--- title: "Packaging & Context Handling — 封装与上下文处理" source: "https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86" wiki: "systems-analysis.info/int" article: "Packaging_&_Context_Handling_—_封装与上下文处理" language: "zh" categories: - "Category:Chinese" - "Category:Large language models" - "Category:Prompt engineering" revision_id: 5394 wiki_created_at: 2026-09-06T23:48:40Z wiki_modified_at: 2026-09-06T23:48:40Z downloaded_at: 2026-09-07T23:08:14Z --- # Packaging & Context Handling — 封装与上下文处理 **Packaging & Context Handling** — 一套在检索增强生成（Retrieval‑Augmented Generation, RAG）中，用于选择、压缩、布局和提供已检索知识片段到 LLM 上下文中的技术。其目标是最大化有限令牌预算的效用，提高答案的准确性和鲁棒性，并确保来源的可追溯引用。“封装”不仅指形成片段列表，还包括它们的压缩、排序、分组以及给模型的指令，包括 *stuff*、*map‑reduce*、*refine* 和 *tree‑of‑chunks* 等策略。^{[\[1\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-lewis2020-1)[\[2\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-langchain-sum-2)} ## 定义与动机在 RAG 系统中，最终答案的质量不仅取决于检索，还取决于所选片段*如何*被送入提示词（prompt）。上下文限制和令牌成本迫使我们在完整性和精确性之间进行权衡：多余的片段会增加“中间迷失”（*lost‑in‑the‑middle*）的风险并延长延迟，而激进的过滤/压缩可能会删除关键证据。^{[\[3\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-lostmiddle-3)} 在经典的 RAG 中，信息源充当外部“非参数化内存”，确保了时效性和可引用性，前提是封装能让 LLM 可靠地处理事实和引用。^{[\[1\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-lewis2020-1)} ## 分块与内容提取分块（*chunking*）策略决定了块（chunk）的大小/重叠、规范化和粒度（*document*→*passage*→*sentence*）。典型方法包括： - **固定大小规则**（按字符/令牌）并设置重叠，以保持块之间的连贯性；^{[\[4\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-lc-splitters-4)[\[5\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-llama-splitters-5)} - **语义分块**（根据嵌入向量的相似度确定边界），以减少“意义断裂”。^{[\[6\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-llama-sem-6)} - **句子窗口检索**（Sentence‑window retrieval）—— 首先对句子建立索引；检索时，提取相关句子及其前后的相邻句子作为*窗口*，以恢复局部上下文。^{[\[7\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-haystack-sentwin-7)} - **段落级索引**（Passage‑level indexing）—— 将维基百科分割成约100词的段落，这已成为开放域问答（DPR）的标准，表明细粒度对初始检索阶段很有帮助。^{[\[8\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-dpr-8)} - **规范化与清理**（删除垃圾内容、页眉/页脚、统一空格），并在元数据层面跟踪来源/页面/偏移量，以实现可追溯性。^{[\[9\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-haystack-pre-9)} - **去重**（Deduplication）候选块（精确和*近乎重复*）：使用 shingles + MinHash/LSH 来减少重复。^{[\[10\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-broder1997-10)} ## 多样化与选择 (MMR 等) 在构建上下文集合时，要求高相关性和低冗余性。经典的“最大边界相关性”（Maximal Marginal Relevance）函数在选择下一个片段时，会同时考虑其与查询的相似度以及与*已选片段*的最大相似度（即对重复内容进行惩罚）： ${MMR}(d_{i}) = \arg\max\limits_{d_{i} \in D \smallsetminus S}\left\lbrack \lambda \cdot {sim}(q,d_{i}) - (1 - \lambda) \cdot \max\limits_{d_{j} \in S}{sim}(d_{i},d_{j}) \right\rbrack,\ \lambda \in \lbrack 0,1\rbrack$。^{[\[11\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-mmr-11)} 信号组合：混合检索（BM25 + dense）→ 融合（例如，倒数排序融合，Reciprocal Rank Fusion, RRF）→ 使用交叉编码器（cross‑encoder）/ColBERT 进行重排： - **RRF**: 一种简单有效的无监督方案，用于合并不同检索器的排名结果。^{[\[12\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-rrf-12)} - **交叉编码器重排器**（BERT/MonoT5/现代商业 API）能显著提高 top‑k 的准确性，但会增加延迟。^{[\[13\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-nogueira2019-13)[\[14\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-cohere-rerank-14)} - **多向量检索器** ColBERT（*late interaction*）在大型语料库上通常作为高效的一级重排器/检索器。^{[\[15\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-colbert-15)} - **混合搜索**（BM25F+向量）已在工业级搜索引擎和库中实现，并带有可配置的权重/融合方法（alpha、RRF 等）。