# 吴师兄大模型(www.wushixiongai.com)

> 围绕 RAG、Agent、推理部署等工程项目拆解面试题：先给 30 秒回答，再讲清完整工程过程，并继续追问职责、选型、指标、失败案例和上线过程。另保留 851 道基础知识题、项目教程，以及公开分享与教学复原形式的面经。作者吴师兄（GitHub MisterBooo，LeetCodeAnimation 作者）。

## 核心入口
- [大模型项目面试题](https://www.wushixiongai.com/questions/project)
- [RAG 项目面试题](https://www.wushixiongai.com/questions/project/rag)：从文档解析、Chunk、混合检索、Rerank、评测与上线出发，把一条 RAG 工程链路讲清楚，再练技术选型、失败排查和效果验证。
- [Deep Research 项目面试题](https://www.wushixiongai.com/questions/project/deep-research)：围绕一个会规划、检索、阅读、校验并生成带引用报告的研究 Agent，深入练长任务设计、工具可靠性、训练方法和评估闭环。
- [Agent 项目面试题](https://www.wushixiongai.com/questions/project/agent)：从一个真正会规划、调用工具、保存状态并处理失败的 Agent 出发，深入练架构取舍、状态恢复、安全边界和多智能体协作。
- [大模型应用开发学习路线图（2026 版，四阶段）](https://www.wushixiongai.com/roadmap)
- [RAG 面试题大全与知识体系](https://www.wushixiongai.com/rag)
- [Agent 面试题大全与开发指南](https://www.wushixiongai.com/agent)
- [转行大模型应用开发指南（路线图 / 岗位选择 / 面试全解）](https://www.wushixiongai.com/guide)
- [按公司查大模型面试题](https://www.wushixiongai.com/companies)
- [主题专题矩阵](https://www.wushixiongai.com/questions)
- [大模型评估与监控面试题大全](https://www.wushixiongai.com/evaluation)
- [Prompt 工程面试题大全](https://www.wushixiongai.com/prompt-engineering)
- [重排与检索优化面试题大全](https://www.wushixiongai.com/rerank)
- [大模型微调与训练面试题大全](https://www.wushixiongai.com/finetune)
- [Transformer 与模型架构面试题大全](https://www.wushixiongai.com/transformer)
- [向量检索面试题大全](https://www.wushixiongai.com/vector-search)
- [知识图谱与 GraphRAG 面试题大全](https://www.wushixiongai.com/knowledge-graph)
- [多模态大模型面试题大全](https://www.wushixiongai.com/multimodal)
- [面试题库（851 道）](https://www.wushixiongai.com/questions)
- [RAG 智能问答系统完整教程](https://www.wushixiongai.com/projects/rag-system)
- [Deep Research Agent 完整教程](https://www.wushixiongai.com/projects/deep-research)
- [AI 编程 · Agent 实操](https://www.wushixiongai.com/ai-coding)
- [面经与教学复原](https://www.wushixiongai.com/mianjing)
- [大模型简历项目模拟面试](https://www.wushixiongai.com/resume)
- [吴师兄的大模型项目与公开作品](https://www.wushixiongai.com/evidence)
- [大模型面试题图解 100（GitHub 公开资料）](https://github.com/MisterBooo/llm-interview-questions)
- [关于网站与作者](https://www.wushixiongai.com/about)

## 项目面试题（已发布 75 篇）
> 每篇包含 30 秒回答、完整工程解释、连续追问与项目准备清单。
- [RAG 已经能搜 PDF，为什么视频不能直接入库？](https://www.wushixiongai.com/questions/project/rag/video-ingestion)：视频文件能上传，不等于内容能检索。先把声音与画面变成可搜索、可定位、可核验的证据片段。
- [PDF、PPT、扫描件为什么不能共用一条解析 Pipeline？](https://www.wushixiongai.com/questions/project/rag/multi-format-parsing-pipeline)：统一的应该是下游数据结构，不是上游解析方法。格式路由做错，后面的 Embedding 只会索引错误内容。
- [扫描 PDF 没有文本层，RAG 怎么做 OCR 与失败降级？](https://www.wushixiongai.com/questions/project/rag/scanned-pdf-ocr-pipeline)：OCR 读出字只是起点；阅读顺序、表格结构、坐标映射和失败页隔离共同决定入库质量。
- [复杂 PDF 的双栏、表格和图片怎么解析，才不会打乱语义？](https://www.wushixiongai.com/questions/project/rag/complex-pdf-layout-tables)：文字都提取出来仍可能是错的。RAG 需要的是结构正确的内容，而不是字数最多的一段文本。
- [MinerU 能直接解决 RAG 文档解析吗？哪些页面必须二次处理？](https://www.wushixiongai.com/questions/project/rag/mineru-post-processing)：开源解析器是基线，不是质量承诺。先识别失败类型，再决定专用模型、规则修复或人工队列。
- [RAG 文档解析质量怎么评估？不能只看“解析成功”](https://www.wushixiongai.com/questions/project/rag/parsing-quality-evaluation)：接口返回 200 只说明程序没崩。解析是否可用，要看内容能否被正确检索、理解和回到原文。
- [RAG Chunk 为什么不能只按固定字数切？](https://www.wushixiongai.com/questions/project/rag/chunk-semantic-boundaries)：切分目标不是让每块一样长，而是让检索命中一个能独立回答问题的最小语义单元。
- [Chunk Size 和 Overlap 怎么选？不要背 512 和 100](https://www.wushixiongai.com/questions/project/rag/chunk-size-overlap)：参数不是背出来的。文档结构、问题粒度、Embedding 上限和上下文预算共同决定 Chunk 与 Overlap。
- [跨页表格、代码块和图片怎么切，才不丢结构？](https://www.wushixiongai.com/questions/project/rag/table-code-image-chunking)：特殊内容不是一律“不可切”。关键是先保住结构，再在超限时按可解释的逻辑单元拆分。
- [Chunk 为什么要带标题、页码和前后片段 ID？](https://www.wushixiongai.com/questions/project/rag/chunk-metadata-traceability)：正文负责语义检索，metadata 负责边界、过滤和追溯。两者缺一，RAG 都难以稳定上线。
- [RAG 知识库如何动态更新？重复上传怎么做到不重复、不漏写？](https://www.wushixiongai.com/questions/project/rag/incremental-ingestion-idempotency)：幂等不是“文件名相同就跳过”。要区分内容稳定 ID、旧版本清理和批量写入的部分失败。
- [RAG 为什么要同时用 BM25 和向量检索？](https://www.wushixiongai.com/questions/project/rag/hybrid-retrieval-bm25-vector)：按面试官的追问顺序，拆解关键词检索、语义检索、融合、重排、故障排查和效果验证。
