{"text": "这期我们重做一版PDF四方测评。不是问哪个工具永远最好,而是问同一份PDF变成MarkDone以后能不能放心交给AI。我先核对了四个工具的官方用法,再在本地跑同一套样本,Mark it down,Duckling,Marker,MinerU。这一次只比较成功输出的效果。运行环境,设备选择,依赖细节,只作为内部诊断。不拿来做视频结论。样本一共七类,原生文本报告,可填写表单,双栏论文,财报表格,掌技术手册,专利PDF,还有换灯片PDF。每类只切代表页。因为Apple 10K,NASA手册这种长文档,如果直接整本跑,测到的很可能是压力测试,不是抽取质量。脚本做三件事,切样本页,跑四个工具,最后检查关键词,标题,表格型,图片引用和可读摘录。分数不是绝对真理,它只是一个质检信号。这份输出有没有资格进入下一步?AI上下文。先看配置边界。Mark it down 用默认离线转换。Darklin输出 Mark down,并把图片设为 Placeholder。Marker 这一轮使用 No OCR, No Image Extraction 的文本层快速配置,所以,它不代表默认 OCR 能力。Miner U 使用 Pipeline,保留表格,但本轮不评估公式解析。总览结果,Mark it down 七个样本全部完成,平均规则分83,平均耗时1.23秒。它向第一道,速度非常快。Darklin 平均 88.9分,平均 9.31秒。它不是最快,但结构感最稳。Marker 在这轮文本层快速配置下,平均 88.1分,平均 36.2秒。它的边界要写清楚,适合有文本层,不需要 OCR 的场景。Miner U 平均 79.1分,平均 27.7分,平均 27.7分。Darklin 平均耗时1.23,平均耗时1.23秒。它向第一道,速度非常快。Darklin 平均 88.9分,平均 9.31秒。Darklin 平均 88.9分,平均 9.31秒。Miner U 平均 79.1分,平均 27.79秒。它更像重型管线,适合你愿意为复杂版面多复配置成本的时候。第一个案例是 Nest 报告。四个工具都能读到目录和 Executive Summary,但结构不一样。Mark it down 把目录更像普通文本串出来。Darklin 和 Marker 更容易形成表格化目录。Miner U 保留了标题,但表格结构弱一些。第二个案例是 W 四表单。这里不是看有没有字,而是看 Step 1 姓名、地址、 Worksheet 这些字段还在不在同一个语意块里。Darklin 和 Marker 的结构最完整。Miner U 给出 HTML Table,Mark it down 很快,但更像线性文本。第三个案例是双栏论文。论文最怕阅读顺序飘移。Mark it down 开头会带一些 AR 14 页面噪声。Darklin, Marker, Minor U 都能稳定抓到标题和作者,但后续图表位置仍然要抽样检查。第四个案例是财报表格。这里三家都拿到高分,但看真实输出会发现差异。Darklin 的年份列,指标行更规整。Marker 也能保留多列表格。Mark it down 可用,但有些列被拆得更碎。Miner U 输出 HTML Table,不过个别行会有重排噪声。第五个案例是长技术手册。长文档不要一上来整本入库,先抽目录、图表说明和章节页。Darklin 和 Marker 的目录结构最清楚。Mark it down 适合快速预览。Miner U 保留部分图像线索。第六个案例是专利,PDF。它看起来像扫描件,但实际带有可抽取文本层。所以,No OCR 配置也能拿到内容。这就是PDF 制简第一步,先判断有没有文本层,不要只凭页面长相决定要不要 OCR。最后是幻灯片,PDF。所有工具都不算轻松。问题不是读不到字,而是幻灯片的语意藏在布局里,大标题,箭头,左右分组,图形关系。转成 Mark it down 后,这些关系很容易被压平成段落。所以,这期的结论不是选一个神工具。如果你只是快速读普通 PDF,Mark it down 很适合做第一刀。如果你要结构、表格和知识库候选,Darklin 更稳。如果你确认 PDF 有文本层,而且不需要 OCR,Marker 的快速配置表现很好,但一定要著名边界。如果你面对扫描件或复杂版面,Miner U值得保留,不过要接受更重的依赖和参数成本。真正可靠的流程,是 PDF 入库前做四个检查,有没有文本层,阅读顺序有没有乱,表格行列有没有保住,图片,公式,扫描页有没有抽样。工具不是答案,质检流程才是答案。下一期如果继续,我们可以把这套测评变成批量 PDF 入库工作流。不仅看 Mark it down 好不好看,还看尽 Rack hole 能不能引用,能不能追溯,能不能查错。不仅看 Mark it down 。不仅看 Mark it down 。不仅看 Mark it down 。不仅看 Mark it down 。不仅看 Mark it down 。不仅看 Mark it down 。不仅看 Mark it down 。不仅看 Mark it down 。不仅看 Mark it down 。不仅看 Mark it down 。不仅看 Mark it down 。不仅看 Mark it down 。不仅看 Mark it down 。不仅看 Mark it down 。不仅看 Mark it down 。不仅看 Mark it down 。不仅看 Mark it down 。不仅看 Mark it down 。不仅看 Mark it down 。不仅看 Mark it down 。", "segments": [{"id": 0, "seek": 0, "start": 0.0, "end": 29.900000000000002, "text": "这期我们重做一版PDF四方测评。不是问哪个工具永远最好,而是问同一份PDF变成MarkDone以后能不能放心交给AI。我先核对了四个工具的官方用法,再在本地跑同一套样本,Mark it down,Duckling,Marker,MinerU。这一次只比较成功输出的效果。运行环境,设备选择,依赖细节,只作为内部诊断。", "tokens": [50365, 5562, 16786, 15003, 12624, 10907, 2257, 42096, 17349, 37, 19425, 9249, 11038, 233, 5233, 226, 1543, 7296, 22064, 17028, 7549, 23323, 39806, 42503, 3316, 250, 8661, 2131, 11, 11070, 1541, 22064, 13089, 2257, 36266, 17349, 37, 2129, 246, 11336, 15168, 35, 546, 3588, 13547, 8225, 28590, 39619, 28455, 23197, 48698, 1543, 1654, 10108, 8987, 116, 8713, 2289, 19425, 7549, 23323, 39806, 1546, 31929, 9249, 9254, 11148, 11, 8623, 3581, 8802, 10928, 32585, 