^{[\[16\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-weav-hybrid-16)[\[17\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-pine-hybrid-17)} ## 上下文压缩在不丢失事实的情况下减少上下文大小，对成本和延迟至关重要： - **抽取式压缩**（提取关键句子/短语）；**生成式摘要**（改写/压缩）。经典观点参见 Nenkova & McKeown。^{[\[18\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-nenkova2011-18)} - **查询导向/指令导向**压缩：根据查询/任务进行总结（突出证据并删除不相关内容）。 - **提示词/上下文压缩**（Prompt/context compression）通过 LLM 过滤/剪枝令牌（例如 LLMLingua/LLMLingua‑2）来减少令牌预算，同时质量损失较小，但需要仔细验证其*忠实性*。^{[\[19\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-llmlingua-19)} - 压缩是*质量↔成本↔延迟*之间的权衡：激进的压缩会增加忽略细微差别/前提的风险，并降低事实归因的准确性。^{[\[20\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-ji2023-20)} ## 封装策略 (stuff/map-reduce/refine/tree) 以下是在提示词中组织信息源的四种基本方案及其典型应用场景（另见对比表）。 **Stuff** (直接填充) 将选定的片段（可能经过压缩）拼接在一起，然后一次性全部送入模型。这种方法简单快捷，但受上下文长度限制，且在长输入上容易出现*中间迷失*问题。^{[\[2\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-langchain-sum-2)} **Map‑Reduce** 在 *map* 阶段，针对每个片段/文档进行局部回答/总结；然后在 *reduce* 阶段进行聚合（比较、投票、合并）。这种方法在处理大量信息源时扩展性好，可降低单个提示词的负担；但如果聚合方式过于简单，可能会丢失信息源之间的交叉关联。^{[\[2\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-langchain-sum-2)[\[21\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-fid-21)} **Refine** 逐步改进：根据第一个片段生成初始答案，然后结合下一个片段进行迭代式*改进*（补充/修正）。当信息源的顺序很重要时，这种方法很方便；但存在“固着”于早期错误并累积偏差的风险。^{[\[22\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-llama-refine-22)} **Tree‑of‑chunks** 分层压缩/摘要：先对块进行局部摘要 → 再在章节层面进行归纳 → 最后生成最终摘要。这种方法对长文档很有用；但需要确保在不同层级间准确传递源标识符，以便正确归因。^{[\[23\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-llama-tree-23)} | 策略 | 思路 | 成本/延迟 | 上下文丢失风险 | 适用场景 | 来源 | |--------------------|--------------------------------|----------------------|------------------------------------------|---------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | **Stuff** | 将所有片段一次性放入一个提示词 | 低（在上下文限制内） | 在长输入上风险高（*lost‑in‑the‑middle*） | 内容量小，问题简单 | ^{[\[2\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-langchain-sum-2)[\[3\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-lostmiddle-3)} | | **Map‑Reduce** | 局部回答 → 聚合 | 中/高（多次调用） | 中（取决于 reduce 阶段的质量） | 信息源多，需要可扩展性 | ^{[\[2\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-langchain-sum-2)[\[21\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-fid-21)} | | **Refine** | 逐步改进答案 | 中 | 依赖顺序，有固化错误的风险 | 当答案的顺序/演进很重要时 | ^{[\[22\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-llama-refine-22)} | | **Tree‑of‑chunks** | 分层摘要 | 中/高 | 在高层级会丢失细节 | 长文档/合集 | ^{[\[23\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-llama-tree-23)} | 封装策略对比 ## 信息源的排序与定位 LLM 对长上下文中部信息的利用效率较低；有用的事实最好放在开头/结尾，按主题/来源分组，并用标题和 ID 标记。考虑到*查询感知*的重要性和多样性进行重排，有助于将关键片段移到更靠前的位置。^{[\[3\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-lostmiddle-3)[\[13\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-nogueira2019-13)} ## RAG 流水线集成 (fusion → rerank → packaging) 典型的多阶段流水线：**混合检索**（BM25 + dense）→ **融合**（RRF/加权混合）→ **重排**（Cross‑Encoder/ColBERT）→ **封装**（采用其中一种策略）→ **生成** + **引用**。混合搜索和 RRF 对不同检索器得分不一致的情况具有鲁棒性；交叉编码器则能提高输入 LLM 的精确度，从而节省令牌。^{[\[16\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-weav-hybrid-16)[\[12\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-rrf-12)[\[14\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-cohere-rerank-14)[\[15\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-colbert-15)} ## 质量评估与消融研究评估在检索、封装和生成等多个层面进行： - **检索**: Recall@k、nDCG@k、MRR — 信息检索领域的标准指标。