- [RRF 和加权求和怎么选？两路检索分数为何不能直接相加？](https://www.wushixiongai.com/questions/project/rag/rrf-vs-weighted-fusion)：一个是词面相关度，一个是向量相似度。直接相加看似简单，实际先要解决分数是否可比。
- [Rerank 为什么更准，却不能对整个知识库运行？](https://www.wushixiongai.com/questions/project/rag/rerank-cost-quality)：Rerank 负责把候选排对，不负责从全库找回遗漏证据。它更精细，也因此更昂贵。
- [Query 重写、扩写和 HyDE，应该在哪些查询上启用？](https://www.wushixiongai.com/questions/project/rag/query-rewrite-expansion-hyde)：Query 优化不是统一加一道 LLM 调用。不同方法解决不同失败类型，也可能改错用户原意。
- [Embedding 和 Rerank 模型怎么选？什么时候值得领域微调？](https://www.wushixiongai.com/questions/project/rag/embedding-rerank-model-selection)：榜单只能筛候选，真实知识库的 Query、Chunk 和资源约束才决定最终模型。
- [RAG 检索效果差怎么排查？从 Query 到 Rerank 逐层定位](https://www.wushixiongai.com/questions/project/rag/rag-retrieval-debugging)：不要一看到错误答案就改 Prompt。先确认正确证据有没有入库、有没有召回、有没有排到前面。
- [资料里没有答案时，RAG 怎么避免强行生成？](https://www.wushixiongai.com/questions/project/rag/rag-no-answer-control)：向量库总能返回最接近的片段。系统还要判断这些片段是否真的足以回答。
- [RAG 引用怎么做到可核验？为什么让模型自己标来源不够？](https://www.wushixiongai.com/questions/project/rag/rag-citation-attribution)：引用不是答案末尾多一个文件名。它必须能证明具体陈述，并能回到正确页码或片段。
- [多轮 RAG 如何理解“这个怎么申请”？](https://www.wushixiongai.com/questions/project/rag/multi-turn-query-rewrite)：把全部历史塞给检索器并不能自动解决多轮问题。先要把当前追问恢复成可独立检索的问题。
- [对话历史太长怎么办？滑动窗口、摘要和长期记忆怎么分工？](https://www.wushixiongai.com/questions/project/rag/conversation-memory-management)：记忆不是把所有聊天记录永久塞进 Prompt，而是选择当前回答真正需要的历史信息。
- [RAG 评测集怎么构建？问题、答案和证据必须怎么标？](https://www.wushixiongai.com/questions/project/rag/rag-evaluation-dataset)：只有问题和答案不够。没有正确证据位置，就无法判断到底是检索错了还是生成错了。
- [Recall@K、MRR、Faithfulness，分别在评估 RAG 哪一层？](https://www.wushixiongai.com/questions/project/rag/rag-evaluation-metrics)：RAG 是一条链路。指标也必须分层，否则最终分数下降时仍然不知道该修哪一段。
- [RAG 首字响应为什么慢？Embedding、检索和生成怎么逐段优化？](https://www.wushixiongai.com/questions/project/rag/rag-first-token-latency)：性能优化的第一步不是上缓存，而是把首字前发生的每个阶段单独计时，找出最长的那一段。
- [一个可上线的 RAG 系统怎么拆？离线入库、在线问答和监控如何闭环？](https://www.wushixiongai.com/questions/project/rag/production-rag-architecture)：能跑通 Demo 只证明主链路存在。可上线还要处理状态、隔离、重试、观测、降级和回归。
- [什么时候该用 Deep Research，而不是普通 RAG？复杂问题怎么分级路由？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-complexity-routing)：Deep Research 不是默认链路。先判断任务最低需要哪一级能力，再用路由评测集验证质量、等待时间和工具代价。
- [Deep Research 拿到一个大问题后，怎么拆成能查证的子问题？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-question-decomposition-replanning)：问题拆解不是列报告目录，而是为每项研究主张定义动作、证据和完成条件，并根据新证据增量调整计划。
- [Deep Research 为什么会一直搜索？什么时候该停？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-stopping-conditions)：最大步数只能防止程序失控。真正的停止机制要区分研究完成、无进展、风险中止和预算耗尽。
- [Deep Research 报告有引用就可信吗？多个来源互相矛盾怎么办？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-citation-cross-validation)：引用不是报告末尾的链接列表。每个关键结论都要绑定原文、位置和适用条件，来源冲突先对齐口径再表达。
- [ReAct 为什么短任务好用，做长程研究却会越跑越乱？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-react-long-horizon)：ReAct 的循环适合短任务。长程研究的问题来自原始 Observation 持续累积、状态不清和错误沿轨迹传播。
- [IterResearch 的演进报告为什么不是普通摘要，怎样避免越改越错？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-iterresearch-evolving-report)：演进报告是驱动下一轮研究的受控状态，不是只追求简短和通顺的对话摘要。
- [ReSum 动态摘要怎样压缩长任务，又不把关键证据丢掉？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-resum-summary-quality)：ReSum 用可控的信息损失换取长任务继续执行，质量关键在保护状态、保存证据和支持回滚。
- [Deep Research 并行任务怎么拆，依赖、尾延迟和综合成本怎样控制？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-parallel-research-synthesis)：并行研究的关键不是多开 Agent，而是只并行无依赖单元，并控制尾部任务、共享工具与综合成本。
- [Deep Research 有 Search、Visit、Scholar 和 Python，Agent 怎样选对工具？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-tool-routing)：工具路由要根据当前证据缺口选择能力，并由程序检查前置条件、权限、参数和重复动作。
- [Function Calling 有了 JSON Schema，为什么工具参数还是会错？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-tool-schema-function-calling)：JSON Schema 能提高结构稳定性，但业务语义、当前状态、权限和副作用仍需程序校验。
- [Search 超时、Visit 受限、Python 报错时，Deep Research 怎样继续而不乱答？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-tool-failure-recovery)：工具容错的目标不是永远返回文本，而是在失败时仍保持证据标准、状态可恢复和边界清楚。
- [缓存能省调用，为什么 Deep Research 反而可能拿旧资料回答“最新”？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-cache-freshness)：缓存命中只说明旧结果可复用候选，不代表它仍满足当前任务的时间、权限和证据要求。