13089, 2257, 1881, 245, 14496, 8802, 11, 15168, 309, 760, 11, 35, 1134, 1688, 11, 15168, 260, 11, 44, 4564, 52, 1543, 5562, 27505, 14003, 11706, 9830, 225, 11336, 32311, 9830, 241, 7781, 1546, 43076, 9319, 1543, 3316, 238, 8082, 8051, 107, 47885, 11, 7422, 122, 1787, 229, 2215, 231, 6852, 102, 11, 3254, 251, 5266, 244, 10115, 228, 45161, 11, 14003, 11914, 13992, 34742, 13470, 5233, 232, 4307, 255, 1543, 51860], "temperature": 0, "avg_logprob": -0.09434267592756715, "compression_ratio": 1.1501597444089458, "no_speech_prob": 7.69795251848171e-12}, {"id": 1, "seek": 3000, "start": 30.0, "end": 59.8, "text": "不拿来做视频结论。样本一共七类,原生文本报告,可填写表单,双栏论文,财报表格,掌技术手册,专利PDF,还有换灯片PDF。每类只切代表页。因为Apple 10K,NASA手册这种长文档,如果直接整本跑,测到的很可能是压力测试,不是抽取质量。脚本做三件事,切样本页,跑四个工具,", "tokens": [50365, 1960, 24351, 6912, 10907, 40656, 39752, 45641, 7422, 118, 1543, 14496, 8802, 2257, 15408, 29920, 22113, 119, 11, 19683, 8244, 17174, 8802, 49817, 16846, 11, 4429, 13331, 104, 5676, 247, 17571, 47446, 11, 2129, 234, 8987, 237, 7422, 118, 17174, 11, 18464, 95, 49817, 17571, 30921, 11, 6900, 234, 32502, 1474, 107, 11389, 5676, 234, 11, 940, 241, 23700, 17349, 37, 11, 35091, 26075, 95, 13290, 107, 16668, 17349, 37, 1543, 23664, 22113, 119, 14003, 23632, 19105, 17571, 10178, 113, 1543, 34627, 9132, 306, 1266, 42, 11, 45, 3160, 32, 11389, 5676, 234, 5562, 39810, 32271, 17174, 21416, 96, 11, 13119, 43297, 27662, 8802, 32585, 11, 11038, 233, 4511, 1546, 4563, 16657, 1541, 5014, 233, 13486, 11038, 233, 5233, 243, 11, 7296, 46022, 29436, 18464, 101, 26748, 1543, 27067, 248, 8802, 10907, 10960, 20485, 6973, 11, 23632, 14496, 8802, 10178, 113, 11, 32585, 19425, 7549, 23323, 39806, 11, 51855], "temperature": 0, "avg_logprob": -0.06603731066975373, "compression_ratio": 1.1616161616161615, "no_speech_prob": 1.2617475640686049e-11}, {"id": 2, "seek": 6000, "start": 60.0, "end": 74.0, "text": "最后检查关键词,标题,表格型,图片引用和可读摘录。分数不是绝对真理,它只是一个质检信号。这份输出有没有资格进入下一步?AI上下文。", "tokens": [50365, 8661, 13547, 16407, 222, 42623, 28053, 23049, 106, 5233, 235, 11, 162, 3921, 30716, 11, 17571, 30921, 39823, 11, 3919, 122, 16668, 41301, 9254, 12565, 4429, 5233, 119, 34783, 246, 7391, 243, 1543, 6627, 33188, 7296, 10115, 251, 8713, 6303, 13876, 11, 11284, 36859, 20182, 18464, 101, 16407, 222, 17665, 26987, 1543, 5562, 36266, 9830, 241, 7781, 2412, 17944, 5266, 226, 30921, 36700, 14028, 45271, 31429, 30, 48698, 5708, 4438, 17174, 1543, 51065], "temperature": 0, "avg_logprob": -0.08769185919510691, "compression_ratio": 1.052325581395349, "no_speech_prob": 1.4613789345307993e-11}, {"id": 3, "seek": 7400, "start": 74.0, "end": 92.0, "text": "先看配置边界。Mark it down 用默认离线转换。Darklin输出 Mark down,并把图片设为 Placeholder。