663e\u3002<\/strong><\/p>\n
      \u4e0d\u8db3\u4e4b\u5904<\/h3>\n
      \u56e0\u4e3a\u68c0\u7d22\u9636\u6bb5\u7ecf\u8fc7\u4e09\u6b21\u68c0\u7d22\uff0c\u52a0\u4e4b\u4f7f\u7528\u672c\u5730\u77e2\u91cf\u6570\u636e\u5e93\u4f7f\u7528 Chroma[1]\uff0c\u6574\u4f53\u7684\u54cd\u5e94\u901f\u7387\u8fd8\u9700\u8fdb\u4e00\u6b65\u6539\u5584\u3002<\/p>\n
      \u66f4\u591a\u5185\u5bb9\u5728\u516c\u53f7\uff1aLLM \u5e94\u7528\u5168\u6808\u5f00\u53d1<\/a>\uff0c\u56de\u590d RAG<\/strong> \u83b7\u53d6\u6e90\u7801<\/p>","protected":false},"excerpt":{"rendered":"
      \u672c\u6587\u9996\u53d1\u4e8e\u535a\u5ba2LLM \u5e94\u7528\u5f00\u53d1\u5b9e\u8df5 \u5148\u770b\u6548\u679c \u6f14\u793a\u6548\u679c \u6574\u4f53\u65b9\u6848 \u6574\u4f53\u65b9\u6848\u5728\u4e8e\u6587\u6863\u9884\u5904\u7406\u9636\u6bb5\u5b9e\u73b0\u6ee1\u8db3\u4e0a\u4e0b\u6587\u7a97\u53e3\u7684\u539f\u59cb\u6587\u672c\u5206\u5757\uff0c\u6587\u6863\u68c0\u7d22\u9636\u6bb5\u5b9e\u73b0\u6587\u672c\u7684\u4e09\u6b21\u68c0\u7d22\uff0c\u4e0b\u9762\u9010\u4e00\u8fdb\u884c\u8bf4\u660e\u3002\u6d4b\u8bd5\u6587\u7ae0\u6765\u81ea \u5927\u8bed\u8a00\u6a21\u578b<\/p>\n","protected":false},"author":1,"featured_media":8282,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"rank_math_title":"","rank_math_description":"","rank_math_focus_keyword":"","views":"5","footnotes":""},"categories":[3],"tags":[136,126,127,128,129],"collection":[],"class_list":["post-1406","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-fenlei2","tag-136","tag-gpt","tag-ai","tag-128","tag-129"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/posts\/1406","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/comments?post=1406"}],"version-history":[{"count":0,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/posts\/1406\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/media\/8282"}],"wp:attachment":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/media?parent=1406"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/categories?post=1406"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/tags?post=1406"},{"taxonomy":"collection","embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/collection?post=1406"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}