^{[\[24\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-manning2008-24)} - **忠实性/有据性**（Faithfulness/groundedness）: 由引文支持的陈述比例；使用自动化框架（RAGAS, TruLens）+ 手动验证归因。^{[\[25\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-ragas-25)[\[26\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-trulens-26)[\[27\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-rashkin2023-27)} - **端到端问答**: EM/F1/ROUGE，具体取决于任务/数据集（NQ/HotpotQA 等）。^{[\[1\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-lewis2020-1)} - **效率**: 延迟 p50/p95、令牌数、美元成本；比较不同封装策略和压缩级别下的*质量↔成本*。 - **消融研究**（Ablations）: 通过关闭 MMR/去重/压缩/改变顺序来衡量每个组件的贡献（截至 2025‑09‑10，RAG 研究实践建议明确记录 k 值、λ、块大小和令牌限制）。^{[\[28\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-rag-survey-28)} ## 实践建议与清单 - **k值与多样化**: 从混合检索中获取 k=20–40 个候选片段开始；应用 MMR，λ≈0.5–0.8；对 URL/ID/文本哈希相同的重复内容进行严格惩罚。^{[\[11\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-mmr-11)[\[16\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-weav-hybrid-16)} - **分块**: 固定大小分块时，使用 200–400 个令牌，重叠 10–20%；对于法律技术类文档，句子/窗口方案通常更有效。^{[\[4\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-lc-splitters-4)[\[7\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-haystack-sentwin-7)} - **压缩**: 使用基于查询的抽取式过滤和谨慎的生成式方法；根据数据上的*忠实度*评估结果，增减 LLMLingua 类方法的使用（必须进行 A/B 验证）。^{[\[19\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-llmlingua-19)[\[27\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-rashkin2023-27)} - **顺序**: 将重要/高置信度的片段放在提示词的开头；按来源/主题分组，并明确标注 ID 和标题；注意*中间迷失*效应（在开头和结尾重复关键事实可能会有帮助）。^{[\[3\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-lostmiddle-3)} - **重排**: 如果预算允许，在封装前对 top-k（k≈50–200）使用 Cross‑Encoder/ColBERT，这可以节省生成令牌并提高准确性。^{[\[13\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-nogueira2019-13)[\[15\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-colbert-15)} - **回退策略**（Fallback）: (1) 事实不足 → 请求额外来源；(2) 超过令牌限制 → 从 *stuff* 切换到 *refine* 或启用压缩；(3) 置信度低/存在矛盾 → 拒绝回答，并明确列出缺失的 ID（见下文模板）。 ### 封装流水线伪代码 # 输入: 查询 q cands = retrieve(q, K_sparse, K_dense) # BM25、DPR 等检索 cands = diversify_MMR(cands, lambda=0.7) # 多样化 (MMR) snips = compress(query=q, items=cands, mode="extractive|abstractive", budget=tokens) pkg = package(snips, strategy="stuff|map_reduce|refine|tree") resp = generate(prompt=build_prompt(q, pkg), citations=True) # 带引用的 LLM ### 提示词模板骨架（片段） [用户查询] {q} [来源] {# 每个片段都包含 ID、标题和链接 #} - [{id}] {title} — {url} {content_snippet} [要求] 1) 仅使用来源中的事实，并引用 [ID]。 2) 如果数据不足，请说明情况并请求澄清或提供额外来源。 3) 保持回答结构，并列出所有使用的 [ID]。 ## 局限性与开放性问题 - **map-reduce/refine 中的幻觉与聚合**: 生成式摘要可能会引入新的事实；明确的归因指令和引文验证机制至关重要。^{[\[20\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-ji2023-20)[\[27\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-rashkin2023-27)} - **激进压缩/分层压缩导致的细节丢失**: 保持到原始来源/页面/偏移量的反向链接很重要。 - **检索器/重排器和压缩器的领域可移植性**: 需要在特定领域的语料库上进行适配/微调。^{[\[28\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-rag-survey-28)} - **隐私/PII 与 LLM 的记忆效应**: 在没有严格*依据*（grounding）的情况下生成内容可能导致私密字符串泄露；应使用过滤器、私有存储和拒绝策略。^{[\[29\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-carlini-29)[\[30\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-shokri-30)} - **可训练的“封装器”**、自适应排序/布局、用于提高*忠实度*的 RLHF/反馈循环、多语言和超长上下文是当前活跃的研究方向。