Marker 这一轮使用 No OCR, No Image Extraction 的文本层快速配置,所以,它不代表默认 OCR 能力。", "tokens": [50365, 10108, 4200, 38846, 34719, 40503, 21185, 1543, 15168, 309, 760, 220, 9254, 6173, 246, 7422, 97, 18614, 119, 16853, 123, 17819, 105, 26075, 95, 1543, 35, 809, 5045, 9830, 241, 7781, 3934, 760, 11, 3509, 114, 16075, 3919, 122, 16668, 7422, 122, 13992, 13637, 20480, 1543, 15168, 260, 220, 5562, 2257, 17819, 106, 22982, 9254, 883, 422, 18547, 11, 883, 29903, 9881, 26766, 27949, 17174, 8802, 9636, 224, 10251, 31217, 38846, 34719, 11, 7239, 11, 11284, 1960, 19105, 17571, 6173, 246, 7422, 97, 422, 18547, 220, 8225, 13486, 1543, 51265], "temperature": 0, "avg_logprob": -0.11322018408006237, "compression_ratio": 1.0454545454545454, "no_speech_prob": 1.4750402288488118e-11}, {"id": 4, "seek": 9200, "start": 92.0, "end": 104.0, "text": "Miner U 使用 Pipeline,保留表格,但本轮不评估公式解析。总览结果,Mark it down 七个样本全部完成,平均规则分83,平均耗时1.23秒。它向第一道,速度非常快。Darklin 平均 88.9分,平均 9.31秒。它不是最快,但结构感最稳。", "tokens": [50365, 44, 4564, 624, 220, 22982, 9254, 35396, 5440, 11, 24302, 24456, 17571, 30921, 11, 8395, 8802, 17819, 106, 1960, 5233, 226, 7384, 108, 13545, 27584, 17278, 7360, 238, 1543, 33440, 6758, 230, 45641, 9319, 11, 15168, 309, 760, 220, 29920, 7549, 14496, 8802, 38714, 41509, 11, 16716, 14872, 229, 6758, 226, 2437, 247, 6627, 31849, 11, 16716, 14872, 229, 4450, 245, 15729, 16, 13, 9356, 20956, 1543, 11284, 24282, 18049, 6025, 11, 31217, 13127, 14392, 10251, 1543, 35, 809, 5045, 220, 16716, 14872, 229, 24587, 13, 24, 6627, 11, 16716, 14872, 229, 1722, 13, 12967, 20956, 1543, 11284, 7296, 8661, 10251, 11, 8395, 45641, 7360, 226, 9709, 8661, 10415, 111, 1543, 50965], "temperature": 0, "avg_logprob": -0.10904312133789062, "compression_ratio": 1.096774193548387, "no_speech_prob": 1.527476929663596e-11}, {"id": 5, "seek": 10400, "start": 104.0, "end": 116.0, "text": "Marker 在这轮文本层快速配置下,平均 88.1分,平均 36.2秒。它的边界要写清楚,适合有文本层,不需要 OCR 的场景。", "tokens": [50365, 15168, 260, 37286, 5562, 17819, 106, 17174, 8802, 9636, 224, 10251, 31217, 38846, 34719, 4438, 11, 16716, 14872, 229, 24587, 13, 16, 6627, 11, 16716, 14872, 229, 8652, 13, 17, 20956, 1543, 45224, 40503, 21185, 4275, 5676, 247, 38433, 11, 2215, 224, 14245, 2412, 17174, 8802, 9636, 224, 11, 1960, 35748, 422, 18547, 27949, 50255, 50218, 1543, 50965], "temperature": 0, "avg_logprob": -0.11720600128173828, "compression_ratio": 1.5407725321888412, "no_speech_prob": 1.8258004483295842e-11}, {"id": 6, "seek": 10400, "start": 116.0, "end": 122.0, "text": "Miner U 平均 79.1分,平均 27.7分,平均 27.7分。", "tokens": [50965, 44, 4564, 624, 220, 16716, 14872, 229, 32803, 13, 16, 6627, 11, 16716, 14872, 229, 7634, 13, 22, 6627, 11, 16716, 14872, 229, 7634, 13, 22, 6627, 1543, 51265], "temperature": 0, "avg_logprob": -0.11720600128173828, "compression_ratio": 1.5407725321888412, "no_speech_prob": 1.8258004483295842e-11}, {"id": 7, "seek": 10400, "start": 122.0, "end": 128.0, "text": "Darklin 平均耗时1.23,平均耗时1.23秒。它向第一道,速度非常快。Darklin 平均 88.9分,平均 9.31秒。", "tokens": [51265, 35, 809, 5045, 220, 16716, 14872, 229, 4450, 245, 15729, 16, 13, 9356, 11, 16716, 14872, 229, 4450, 245, 15729, 16, 13, 9356, 20956, 1543, 11284, 24282, 18049, 6025, 11, 31217, 13127, 14392, 10251, 1543, 35, 809, 5045, 220, 16716, 14872, 229, 24587, 13, 24, 6627, 11, 16716, 14872, 229, 1722, 13, 12967, 20956, 1543, 51565], "temperature": 0, "avg_logprob": -0.11720600128173828, "compression_ratio": 1.5407725321888412, "no_speech_prob": 1.8258004483295842e-11}, {"id": 8, "seek": 10400, "start": 128.0, "end": 130.0, "text": "Darklin 平均 88.9分,平均 9.31秒。", "tokens": [51565, 35, 809, 5045, 220, 16716, 14872, 229, 24587, 13, 24, 6627, 11, 16716, 14872, 229, 1722, 13, 12967, 20956, 1543, 51665], "temperature": 0, "avg_logprob": -0.11720600128173828, "compression_ratio": 1.5407725321888412, "no_speech_prob": 1.8258004483295842e-11}, {"id": 9, "seek": 13000, "start": 130.0, "end": 156.0, "text": "Miner U 平均 79.1分,平均 27.79秒。