- [Deep Research 训练的种子问题从哪里来，怎样避免一开始就把数据做偏？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-seed-data-design)：种子问题决定后续轨迹与训练分布，任何来源都要经过证据、合规、去重和难度验收。
- [简单问答怎样升级成真正需要多步研究的问题，又不把题目编坏？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-complex-qa-generation)：复杂问答的难度来自不可省略的证据路径，而不是题目更长；每次变换后都要独立求解并重新验收。
- [Teacher 生成的 Agent 轨迹为什么不能直接拿去训练，怎样筛出高质量样本？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-trajectory-generation-quality)：Teacher 只生成候选轨迹，工具观察、行动进展、证据答案与数据分布都要独立验证。
- [轨迹里工具返回为什么不能参与 Loss，Observation Mask 做错会怎样？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-observation-loss-mask)：工具返回要作为上下文被模型看见，但不能成为模型的生成目标；Mask 必须在真实模板和 token 边界上验证。
- [首次 SFT 后格式稳定了，为什么工具选择和答案仍然不行？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-first-sft-format-vs-policy)：格式稳定只说明冷启动开始生效，工具选择、证据利用和答案完整度仍需独立评测与补数。
- [训练数据里 Search 太多，为什么 Agent 遇到什么都想先搜？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-tool-selection-data-bias)：工具偏置来自状态与动作边界失真，不能靠各工具数量相等解决，也不能预先承诺调比例后的收益。
- [Deep Research Agent 训练中，CPT、SFT 和 GRPO 应该怎样取舍？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-multistage-agent-training)：CPT 注入行动偏置，SFT 建立可执行起点，GRPO 优化环境策略；是否进入每一阶段必须由能力缺口和验收证据决定。
- [GRPO 为什么同一道题要采样一组答案，组内相对优势到底怎么算？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-grpo-group-relative-advantage)：GRPO 用同组 reward 统计替代可学习 value model，但学习质量完全依赖组内可比、reward 与环境可信。
- [长轨迹被截断就算失败吗，为什么这样做会把 Agent 训成提前乱答？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-truncated-trajectory-negative-samples)：没有 final answer 只说明未完成；策略错误、环境失败、预算耗尽与技术截断必须分别处理。
- [Deep Research 一跑几分钟，为什么不能用普通同步接口硬等？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-long-running-task-serving)：长任务需要把客户端连接、任务生命周期和模型工具资源解耦，并用持久化状态机管理恢复。
- [一次 Deep Research 到底花了多少钱，为什么只看总 Token 找不到成本黑洞？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-cost-observability)：成本必须落到任务、步骤、模型、工具、重试与结果质量，在线估算和真实账单也要分开。
- [Deep Research 评估为什么不能只看最终答案，过程合理又该怎么量？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-evaluation-result-process)：最终答案和研究过程是两条独立证据，需要分层评估、共同归因，并防止效率与可读性掩盖事实错误。
- [LLM Judge 分数变高，为什么 Deep Research 可能反而更会刷分？](https://www.wushixiongai.com/questions/project/deep-research/deep-research-llm-judge-goodhart-loop)：Judge 分数只表示符合当前评估器，必须用规则、换位、多 Judge、人工和独立 holdout 校准。
- [LLM、Workflow 和 Agent 怎么选？什么任务根本不该做成 Agent](https://www.wushixiongai.com/questions/project/agent/agent-vs-workflow-decision)：是否使用 Agent，不取决于技术是否热门，而取决于任务是否需要在不确定环境中多步决策，以及系统能否约束它的权限、成本和停止条件。
- [Agent 规划应该写在 Prompt 里，还是固定在业务代码里？](https://www.wushixiongai.com/questions/project/agent/agent-planning-prompt-vs-code)：生产级规划不是把一切交给模型，也不是把所有路径写死，而是用代码固定边界和主干，只开放能够验证、能够回退的局部决策。
- [ReAct Agent 为什么会死循环？停止条件和步数上限怎么设计](https://www.wushixiongai.com/questions/project/agent/react-loop-stop-conditions)：ReAct 的关键不是让模型多思考几轮，而是用业务完成条件、失败条件、预算和重复检测约束循环，并为无法完成的任务准备安全出口。
- [Agent Planner 生成的计划为什么不能直接执行？怎么校验和动态重规划](https://www.wushixiongai.com/questions/project/agent/agent-plan-validation-replanning)：LLM 生成的是候选计划，不是可直接执行的授权。计划必须先验证结构、依赖、工具、权限、预算和完成条件，环境变化后只重规划受影响部分。
- [Agent 反思机制怎么落地？Reflector 和 Critic 不能只让模型“再想一次”](https://www.wushixiongai.com/questions/project/agent/agent-reflection-critic-quality-gate)：有效反思不是让同一个模型重复评价自己，而是用独立证据判断结果缺口，并把修复范围、次数和出口约束成可回放流程。
- [LangGraph 的 StateGraph 怎么设计？State、Node、Edge 到底各负责什么](https://www.wushixiongai.com/questions/project/agent/langgraph-stategraph-design)：StateGraph 不是把多个模型节点连起来，而是把状态、处理责任、条件路由和失败出口显式化。
- [Agent 任务中断后怎么恢复？Checkpoint、断点续跑和人工审批怎样设计](https://www.wushixiongai.com/questions/project/agent/agent-checkpoint-resume-human-approval)：可靠恢复需要一致快照、幂等副作用和审批协议共同工作，而不是只保存聊天记录。
- [Function Calling 到底怎么跑？模型不会执行函数，完整工具结果循环怎么实现](https://www.wushixiongai.com/questions/project/agent/function-calling-tool-result-loop)：模型只提出工具调用意图，宿主程序校验并执行，再把结果回传给模型继续决策。
- [Agent 工具怎么设计和注册？Tool Schema、Registry 与执行器如何解耦](https://www.wushixiongai.com/questions/project/agent/agent-tool-schema-registry)：Schema 帮模型理解工具，Registry 管理执行器和策略，网关负责真实校验、执行与审计。
- [Agent 有几十上百个工具时怎么选？大工具集路由怎样设计和验收](https://www.wushixiongai.com/questions/project/agent/agent-large-tool-routing)：大工具集先按身份和任务过滤，再召回与重排小候选集，最终由网关重新授权和执行。
