5.1 缓存命中率
经过测试，我们的缓存命中率普遍在80%以上，平均命中率在85%以上。这将使您可以使用极少的token消耗来处理大文档。
5.2 生成耗时
在我们的测试中，平均每个文档大约需要100 秒左右的处理时间，超长文档（切片数量超过200 个）平均耗时在300 秒左右。
5.3 效果评测
由于正确切片和实际切片的之间的差异，我们无法直接判断切片是否召回，因此我们定义了一个token 级别的召回率，用于判断切片是否正确召回。当token召回率大于一定阈值（通常是0.8）时，我们将该切片视为正确召回。
chunk_recall = len (relevant_token_set & retrieved_token_set) / len (relevant_token_set)
检索召回率（Recall）：相关片段被成功召回的比例，衡量系统覆盖相关内容的能力。 检索精准率（Precision）：检索片段中命中相关片段的比例，衡量系统返回结果的纯度。 F1 Score：Precision 与 Recall 的调和平均，用于在两者存在权衡时提供一个综合评价指标，尤其适用于正负样本分布不均衡的检索与分类任务。 端到端语义一致性准确率（End-to-End Semantic Accuracy）：模型基于检索结果生成的最终答案，与参考答案在语义层面的等价性或一致性比例。通常通过语义相似度模型、判别式评估器或人工标注判断，用于衡量完整 RAG 流程（检索 + 生成）的真实回答质量。
我们分别在三个场景下做了评测，学术论文场景、车书问答场景、保险条款场景，每一个场景都有不同程度的提升。
文档类型 基准召回率 上下文增强召回率 提升幅度 失败率下降 学术论文 33% 56.25% +23.25% -34.75% 车书问答 68% 80% +12% -37.5% 保险条款 53.60% 71.43% +17.83% -38.4%
后续我们将会在更多评测集上进行评测，相关评测数据会更新到文档中来，并形成一个上下文增强的最佳实践。
5.4 生成样例
• 样例一：学术论文
1.2.3.5 基于机器学习的轨迹规划方法 模型预测生成方法根据采样轨迹生成的模型预测生成轨迹。图4-12(b)为采样轨迹和模型预测生成轨迹的曲率变化率比较。采样轨迹曲率变化率反复振荡，最大 值为0.004 0 m-2,最小为 -0.004 2 m-2，标准差为2.04×10-3 m-2。模型预测生成方法轨迹的曲率变化率较为平缓，最大值为0.002 1 m-2,最小为 -0.001 5 m-2，标准差为1.34×10-3 m-2。采样轨迹的曲率变化率振荡明显，影响车辆行驶的舒适性。 基于模型优化的方法的曲率变化率标准差相对于采样轨迹降低了34.3%，显著降 低了轨迹的曲率波动，提升舒适性。 基于采样方法的轨迹曲率变化率反复震荡，主要原因是采样依据S-L 坐标系 中心参考线进行。采样轨迹的横向偏移虽然是多项式，但是依赖于中心参考线的 平滑程度。受限于真实高精地图的中心参考线曲率，采样轨迹的曲率可能发生突 变和振荡。相比之下，模型生成轨迹的曲率平滑，主要原因是模型轨迹基于笛卡尔坐标系和客车车辆运动模型生成，提高了舒适性。 该案例测试说明，本文提出的轨迹规划框架可以提供一条安全、舒适的轨迹， 可提供给下层的轨迹跟踪模块以供控制。 (a) 多阶段规划轨迹结果 图4-12 采样和优化轨迹对比 
该片段出自"《21-周扬-硕士论文-自动驾驶客车结构化道路高实时性轨迹规划算法研究》第四章高实时性动作规划算法 与实验研究"，对比了采样轨迹与模型预测生成轨迹的曲率变化率，指出模型优化方法使曲率变化率标准差降低34.3%， 显著提升了轨迹舒适性，原因是模型轨迹基于笛卡尔坐标系和车辆运动模型生成，而非依赖S-L坐标系中心参考线。
• 样例二：金融报告
1.500% Notes due 2031MMM31New York Stock Exchange, Inc. Note: The common stock of the registrant is also traded on the SWX Swiss Exchange. Securities registered pursuant to section 12(g) of the Act: None Indicate by check mark if the registrant is a well-known seasoned issuer, as defined in Rule 405 of the Securities Act. Yes ☒ No ☐ Indicate by check mark if the registrant is not required to file reports pursuant to Section 13 or Section 15(d) of the Act. Yes ☐ No ☒ Indicate by check mark whether the registrant (1) has filed all reports required to be filed by Section 13 or 15(d) of the Securities Exchange Act of 1934 during thepreceding 12 months (or for such shorter period that the registrant was required to file such reports), and (2) has been subject to such filing requirements for the past 90 days. Yes ☒ No ☐ Indicate by check mark whether the registrant has submitted electronically every Interactive Data File required to be submitted pursuant to Rule 405 of Regulation S-T (§ 232.405 of this chapter) during the preceding 12 months (or for such shorter period that the registrant was required to submit such files). Yes ☒ No ☐ Indicate by check mark whether the registrant is a large accelerated filer, an accelerated filer, a non-accelerated filer, a smaller reporting company, or an emerginggrowth company. See the