它更像重型管线,适合你愿意为复杂版面多复配置成本的时候。第一个案例是 Nest 报告。四个工具都能读到目录和 Executive Summary,但结构不一样。Mark it down 把目录更像普通文本串出来。Darklin 和 Marker 更容易形成表格化目录。", "tokens": [50365, 44, 4564, 624, 220, 16716, 14872, 229, 32803, 13, 16, 6627, 11, 16716, 14872, 229, 7634, 13, 32042, 20956, 1543, 11284, 19002, 12760, 12624, 39823, 23131, 16853, 123, 11, 2215, 224, 14245, 2166, 4820, 123, 9042, 13992, 1787, 235, 4422, 224, 42096, 8833, 6392, 1787, 235, 38846, 34719, 11336, 8802, 49873, 1543, 18049, 7549, 28899, 17797, 1541, 31581, 220, 49817, 16846, 1543, 19425, 7549, 23323, 39806, 7182, 8225, 5233, 119, 4511, 11386, 7391, 243, 12565, 20658, 8626, 76, 822, 11, 8395, 45641, 7360, 226, 1960, 2257, 14496, 1543, 15168, 309, 760, 220, 16075, 11386, 7391, 243, 19002, 12760, 29993, 19550, 17174, 8802, 940, 110, 44561, 1543, 35, 809, 5045, 220, 12565, 3934, 260, 220, 19002, 49212, 30900, 11336, 17571, 30921, 23756, 11386, 7391, 243, 1543, 51665], "temperature": 0, "avg_logprob": -0.06742366524629814, "compression_ratio": 1.1137123745819397, "no_speech_prob": 1.569205355933523e-11}, {"id": 10, "seek": 15600, "start": 156.0, "end": 163.0, "text": "Miner U 保留了标题,但表格结构弱一些。第二个案例是 W 四表单。", "tokens": [50365, 44, 4564, 624, 220, 24302, 24456, 2289, 162, 3921, 30716, 11, 8395, 17571, 30921, 45641, 7360, 226, 5702, 109, 38515, 1543, 19693, 7549, 28899, 17797, 1541, 343, 220, 19425, 17571, 47446, 1543, 50715], "temperature": 0, "avg_logprob": -0.10799484050020258, "compression_ratio": 1.0695652173913044, "no_speech_prob": 2.110269424482336e-11}, {"id": 11, "seek": 15600, "start": 163.0, "end": 175.0, "text": "这里不是看有没有字,而是看 Step 1 姓名、地址、 Worksheet 这些字段还在不在同一个语意块里。Darklin 和 Marker 的结构最完整。", "tokens": [50715, 35102, 7296, 4200, 2412, 17944, 22381, 11, 11070, 1541, 4200, 5470, 502, 220, 7795, 241, 15940, 1231, 10928, 14872, 222, 1231, 6603, 9611, 302, 220, 5562, 13824, 22381, 28427, 14852, 3581, 1960, 3581, 13089, 20182, 5233, 255, 9042, 47734, 15759, 1543, 35, 809, 5045, 220, 12565, 3934, 260, 27949, 45641, 7360, 226, 8661, 14128, 27662, 1543, 51315], "temperature": 0, "avg_logprob": -0.10799484050020258, "compression_ratio": 1.0695652173913044, "no_speech_prob": 2.110269424482336e-11}, {"id": 12, "seek": 17500, "start": 175.0, "end": 198.0, "text": "Miner U 给出 HTML Table,Mark it down 很快,但更像线性文本。第三个案例是双栏论文。论文最怕阅读顺序飘移。Mark it down 开头会带一些 AR 14 页面噪声。