^{[\[28\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-rag-survey-28)[\[3\]](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_note-lostmiddle-3)} ## 外部链接 - LangChain: *Summarization* (stuff/map_reduce/refine). [29] - LangChain: *Text splitters*. [30] - LlamaIndex: *Response Synthesizers (refine/tree)*. [31] - LlamaIndex: *Node Parsers / SentenceSplitter / SemanticSplitter*. [32] - Haystack: *SentenceWindowRetriever*. [33] - Haystack: *PreProcessors / DocumentSplitter*. [34] - Weaviate: *Hybrid search*. [35] - Pinecone: *Hybrid search*. [36] - Cohere: *Rerank API*. [37] - RAGAS (repo/docs). [38] [39] - TruLens (docs). [40] ## 参考文献 - Manning, C. D., Raghavan, P., Schütze, H. (2008). *Introduction to Information Retrieval*. Cambridge University Press. ISBN 978‑0521865715. - Nenkova, A., McKeown, K. (2011). *Automatic Summarization*. FnT IR, 5(2–3), 103–233. DOI:10.1561/1500000015. - Lewis, P., et al. (2020). *Retrieval‑Augmented Generation for Knowledge‑Intensive NLP Tasks*. NeurIPS. arXiv:2005.11401. - Khattab, O., Zaharia, M. (2020). *ColBERT*. SIGIR’20. DOI:10.1145/3397271.3401075. - Izacard, G., Grave, E. (2021). *Fusion‑in‑Decoder*. EACL. arXiv:2007.01282. - Ji, Z., et al. (2023). *Survey of Hallucination in NLG*. ACM CS. DOI:10.1145/3571730. - Gao, S., et al. (2024). *RAG for LLM: A Survey*. arXiv:2312.10997. ## 注释 1. ↑ ^{[1.0](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-lewis2020_1-0)} ^{[1.1](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-lewis2020_1-1)} ^{[1.2](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-lewis2020_1-2)} Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., et al. (2020). *Retrieval‑Augmented Generation for Knowledge‑Intensive NLP Tasks*. NeurIPS. arXiv:2005.11401. [1] 2. ↑ ^{[2.0](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-langchain-sum_2-0)} ^{[2.1](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-langchain-sum_2-1)} ^{[2.2](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-langchain-sum_2-2)} ^{[2.3](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-langchain-sum_2-3)} ^{[2.4](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-langchain-sum_2-4)} LangChain Docs. *Summarization* (stuff/map_reduce/refine/map_rerank). (доступ: 2025‑09‑10). [2] 3. ↑ ^{[3.0](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-lostmiddle_3-0)} ^{[3.1](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-lostmiddle_3-1)} ^{[3.2](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-lostmiddle_3-2)} ^{[3.3](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-lostmiddle_3-3)} ^{[3.4](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-lostmiddle_3-4)} Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., Liang, P. (2024). *Lost in the Middle: How Language Models Use Long Contexts*. TACL. arXiv:2307.03172. [3] 4. ↑ ^{[4.0](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-lc-splitters_4-0)} ^{[4.1](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-lc-splitters_4-1)} LangChain Docs. *Text splitters* (RecursiveCharacter/TokenTextSplitter). (доступ: 2025‑09‑10). [4] 5. [↑](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-llama-splitters_5-0) LlamaIndex Docs. *SentenceSplitter / TokenTextSplitter / SemanticSplitter*. (доступ: 2025‑09‑10). [5] [6] [7] 