definitions of “large accelerated filer,” “accelerated filer,” “smaller reporting company,” and “emerging growth company” in Rule 12b-2of the Exchange Act. Large accelerated filer ☒Accelerated filer ☐Non-accelerated filer ☐ Smaller reporting company ☐Emerging growth company ☐ If an emerging growth company, indicate by check mark if the registrant has elected not to use the extended transition period for complying with any new or revisedfinancial accounting standards provided pursuant to Section 13(a) of the Exchange Act. ☐ Indicate by check mark whether the registrant is a shell company (as defined in Rule 12b-2 of the Act). Yes ☐ No The aggregate market value of voting stock held by nonaffiliates of the registrant, computed by reference to the closing price and shares outstanding, was approximately \$91.3 billion as of January 31, 2020 (approximately \$ 99.7 billion as of June 30, 2019, the last business day of the registrant’s most recently completedsecond quarter). Shares of common stock outstanding at January 31, 2020: 575.6 million DOCUMENTS INCORPORATED BY REFERENCE Parts of the Company’s definitive proxy statement (to be filed pursuant to Regulation 14A within 120 days after Registrant’s fiscal year-end of December 31, 2019) for its annual meeting to be held on May 12, 2020, are incorporated by reference in this Form 10-K in response to Part III, Items10, 11, 12, 13 and 14. 3M COMPANY FORM 10-K For the Year Ended December 31, 2019 Pursuant to Part IV, Item 16, a summary of Form 10-K content follows, including hyperlinked cross-references (in the EDGAR filing). This allows users toeasily locate the corresponding items in Form 10-K, where the disclosure is fully presented. The summary does not include certain Part III information thatwill be incorporated by reference from the proxy statement, which will be filed after this Form 10-K filing. Beginning Page PART I ITEM 1Business4 ITEM 1ARisk Factors10 ITEM 1BUnresolved Staff Comments13 ITEM 2Properties13 ITEM 3Legal Proceedings13 ITEM 4Mine Safety Disclosures13 PART II ITEM 5Market for Registrant’s Common Equity, Related Stockholder Matters and Issuer Purchases of Equity Securities14 ITEM 6Selected Financial Data15 ITEM 7Management’s Discussion and Analysis of Financial Condition and Results of Operations16 MD&A is designed to provide a reader of 3M’s financial statements with a narrative from the perspective of
This chunk is from "3M 2019 Form 10-K Annual Report, Section: Securities Registration and Company Information." It lists 3M's 1.500% Notes due 2031 trading as MMM31 on the NYSE, indicates the common stock is also traded on the SWX Swiss Exchange, and provides regulatory compliance status including that 3M is a well-known seasoned issuer and large accelerated filer. The chunk also includes market value data of voting stock held by nonaffiliates (\$91.3 billion as of January 31, 2020) and shares outstanding information (575.6 million shares).