Darklin, Marker, Minor U 都能稳定抓到标题和作者,但后续图表位置仍然要抽样检查。", "tokens": [50365, 44, 4564, 624, 220, 23197, 7781, 17995, 25535, 11, 15168, 309, 760, 26029, 10251, 11, 8395, 19002, 12760, 16853, 123, 21686, 17174, 8802, 1543, 35878, 7549, 28899, 17797, 1541, 2129, 234, 8987, 237, 7422, 118, 17174, 1543, 7422, 118, 17174, 8661, 21164, 10034, 227, 5233, 119, 10178, 118, 6346, 237, 11808, 246, 8204, 119, 1543, 15168, 309, 760, 220, 18937, 39862, 12949, 4845, 99, 38515, 8943, 3499, 220, 10178, 113, 8833, 21250, 103, 32045, 1543, 35, 809, 5045, 11, 3934, 260, 11, 36117, 624, 220, 7182, 8225, 10415, 111, 12088, 43069, 4511, 162, 3921, 30716, 12565, 11914, 12444, 11, 8395, 13547, 10115, 255, 3919, 122, 17571, 11160, 34719, 1550, 235, 5823, 4275, 46022, 14496, 16407, 222, 42623, 1543, 51515], "temperature": 0, "avg_logprob": -0.07869611020948065, "compression_ratio": 1.0732600732600732, "no_speech_prob": 1.6195407862018563e-11}, {"id": 13, "seek": 19800, "start": 198.0, "end": 220.0, "text": "第四个案例是财报表格。这里三家都拿到高分,但看真实输出会发现差异。Darklin 的年份列,指标行更规整。Marker 也能保留多列表格。Mark it down 可用,但有些列被拆得更碎。Miner U 输出 HTML Table,不过个别行会有重排噪声。", "tokens": [50365, 7536, 19425, 7549, 28899, 17797, 1541, 18464, 95, 49817, 17571, 30921, 1543, 35102, 10960, 5155, 7182, 24351, 4511, 12979, 6627, 11, 8395, 4200, 6303, 24726, 9830, 241, 7781, 12949, 28926, 20204, 21679, 5702, 224, 1543, 35, 809, 5045, 27949, 5157, 36266, 43338, 11, 25922, 162, 3921, 8082, 19002, 6758, 226, 27662, 1543, 15168, 260, 220, 6404, 8225, 24302, 24456, 6392, 43338, 17571, 30921, 1543, 15168, 309, 760, 220, 4429, 9254, 11, 8395, 2412, 13824, 43338, 23238, 6852, 228, 5916, 19002, 16337, 236, 1543, 44, 4564, 624, 220, 9830, 241, 7781, 17995, 25535, 11, 1960, 16866, 7549, 18453, 8082, 12949, 2412, 12624, 44647, 21250, 103, 32045, 1543, 51465], "temperature": 0, "avg_logprob": -0.07172709378329191, "compression_ratio": 1.0919540229885059, "no_speech_prob": 2.0427295271963075e-11}, {"id": 14, "seek": 22000, "start": 220.0, "end": 237.0, "text": "第五个案例是长技术手册。长文档不要一上来整本入库,先抽目录、图表说明和章节页。Darklin 和 Marker 的目录结构最清楚。Mark it down 适合快速预览。Miner U 保留部分图像线索。", "tokens": [50365, 7536, 21001, 7549, 28899, 17797, 1541, 32271, 32502, 1474, 107, 11389, 5676, 234, 1543, 32271, 17174, 21416, 96, 11962, 2257, 5708, 6912, 27662, 8802, 14028, 6346, 241, 11, 10108, 46022, 11386, 7391, 243, 1231, 3919, 122, 17571, 8090, 11100, 12565, 11957, 254, 45161, 10178, 113, 1543, 35, 809, 5045, 220, 12565, 3934, 260, 27949, 11386, 7391, 243, 45641, 7360, 226, 8661, 38433, 1543, 15168, 309, 760, 220, 2215, 224, 14245, 10251, 31217, 12501, 226, 6758, 230, 1543, 44, 4564, 624, 220, 24302, 24456, 32174, 3919, 122, 12760, 16853, 123, 7732, 95, 1543, 51215], "temperature": 0, "avg_logprob": -0.08328752716382344, "compression_ratio": 1.0360360360360361, "no_speech_prob": 1.805237036245355e-11}, {"id": 15, "seek": 23700, "start": 237.0, "end": 249.0, "text": "第六个案例是专利,PDF。它看起来像扫描件,但实际带有可抽取文本层。所以,No OCR 配置也能拿到内容。", "tokens": [50365, 7536, 30566, 7549, 28899, 17797, 1541, 940, 241, 23700, 11, 17349, 37, 1543, 11284, 4200, 9147, 6912, 12760, 3416, 104, 11673, 237, 20485, 11, 8395, 24726, 8842, 227, 4845, 99, 2412, 4429, 46022, 29436, 17174, 8802, 9636, 224, 1543, 7239, 11, 4540, 422, 18547, 220, 38846, 34719, 6404, 8225, 24351, 4511, 34742, 25750, 1543, 50965], "temperature": 0, "avg_logprob": -0.10492443216258082, "compression_ratio": 0.9236111111111112, "no_speech_prob": 1.862242651751167e-11}, {"id": 16, "seek": 24900, "start": 249.0, "end": 271.16, "text": "这就是PDF 制简第一步,先判断有没有文本层,不要只凭页面长相决定要不要 OCR。