- [Agent 工具返回成功就算完成了吗？调用、执行、业务成功要分三层判断](https://www.wushixiongai.com/questions/project/agent/agent-tool-three-layer-success)：工具不报错不代表任务完成，协议、真实执行和业务目标必须分别判断并记录证据。
- [Agent 工具超时怎么办？重试、换工具、降级和停止怎样设计](https://www.wushixiongai.com/questions/project/agent/agent-tool-timeout-retry-fallback)：工具超时先识别错误、副作用和剩余预算，再选择有限重试、兼容降级或安全停止。
- [Agent 调用写操作怎么防误执行？副作用分级、幂等、确认和审计怎样做](https://www.wushixiongai.com/questions/project/agent/agent-side-effect-safety-idempotency)：模型只能建议写操作，真实执行必须经过权限、确认、幂等状态机和完整审计。
- [MCP 和 Function Calling 有什么区别？Agent 工具接入时两者怎样配合](https://www.wushixiongai.com/questions/project/agent/mcp-vs-function-calling-agent-tools)：Function Calling 表达模型调用意图，MCP 标准化 Host 到外部工具的发现、连接与结果传输。
- [Agent 短任务和长任务怎么记？工作记忆、长期记忆与任务状态怎样分层](https://www.wushixiongai.com/questions/project/agent/agent-short-long-task-memory)：工作记忆服务当前推理，任务状态保证长任务连续，长期记忆只保存未来仍值得复用的信息。
- [Agent 长期记忆什么时候写入？从候选提取到更新、冲突和遗忘的完整生命周期](https://www.wushixiongai.com/questions/project/agent/agent-memory-write-lifecycle)：长期记忆不是自动保存对话，而是一条带授权、来源、版本、权限和删除能力的数据生命周期。
- [Agent 从海量记忆里怎么找对内容？相关性、时间性和重要性怎样组合](https://www.wushixiongai.com/questions/project/agent/agent-memory-relevance-recency-importance)：先过滤不能用的记忆，再用相关性、时间性、重要性和来源选择真正值得进入上下文的少量内容。
- [Agent 记忆数据怎么存？关系库、向量库、对象存储与 Schema 怎样设计](https://www.wushixiongai.com/questions/project/agent/agent-memory-storage-schema)：关系库保存事实和生命周期，向量库只做语义索引，对象存储放大工件，缓存不承担真相。
- [复杂任务一定要拆成多 Agent 吗？单 Agent 与多 Agent 怎么选](https://www.wushixiongai.com/questions/project/agent/single-agent-vs-multi-agent)：先用单 Agent 建立失败基线，只有拆分能带来可验证的专业化、隔离或并行时才引入多 Agent。
- [多 Agent 怎么编排？固定流水线和 Supervisor 监督模式怎么选](https://www.wushixiongai.com/questions/project/agent/multi-agent-pipeline-vs-supervisor)：固定依赖交给图，运行时未知分支才交给 Supervisor，多数可靠方案是外层 Pipeline 加局部动态监督。
- [Text2SQL 面对上百张表，怎样选对 Schema 并处理业务歧义](https://www.wushixiongai.com/questions/project/agent/text2sql-schema-selection-ambiguity)：先解析业务口径并构造最小连通 Schema，无法唯一确定指标、时间或范围时再做最小澄清。
- [Text2SQL 生成的 SQL 敢直接执行吗？安全校验与执行边界怎么设计](https://www.wushixiongai.com/questions/project/agent/text2sql-sql-safety-validation)：模型只生成不可信查询计划，系统用 AST、权限、资源、隔离执行和结果审查决定能否运行。
- [Agent 能执行 Python 代码，怎样设计 Code Interpreter 沙箱才不会伤到宿主机](https://www.wushixiongai.com/questions/project/agent/agent-code-interpreter-sandbox)：源码检查只减少误用，真正安全边界是短生命周期执行环境、最小权限、网络文件隔离和资源限制。
- [Agent 线上答错一次，怎样用 Trace 和 Replay 定位到底哪一步坏了](https://www.wushixiongai.com/questions/project/agent/agent-trace-replay-observability)：Trace 还原任务事实链，Replay 复现和验证修复，最终从用户失败定位到最早错误步骤。
- [Agent 怎么评估才不只看最终答案？任务、执行轨迹与成本要分层验收](https://www.wushixiongai.com/questions/project/agent/agent-evaluation-task-trajectory-cost)：结果证明事情做成，轨迹证明过程可靠，成本证明方案值得，安全与关键事实不能被综合分抵消。

## 简历项目模拟面试（10 组）
> 以下页面展示教学用项目描述和连续面试问题，不包含个性化回答。
- [企业知识库 RAG 架构项目深挖题](https://www.wushixiongai.com/resume-interview/enterprise-rag-architecture)：适合简历中写过企业知识库或 RAG 项目的人，重点练自己负责的部分、项目架构、检索效果和上线问题。
- [PDF、表格、OCR 与文档解析项目深挖题](https://www.wushixiongai.com/resume-interview/document-parsing-pipeline)：适合写过 PDF、表格或 OCR 解析的人，重点练处理流程、质量评估、错误页面和失败兜底。
- [混合检索、RRF 与重排项目深挖题](https://www.wushixiongai.com/resume-interview/hybrid-retrieval-rrf-rerank)：适合写过混合检索、RRF 或重排的人，重点练为什么这样选、效果提升来自哪里、延迟增加多少。
- [RAG 评测集与 bad case 项目深挖题](https://www.wushixiongai.com/resume-interview/rag-evaluation-bad-cases)：适合做过 RAG 评测的人，重点练评测数据怎么来、指标怎么算、错误怎么定位，以及怎样持续改进。
- [多轮对话、引用与权限隔离项目深挖题](https://www.wushixiongai.com/resume-interview/multi-turn-citation-permission)：适合做过多轮对话、引用或权限控制的人，重点练上下文管理、引用准确性和不同用户的数据隔离。
- [Agent 工具调用与 Schema 设计项目深挖题](https://www.wushixiongai.com/resume-interview/agent-tool-schema)：适合做过 Agent 工具调用的人，重点练接口设计、参数校验、失败重试、重复执行和安全控制。
- [Agent 工作流、状态持久化与人工介入项目深挖题](https://www.wushixiongai.com/resume-interview/agent-workflow-state-human)：适合做过 Agent 工作流的人，重点练状态如何保存、失败后怎样恢复，以及什么时候需要人工确认。
- [Agent 超时、重试、幂等、安全与降级项目深挖题](https://www.wushixiongai.com/resume-interview/agent-reliability-safety)：适合处理过 Agent 超时、重试或降级的人，重点练故障处理、避免重复执行、安全防护和失败兜底。
- [vLLM 推理服务、吞吐、延迟与成本项目深挖题](https://www.wushixiongai.com/resume-interview/vllm-serving-performance-cost)：适合部署过大模型推理服务的人，重点练服务选型、吞吐、延迟、显存、容量和成本。