6. [↑](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-llama-sem_6-0) LlamaIndex Docs. *SemanticSplitterNodeParser*. (доступ: 2025‑09‑10). [8] 7. ↑ ^{[7.0](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-haystack-sentwin_7-0)} ^{[7.1](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-haystack-sentwin_7-1)} Haystack Docs. *SentenceWindowRetriever*. (доступ: 2025‑09‑10). [9] 8. [↑](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-dpr_8-0) Karpukhin, V., Oguz, B., Min, S., Lewis, P., Wu, L., Edunov, S., Chen, D., Yih, W.‑T. (2020). *Dense Passage Retrieval for Open‑Domain Question Answering*. EMNLP. arXiv:2004.04906. [10] 9. [↑](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-haystack-pre_9-0) Haystack Docs. *PreProcessors / DocumentSplitter*. (доступ: 2025‑09‑10). [11] [12] 10. [↑](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-broder1997_10-0) Broder, A. Z. (1997). *On the Resemblance and Containment of Documents*. Compression and Complexity of Sequences. [13] 11. ↑ ^{[11.0](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-mmr_11-0)} ^{[11.1](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-mmr_11-1)} Carbonell, J., Goldstein, J. (1998). *The Use of MMR, Diversity‑Based Reranking for Reordering Documents and Producing Summaries*. SIGIR’98, pp. 335–336. DOI:10.1145/290941.291025. 12. ↑ ^{[12.0](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-rrf_12-0)} ^{[12.1](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-rrf_12-1)} Cormack, G. V., Clarke, C. L. A., Büttcher, S. (2009). *Reciprocal Rank Fusion outperforms Condorcet and Individual Rank Learning Methods*. SIGIR’09, pp. 758–759. DOI:10.1145/1571941.1572114. [14] 13. ↑ ^{[13.0](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-nogueira2019_13-0)} ^{[13.1](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-nogueira2019_13-1)} ^{[13.2](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-nogueira2019_13-2)} Nogueira, R., Cho, K. (2019). *Passage Re‑ranking with BERT*. arXiv:1901.04085. [15] 14. ↑ ^{[14.0](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-cohere-rerank_14-0)} ^{[14.1](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-cohere-rerank_14-1)} Cohere Docs. *Rerank API overview*. (доступ: 2025‑09‑10). [16] 15. ↑ ^{[15.0](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-colbert_15-0)} ^{[15.1](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-colbert_15-1)} ^{[15.2](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-colbert_15-2)} Khattab, O., Zaharia, M. (2020). *ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT*. SIGIR’20, pp. 39–48. DOI:10.1145/3397271.3401075. arXiv:2004.12832. 16. ↑ ^{[16.0](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-weav-hybrid_16-0)} ^{[16.1](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-weav-hybrid_16-1)} ^{[16.2](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-weav-hybrid_16-2)} Weaviate Docs. *Hybrid search (BM25F + vector)*. (доступ: 2025‑09‑10). [17] 17. [↑](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-pine-hybrid_17-0) Pinecone Docs. *Hybrid search*. (доступ: 2025‑09‑10). [18] 18. [↑](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-nenkova2011_18-0) Nenkova, A., McKeown, K. (2011). *Automatic Summarization*. Foundations and Trends in Information Retrieval, 5(2–3), 103–233. DOI:10.1561/1500000015. 