最后是幻灯片,PDF。所有工具都不算轻松。问题不是读不到字,而是幻灯片的语意藏在布局里,大标题,箭头,左右分组,图形关系。", "tokens": [50365, 5562, 5620, 17349, 37, 220, 25491, 11249, 222, 18049, 31429, 11, 10108, 2437, 97, 4307, 255, 2412, 17944, 17174, 8802, 9636, 224, 11, 11962, 14003, 6336, 255, 10178, 113, 8833, 32271, 15106, 5676, 111, 12088, 4275, 11962, 422, 18547, 1543, 8661, 13547, 1541, 3509, 119, 13290, 107, 16668, 11, 17349, 37, 1543, 39300, 23323, 39806, 7182, 1960, 19497, 17819, 119, 36897, 1543, 34069, 7296, 5233, 119, 35728, 22381, 11, 11070, 1541, 3509, 119, 13290, 107, 16668, 1546, 5233, 255, 9042, 21782, 237, 3581, 34688, 34703, 15759, 11, 3582, 162, 3921, 30716, 11, 11249, 255, 39862, 11, 29457, 6627, 10115, 226, 11, 3919, 122, 30900, 28053, 25368, 1543, 51473], "temperature": 0, "avg_logprob": -0.06764050646945163, "compression_ratio": 1.1652173913043478, "no_speech_prob": 1.6936098357067664e-11}, {"id": 17, "seek": 27116, "start": 271.16, "end": 285.16, "text": "转成 Mark it down 后,这些关系很容易被压平成段落。所以,这期的结论不是选一个神工具。如果你只是快速读普通 PDF,Mark it down 很适合做第一刀。", "tokens": [50365, 17819, 105, 11336, 3934, 309, 760, 220, 13547, 11, 5562, 13824, 28053, 25368, 4563, 49212, 23238, 5014, 233, 16716, 11336, 28427, 30848, 1543, 7239, 11, 5562, 16786, 1546, 45641, 7422, 118, 7296, 2215, 231, 20182, 20432, 23323, 39806, 1543, 45669, 36859, 10251, 31217, 5233, 119, 29993, 19550, 17752, 11, 15168, 309, 760, 26029, 2215, 224, 14245, 10907, 18049, 2437, 222, 1543, 51065], "temperature": 0, "avg_logprob": -0.06942011958868928, "compression_ratio": 1.2207357859531773, "no_speech_prob": 1.2561636626717831e-11}, {"id": 18, "seek": 27116, "start": 285.16, "end": 299.16, "text": "如果你要结构、表格和知识库候选,Darklin 更稳。如果你确认 PDF 有文本层,而且不需要 OCR,Marker 的快速配置表现很好,但一定要著名边界。", "tokens": [51065, 13119, 32085, 45641, 7360, 226, 1231, 17571, 30921, 12565, 6498, 5233, 228, 6346, 241, 11238, 2215, 231, 11, 35, 809, 5045, 220, 19002, 10415, 111, 1543, 45669, 38114, 106, 7422, 97, 17752, 21461, 17174, 8802, 9636, 224, 11, 22942, 1960, 35748, 422, 18547, 11, 15168, 260, 27949, 10251, 31217, 38846, 34719, 17571, 20204, 23801, 11, 8395, 48161, 19382, 15940, 40503, 21185, 1543, 51765], "temperature": 0, "avg_logprob": -0.06942011958868928, "compression_ratio": 1.2207357859531773, "no_speech_prob": 1.2561636626717831e-11}, {"id": 19, "seek": 29916, "start": 299.16, "end": 308.16, "text": "如果你面对扫描件或复杂版面,Miner U值得保留,不过要接受更重的依赖和参数成本。", "tokens": [50365, 45669, 8833, 8713, 3416, 104, 11673, 237, 20485, 19780, 1787, 235, 4422, 224, 42096, 8833, 11, 44, 4564, 624, 40242, 5916, 24302, 24456, 11, 1960, 16866, 4275, 14468, 23151, 19002, 12624, 1546, 3254, 251, 5266, 244, 12565, 2129, 224, 33188, 11336, 8802, 1543, 50815], "temperature": 0, "avg_logprob": -0.07451046915615306, "compression_ratio": 1.205992509363296, "no_speech_prob": 1.6886452305908684e-11}, {"id": 20, "seek": 29916, "start": 308.16, "end": 325.16, "text": "真正可靠的流程,是 PDF 入库前做四个检查,有没有文本层,阅读顺序有没有乱,表格行列有没有保住,图片,公式,扫描页有没有抽样。