- [Trace、线上监控与问题复盘项目深挖题](https://www.wushixiongai.com/resume-interview/trace-monitoring-incident-review)：适合建设过大模型监控或 Trace 的人，重点练日志、指标、故障定位，以及如何避免同类问题再次发生。

## 说明
- 全站题目页、教程章节页与简历项目模拟面试页均为静态 HTML（正文全文可抓取引用）；全部题目直链见 [llms-full.txt](https://www.wushixiongai.com/llms-full.txt)
- Sitemap: https://www.wushixiongai.com/sitemap.xml

## 全部面试题(851 道)
- [Flash Attention 原理怎么理解?](https://www.wushixiongai.com/q/arch-flash-attention-io-optimization)
- [DDPM vs Flow Matching: 采样效率与优缺点?](https://www.wushixiongai.com/q/train-ddpm-vs-flow-matching)
- [DDPM 闭式采样为何无需展开前向链?](https://www.wushixiongai.com/q/train-ddpm-random-timestep-sampling)
- [PPO vs DPO vs GRPO 怎么选?](https://www.wushixiongai.com/q/rl-ppo-dpo-alignment-overview)
- [DPO vs PPO 怎么选?](https://www.wushixiongai.com/q/rl-dpo-vs-ppo-alignment)
- [模型优化方法有哪些？](https://www.wushixiongai.com/q/infra-model-optimization-methods)
- [大模型面临哪些核心挑战?](https://www.wushixiongai.com/q/arch-llm-challenges-bottlenecks)
- [Reward 函数设计关键因素?](https://www.wushixiongai.com/q/rl-reward-function-key-factors)
- [长周期 Agent 如何建模与规划?](https://www.wushixiongai.com/q/rl-long-horizon-modeling-training)
- [Transformer 为何用 Layer Norm?](https://www.wushixiongai.com/q/arch-layernorm-vs-batchnorm-rationale)
- [SFT 置信度下降影响](https://www.wushixiongai.com/q/train-sft-confidence-distribution-shift)
- [位置编码为什么重要?](https://www.wushixiongai.com/q/arch-positional-encoding-importance)
- [奖励函数设计与投机陷阱](https://www.wushixiongai.com/q/rl-reward-function-design-principles)
- [长视野任务的信用分配与 Option](https://www.wushixiongai.com/q/rl-long-horizon-sparse-reward-credit)
- [LLaMA为何选RMSNorm?](https://www.wushixiongai.com/q/arch-layernorm-vs-rmsnorm-tradeoffs)
- [LayerNorm vs BatchNorm 怎么选?](https://www.wushixiongai.com/q/arch-layernorm-over-batchnorm-reasons)
- [DDPM 为什么随机采样时间步?](https://www.wushixiongai.com/q/train-ddpm-timestep-sampling-rationale)
- [RMSNorm vs LayerNorm：训练效率怎么比?](https://www.wushixiongai.com/q/arch-layernorm-rmsnorm-modern-models)
- [Transformer 为何要归一化?](https://www.wushixiongai.com/q/arch-normalization-purpose-methods)
- [PyTorch 动态图怎么自动微分?](https://www.wushixiongai.com/q/train-pytorch-dynamic-graph-autograd)
- [Agent 奖励如何对齐目标并防作弊?](https://www.wushixiongai.com/q/rl-agent-reward-function-design)
- [长视界任务怎么结合环境训练?](https://www.wushixiongai.com/q/rl-long-horizon-credit-assignment)
- [vLLM vs TensorRT-LLM 怎么选?](https://www.wushixiongai.com/q/infra-inference-frameworks-comparison)
- [大模型任务表现差怎么排查？](https://www.wushixiongai.com/q/project-model-underperformance-troubleshooting)
- [CNN/RNN/Transformer 怎么选?](https://www.wushixiongai.com/q/arch-neural-network-types-overview)
- [逻辑回归原理怎么理解?](https://www.wushixiongai.com/q/arch-logistic-regression-principle)
- [主流图像生成模型有哪些?](https://www.wushixiongai.com/q/multimodal-image-generation-models)
- [多模态大模型面临哪些挑战?](https://www.wushixiongai.com/q/multimodal-challenges-alignment-consistency)
- [Focal Loss 公式与作用机制](https://www.wushixiongai.com/q/train-focal-loss-class-imbalance)
- [训练数据清洗：规则 vs LLM](https://www.wushixiongai.com/q/train-data-cleaning-rule-vs-llm)
- [LoRA 低秩分解 vs 全量微调怎么选?](https://www.wushixiongai.com/q/train-lora-low-rank-principle)
- [参数量增长如何影响模型能力?](https://www.wushixiongai.com/q/train-parameter-scaling-effects)
- [项目局限性与改进方向](https://www.wushixiongai.com/q/project-limitations-reflection)
- [参数规模怎么影响大模型?](https://www.wushixiongai.com/q/arch-parameter-scale-generalization)
- [二分类交叉熵损失函数怎么推导?](https://www.wushixiongai.com/q/train-binary-cross-entropy-loss)
- [MSE 损失函数怎么用?](https://www.wushixiongai.com/q/train-mse-loss-regression)
- [Encoder vs Decoder 功能区别?](https://www.wushixiongai.com/q/arch-encoder-vs-decoder-roles)
- [LLM 准确性提升四大路径](https://www.wushixiongai.com/q/eval-llm-accuracy-improvement-paths)
- [LLM 推理框架综合选型](https://www.wushixiongai.com/q/infra-inference-frameworks-overview)
- [评测数据集怎么选?](https://www.wushixiongai.com/q/project-evaluation-dataset-selection)
- [Layer Norm 怎么稳定训练?](https://www.wushixiongai.com/q/arch-layernorm-mechanism-sublayers)
- [实习项目怎么介绍最加分?](https://www.wushixiongai.com/q/project-representative-internship-intro)