19. ↑ ^{[19.0](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-llmlingua_19-0)} ^{[19.1](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-llmlingua_19-1)} Zhu, Y., Shao, Z., Li, M., et al. (2023). *LLMLingua: Compressing Prompts for Accelerating LLM Inference*. arXiv:2310.05736. [19] 20. ↑ ^{[20.0](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-ji2023_20-0)} ^{[20.1](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-ji2023_20-1)} Ji, Z., Lee, N., Frieske, R., et al. (2023). *Survey of Hallucination in Natural Language Generation*. ACM Computing Surveys, 55(12), Art.248. DOI:10.1145/3571730. 21. ↑ ^{[21.0](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-fid_21-0)} ^{[21.1](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-fid_21-1)} Izacard, G., Grave, E. (2021). *Leveraging Passage Retrieval with Generative Models for Open‑Domain QA (Fusion‑in‑Decoder)*. EACL. arXiv:2007.01282. [20] 22. ↑ ^{[22.0](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-llama-refine_22-0)} ^{[22.1](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-llama-refine_22-1)} LlamaIndex Docs. *Response Synthesizers: refine*. (доступ: 2025‑09‑10). [21] 23. ↑ ^{[23.0](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-llama-tree_23-0)} ^{[23.1](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-llama-tree_23-1)} LlamaIndex Docs. *Tree Summarize*. (доступ: 2025‑09‑10). [22] 24. [↑](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-manning2008_24-0) Manning, C. D., Raghavan, P., Schütze, H. (2008). *Introduction to Information Retrieval*. Cambridge Univ. Press. (см. главы о nDCG/MRR). [23] 25. [↑](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-ragas_25-0) Es, S., et al. (2023). *RAGAS: Automated Evaluation of Retrieval‑Augmented Generation*. arXiv:2309.15217. [24] 26. [↑](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-trulens_26-0) TruLens Docs. *Evaluating RAG (groundedness, relevance)*. (доступ: 2025‑09‑10). [25] 27. ↑ ^{[27.0](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-rashkin2023_27-0)} ^{[27.1](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-rashkin2023_27-1)} ^{[27.2](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-rashkin2023_27-2)} Rashkin, H., Nakov, P., et al. (2023). *Measuring Attribution in Natural Language Generation*. Computational Linguistics, 49(4), 1207–1261. DOI:10.1162/coli_a_00486. 28. ↑ ^{[28.0](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-rag-survey_28-0)} ^{[28.1](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-rag-survey_28-1)} ^{[28.2](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-rag-survey_28-2)} Gao, S., et al. (2024). *Retrieval‑Augmented Generation for Large Language Models: A Survey*. arXiv:2312.10997. [26] 29. [↑](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-carlini_29-0) Carlini, N., Tramèr, F., et al. (2021). *Extracting Training Data from Large Language Models*. USENIX Security. [27] 30. [↑](https://systems-analysis.info/int/Packaging_%26_Context_Handling_%E2%80%94_%E5%B0%81%E8%A3%85%E4%B8%8E%E4%B8%8A%E4%B8%8B%E6%96%87%E5%A4%84%E7%90%86#cite_ref-shokri_30-0) Shokri, R., Stronati, M., Song, C., Shmatikov, V. (2017). *Membership Inference Attacks Against ML Models*. IEEE S&P. [28]