工具不是答案,质检流程才是答案。", "tokens": [50815, 6303, 15789, 4429, 5363, 254, 1546, 27854, 29649, 11, 1541, 17752, 220, 14028, 6346, 241, 8945, 10907, 19425, 7549, 16407, 222, 42623, 11, 2412, 17944, 17174, 8802, 9636, 224, 11, 10034, 227, 5233, 119, 10178, 118, 6346, 237, 2412, 17944, 2930, 109, 11, 17571, 30921, 8082, 43338, 2412, 17944, 24302, 21632, 11, 3919, 122, 16668, 11, 13545, 27584, 11, 3416, 104, 11673, 237, 10178, 113, 2412, 17944, 46022, 14496, 1543, 23323, 39806, 7296, 28223, 28899, 11, 18464, 101, 16407, 222, 27854, 29649, 18888, 1541, 28223, 28899, 1543, 51665], "temperature": 0, "avg_logprob": -0.07451046915615306, "compression_ratio": 1.205992509363296, "no_speech_prob": 1.6886452305908684e-11}, {"id": 21, "seek": 32516, "start": 325.16, "end": 332.16, "text": "下一期如果继续,我们可以把这套测评变成批量 PDF 入库工作流。", "tokens": [50365, 45271, 16786, 13119, 10115, 100, 10115, 255, 11, 15003, 6723, 16075, 5562, 1881, 245, 11038, 233, 5233, 226, 2129, 246, 11336, 3416, 117, 26748, 17752, 220, 14028, 6346, 241, 41315, 27854, 1543, 50715], "temperature": 0, "avg_logprob": -0.1937423732182751, "compression_ratio": 1.0497237569060773, "no_speech_prob": 1.928311156584872e-11}, {"id": 22, "seek": 32516, "start": 332.16, "end": 338.16, "text": "不仅看 Mark it down 好不好看,还看尽 Rack hole 能不能引用,能不能追溯,能不能查错。", "tokens": [50715, 1960, 1550, 227, 4200, 3934, 309, 760, 12202, 15769, 4200, 11, 14852, 4200, 1530, 121, 497, 501, 5458, 220, 8225, 28590, 41301, 9254, 11, 8225, 28590, 39947, 16084, 107, 11, 8225, 28590, 42623, 29900, 1543, 51015], "temperature": 0, "avg_logprob": -0.1937423732182751, "compression_ratio": 1.0497237569060773, "no_speech_prob": 1.928311156584872e-11}, {"id": 23, "seek": 33816, "start": 338.16, "end": 339.16, "text": "不仅看 Mark it down 。", "tokens": [50365, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 50415], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}, {"id": 24, "seek": 33816, "start": 339.16, "end": 340.16, "text": "不仅看 Mark it down 。", "tokens": [50415, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 50465], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}, {"id": 25, "seek": 33816, "start": 340.16, "end": 341.16, "text": "不仅看 Mark it down 。", "tokens": [50465, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 50515], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}, {"id": 26, "seek": 33816, "start": 341.16, "end": 342.16, "text": "不仅看 Mark it down 。", "tokens": [50515, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 50565], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}, {"id": 27, "seek": 33816, "start": 342.16, "end": 343.16, "text": "不仅看 Mark it down 。", "tokens": [50565, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 50615], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}, {"id": 28, "seek": 33816, "start": 343.16, "end": 344.16, "text": "不仅看 Mark it down 。", "tokens": [50615, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 50665], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}, {"id": 29, "seek": 33816, "start": 344.16, "end": 345.16, "text": "不仅看 Mark it down 