- [训练曲线异常怎么排查？](https://www.wushixiongai.com/q/train-training-curves-interpretation)
- [MoE 怎么提升 Agent 能力?](https://www.wushixiongai.com/q/arch-moe-for-agent-capabilities)
- [ML项目流程：特征工程到评估](https://www.wushixiongai.com/q/project-ml-full-pipeline-experience)
- [Prompt vs Seq2Seq，怎么转？](https://www.wushixiongai.com/q/train-discriminative-to-generative-methods)
- [深度学习项目怎么介绍?](https://www.wushixiongai.com/q/project-deep-learning-experience)
- [VQ-VAE 原理与完整结构](https://www.wushixiongai.com/q/arch-vqvae-principle-structure)
- [稀疏与延迟奖励的塑形与权衡](https://www.wushixiongai.com/q/rl-reward-mechanism-factors)
- [扩散模型 SNR 怎么用?](https://www.wushixiongai.com/q/train-diffusion-snr-scheduling)
- [RL 训练环境与 Sim2Real](https://www.wushixiongai.com/q/rl-training-environment-setup)
- [RMSNorm vs LayerNorm：数值稳定性怎么比?](https://www.wushixiongai.com/q/arch-layernorm-rmsnorm-comparison)
- [格式化输出 vs 思维链，怎么选？](https://www.wushixiongai.com/q/arch-discriminative-to-generative-approaches)
- [多模态大模型有哪些技术挑战?](https://www.wushixiongai.com/q/multimodal-key-technical-challenges)
- [Transformer 为何用 LayerNorm 而非 BatchNorm?](https://www.wushixiongai.com/q/arch-transformer-layernorm-not-batchnorm)
- [批量文本清洗如何自动质检?](https://www.wushixiongai.com/q/train-batch-data-cleaning-pipeline)
- [Transformer归一化方法有哪些?](https://www.wushixiongai.com/q/arch-normalization-methods-training-stability)
- [13B模型INT8/INT4量化后多大?](https://www.wushixiongai.com/q/infra-quantization-storage-inference-impact)
- [Agent 商业落地评估框架](https://www.wushixiongai.com/q/interview-agent-capability-commercialization)
- [客服办公导购 Agent 选型](https://www.wushixiongai.com/q/interview-agent-application-scenario-outlook)
- [合成数据怎么系统构建?](https://www.wushixiongai.com/q/train-synthetic-data-construction)
- [VQ-VAE 码本塌缩治理](https://www.wushixiongai.com/q/arch-vqvae-codebook-discrete-latent)
- [实习项目被追问哪些技术细节?](https://www.wushixiongai.com/q/interview-project-deep-dive-topics)
- [判别式转生成式有哪些方法？](https://www.wushixiongai.com/q/arch-task-reformulation-generative-methods)
- [推理框架硬件支持对比](https://www.wushixiongai.com/q/infra-inference-frameworks-hardware-support)
- [技术项目怎么讲给非技术人?](https://www.wushixiongai.com/q/interview-project-non-technical-explanation)
- [项目难题怎么定位与解决?](https://www.wushixiongai.com/q/project-challenge-problem-solving)
- [FAISS 怎么实现高效 ANN 搜索?](https://www.wushixiongai.com/q/infra-faiss-similarity-search)
- [SFT vs 预训练:核心区别在哪?](https://www.wushixiongai.com/q/train-sft-vs-pretraining-differences)
- [信息抽取：规则 vs LLM](https://www.wushixiongai.com/q/train-rule-vs-llm-information-extraction)
- [Encoder vs Decoder 核心区别?](https://www.wushixiongai.com/q/arch-encoder-decoder-design-applications)
- [LayerNorm 如何控制激活尺度?](https://www.wushixiongai.com/q/arch-layernorm-role-gradient-stability)
- [LoRA 微调效果差怎么排查?](https://www.wushixiongai.com/q/train-lora-underperformance-diagnosis)
- [Agent 工具调用奖励设计](https://www.wushixiongai.com/q/rl-agent-reward-design-guidelines)
- [超长上下文怎么处理?](https://www.wushixiongai.com/q/infra-long-context-handling)
- [LoRA 有哪些局限性?](https://www.wushixiongai.com/q/train-lora-limitations-alternatives)
- [MSE vs 交叉熵损失怎么选?](https://www.wushixiongai.com/q/train-mse-vs-cross-entropy)
- [Few-shot 推理为何有效?](https://www.wushixiongai.com/q/arch-few-shot-inference-mechanism)
- [MSE vs Focal Loss 分类怎么选?](https://www.wushixiongai.com/q/train-classification-mse-vs-focal)
- [自回归生成用 token id 还是概率分布?](https://www.wushixiongai.com/q/arch-autoregressive-token-feeding)
- [大模型应用怎么保障可靠性?](https://www.wushixiongai.com/q/eval-output-reliability-assurance)
- [自动 vs 人工评价怎么选?](https://www.wushixiongai.com/q/eval-automatic-vs-human-metrics)
- [模型收敛怎么判断?](https://www.wushixiongai.com/q/train-convergence-judgment-metrics)
- [SFT 过拟合怎么判断和缓解?](https://www.wushixiongai.com/q/train-sft-overfitting-detection)
- [SFT 后置信度下降怎么办?](https://www.wushixiongai.com/q/train-sft-prediction-confidence-drop)
- [SFT 后置信度下降原因](https://www.wushixiongai.com/q/train-sft-confidence-drop-causes)
- [LLM 幻觉与安全风险怎么缓解?](https://www.wushixiongai.com/q/eval-hallucination-safety-mitigation)
- [PyTorch 计算图怎么支持自动微分?](https://www.wushixiongai.com/q/infra-pytorch-dynamic-vs-static-graph)
- [大模型微调方法怎么选?](https://www.wushixiongai.com/q/project-finetuning-practical-experience)
- [大模型 vs 规则引擎怎么选?](https://www.wushixiongai.com/q/project-llm-vs-rule-engine)
- [验证集分布：以部署目标为准](https://www.wushixiongai.com/q/eval-validation-set-class-distribution)
- [Coding/GUI/Search 商业化先后](https://www.wushixiongai.com/q/interview-agent-capabilities-coding-gui-search)
- [判别式转生成式哪种方法更优?](https://www.wushixiongai.com/q/eval-task-reformulation-method-comparison)
- [固定间隔vs关键帧怎么选？](https://www.wushixiongai.com/q/multimodal-video-frame-sampling-strategies)
- [FFN 隐藏层维度为何扩 4 倍?](https://www.wushixiongai.com/q/arch-ffn-hidden-dimension-expansion)
- [SFT prompt 模板设计原则与结构](https://www.wushixiongai.com/q/train-sft-prompt-template-design)
- [面试反问怎么问才加分?](https://www.wushixiongai.com/q/interview-questions-to-ask-team)
- [环境训练 vs 监督学习怎么选?](https://www.wushixiongai.com/q/rl-environment-based-vs-supervised)
- [模型收敛必须用验证集吗？](https://www.wushixiongai.com/q/train-convergence-validation-set-necessity)
- [扫描PDF怎么解析内容?](https://www.wushixiongai.com/q/project-scanned-pdf-parsing-pipeline)
- [SFT 硬件资源怎么配?](https://www.wushixiongai.com/q/train-sft-hardware-requirements)
- [SFT 复读机问题怎么解?](https://www.wushixiongai.com/q/train-sft-repetition-problem)
- [LLM 训练哪步最难?](https://www.wushixiongai.com/q/train-most-challenging-stage)
- [成就感经历 STAR 回答](https://www.wushixiongai.com/q/interview-most-rewarding-experience)
- [PPO vs GRPO 怎么选?](https://www.wushixiongai.com/q/rl-ppo-vs-grpo-comparison)
- [GSPO 与 SFT、PPO、GRPO 区别](https://www.wushixiongai.com/q/rl-gspo-importance-policy-learning)
- [Agent 哪项能力最先商业化?](https://www.wushixiongai.com/q/interview-agent-core-capability-priority)
- [环境训练方法怎么选?](https://www.wushixiongai.com/q/rl-environment-based-training-methods)
- [主流生成式模型有哪些?](https://www.wushixiongai.com/q/multimodal-generative-models-overview)
- [模拟器与 World Model 的偏差和迁移](https://www.wushixiongai.com/q/rl-environment-simulation-long-sequence)
- [Agent 未来规模化方向](https://www.wushixiongai.com/q/interview-agent-direction-near-term)
- [均匀采样vs关键帧如何影响模型？](https://www.wushixiongai.com/q/multimodal-video-frame-sampling-selection)
- [数据预处理关键步骤与目的](https://www.wushixiongai.com/q/project-data-preprocessing-experience)
- [实习项目业务场景怎么讲?](https://www.wushixiongai.com/q/project-internship-business-scenario)
- [RLHF 中 Reference Model 不训练?](https://www.wushixiongai.com/q/rl-rlhf-reference-model-role)
- [图像数据超参数怎么调?](https://www.wushixiongai.com/q/train-image-data-hyperparameter-tuning)
- [AI 论文怎么读?核心思想+创新点](https://www.wushixiongai.com/q/interview-recent-paper-reading)
- [大规模数据清洗自动化关键步骤](https://www.wushixiongai.com/q/train-business-data-cleaning-automation)
- [MoE 怎么增强 Agent 能力?](https://www.wushixiongai.com/q/arch-moe-agent-task-routing)
- [Flow Matching 线性目标速度](https://www.wushixiongai.com/q/train-flow-matching-gaussian-velocity)
- [高噪 vs 低噪专家怎么划分?](https://www.wushixiongai.com/q/multimodal-diffusion-noise-expert-routing)
- [可视化怎么诊断训练问题?](https://www.wushixiongai.com/q/train-visualization-diagnostics)
- [多模态异构数据怎么统一存储?](https://www.wushixiongai.com/q/multimodal-heterogeneous-storage-representation)
- [幻觉一致性延迟诊断](https://www.wushixiongai.com/q/infra-model-defects-solutions)
- [大模型项目缺陷修复](https://www.wushixiongai.com/q/infra-model-defects-mitigation)
- [Environment-based 训练怎么用?](https://www.wushixiongai.com/q/rl-environment-based-concept-comparison)
- [数据分层为何不用大模型?](https://www.wushixiongai.com/q/train-data-stratification-model-choice)
- [VQ-VAE 与传统 VAE 有何不同?](https://www.wushixiongai.com/q/arch-vqvae-architecture-workflow)
- [LlamaFactory 怎么构造训练数据?](https://www.wushixiongai.com/q/train-finetuning-frameworks-llamafactory)
- [无标注数据怎么构建高质量数据集?](https://www.wushixiongai.com/q/train-annotated-dataset-construction)
- [视频切帧怎么平衡效率与信息?](https://www.wushixiongai.com/q/multimodal-video-frame-efficiency-tradeoff)
- [工程化核心目标是什么?](https://www.wushixiongai.com/q/infra-engineering-objectives)