。", "tokens": [50665, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 50715], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}, {"id": 30, "seek": 33816, "start": 345.16, "end": 346.16, "text": "不仅看 Mark it down 。", "tokens": [50715, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 50765], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}, {"id": 31, "seek": 33816, "start": 346.16, "end": 347.16, "text": "不仅看 Mark it down 。", "tokens": [50765, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 50815], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}, {"id": 32, "seek": 33816, "start": 347.16, "end": 348.16, "text": "不仅看 Mark it down 。", "tokens": [50815, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 50865], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}, {"id": 33, "seek": 33816, "start": 348.16, "end": 349.16, "text": "不仅看 Mark it down 。", "tokens": [50865, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 50915], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}, {"id": 34, "seek": 33816, "start": 349.16, "end": 350.16, "text": "不仅看 Mark it down 。", "tokens": [50915, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 50965], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}, {"id": 35, "seek": 33816, "start": 350.16, "end": 351.16, "text": "不仅看 Mark it down 。", "tokens": [50965, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 51015], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}, {"id": 36, "seek": 33816, "start": 351.16, "end": 352.16, "text": "不仅看 Mark it down 。", "tokens": [51015, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 51065], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}, {"id": 37, "seek": 33816, "start": 352.16, "end": 353.16, "text": "不仅看 Mark it down 。", "tokens": [51065, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 51115], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}, {"id": 38, "seek": 33816, "start": 353.16, "end": 354.16, "text": "不仅看 Mark it down 。", "tokens": [51115, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 51165], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}, {"id": 39, "seek": 33816, "start": 354.16, "end": 355.16, "text": "不仅看 Mark it down 。", "tokens": [51165, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 51215], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}, {"id": 40, "seek": 33816, "start": 355.16, "end": 356.16, "text": "不仅看 Mark it down 。", "tokens": [51215, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 51265], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}, {"id": 41, "seek": 33816, "start": 356.16, "end": 357.16, "text": "不仅看 Mark it down 。", "tokens": [51265, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 51315], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}, {"id": 42, "seek": 33816, "start": 357.16, "end": 358.16, "text": "不仅看 Mark it down 。", "tokens": [51315, 1960, 1550, 227, 4200, 3934, 309, 760, 220, 1543, 51365], "temperature": 0, "avg_logprob": -0.1769937223142332, "compression_ratio": 13.0, "no_speech_prob": 2.3702631871125313e-11}], "language": "zh"}