- [训练数据来源怎么选?](https://www.wushixiongai.com/q/train-data-sources-quality-compliance)
- [VecEnv 怎么加速强化学习训练?](https://www.wushixiongai.com/q/rl-vectorized-environment-acceleration)
- [GSPO 序列级比率与裁剪](https://www.wushixiongai.com/q/rl-gspo-supervised-policy-optimization)
- [数据难度分层怎么验证?](https://www.wushixiongai.com/q/train-data-difficulty-stratification-validation)
- [普通经历如何讲成就感](https://www.wushixiongai.com/q/interview-proudest-achievement-significance)
- [视频切帧方法怎么选?](https://www.wushixiongai.com/q/multimodal-video-frame-extraction-methods)
- [视频切帧优化：语义+动态采样](https://www.wushixiongai.com/q/multimodal-video-frame-sampling-optimization)
- [LLM 应用场景价值怎么分析?](https://www.wushixiongai.com/q/interview-llm-application-value)
- [多模态 vs 纯文本大模型怎么选?](https://www.wushixiongai.com/q/interview-multimodal-vs-text-preference)
- [MoE 专家网络怎么动态路由?](https://www.wushixiongai.com/q/multimodal-diffusion-expert-selection)
- [LayerNorm 前向实现：数值稳定性陷阱](https://www.wushixiongai.com/q/arch-layernorm-numpy-implementation)
- [LoRA 低秩分解的数学原理是什么?](https://www.wushixiongai.com/q/train-lora-math-structure)
- [Self-Attention 怎么计算 QKV?](https://www.wushixiongai.com/q/arch-self-attention-qkv-mechanism)
- [Transformer 架构核心组件原理](https://www.wushixiongai.com/q/arch-transformer-overall-architecture)
- [Multi-Head Attention 维度变化](https://www.wushixiongai.com/q/arch-multi-head-attention-coding)
- [显存优化技术原理与对比](https://www.wushixiongai.com/q/infra-gpu-memory-optimization)
- [GPT vs BERT 核心区别在哪?](https://www.wushixiongai.com/q/arch-gpt-vs-bert-comparison)
- [技术项目怎么讲出亮点?](https://www.wushixiongai.com/q/project-technical-project-introduction)
- [Reward Model 训练流程详解](https://www.wushixiongai.com/q/rl-reward-model-training-pipeline)
- [SFT 人工标注数据怎么构建?](https://www.wushixiongai.com/q/train-sft-annotation-dataset-design)
- [TRL Agent 组件交互机制](https://www.wushixiongai.com/q/rl-trl-search-agent-components)
- [KV-Cache 空间复杂度怎么算?](https://www.wushixiongai.com/q/arch-kv-cache-space-complexity)
- [DeepSpeed 核心特性与架构怎么用?](https://www.wushixiongai.com/q/infra-deepspeed-framework)
- [ViT 图像块嵌入原理与实现](https://www.wushixiongai.com/q/multimodal-vit-patch-embedding)
- [On-policy、Off-policy 与 DPO 怎么选?](https://www.wushixiongai.com/q/rl-on-policy-vs-off-policy)
- [Decoder-only 架构为何是主流?](https://www.wushixiongai.com/q/arch-decoder-only-rationale)
- [重参数化技巧怎么解决梯度问题?](https://www.wushixiongai.com/q/train-reparameterization-trick)
- [Flow Matching 原理是什么?](https://www.wushixiongai.com/q/train-flow-matching-principle)
- [MLA 与 KV Cache 怎么用?](https://www.wushixiongai.com/q/arch-multi-head-latent-attention)
- [验证集和测试集为何要分开?](https://www.wushixiongai.com/q/eval-validation-test-split-rationale)
- [BERT+CRF 怎么联合训练?](https://www.wushixiongai.com/q/arch-bert-crf-ner)
- [训练数据怎么采集和预处理?](https://www.wushixiongai.com/q/train-pretraining-data-pipeline)
- [PPO vs GRPO vs DPO 适用场景?](https://www.wushixiongai.com/q/rl-ppo-grpo-dpo-principles)
- [PPO vs GRPO vs DPO 原理区别?](https://www.wushixiongai.com/q/rl-ppo-grpo-dpo-comparison)
- [LayerNorm vs RMSNorm 原理差异?](https://www.wushixiongai.com/q/arch-normalization-layernorm-rmsnorm-principle)
- [ZeRO 与 FlashAttention 各优化什么?](https://www.wushixiongai.com/q/arch-attention-optimization-methods)
- [多模态大模型怎么理解？](https://www.wushixiongai.com/q/multimodal-architecture-representative-models)
- [全参数 vs 参数高效微调怎么选?](https://www.wushixiongai.com/q/train-full-vs-parameter-efficient-finetuning)
- [KL散度本质与非对称性怎么用?](https://www.wushixiongai.com/q/arch-kl-divergence-applications)
- [Agent 工具调用训练数据怎么构建?](https://www.wushixiongai.com/q/rl-tool-agent-training-data-construction)
- [面试高频论文怎么选?](https://www.wushixiongai.com/q/interview-research-papers-discussed)
- [SFT、偏好优化与 RL 怎么衔接?](https://www.wushixiongai.com/q/rl-sft-rl-alternating-training)
- [多头注意力原理与QKV拆分陷阱](https://www.wushixiongai.com/q/arch-multi-head-attention-implementation)
- [面试反问环节怎么准备?](https://www.wushixiongai.com/q/interview-candidate-questions-session)
- [NDCG 怎么计算?](https://www.wushixiongai.com/q/eval-ndcg-calculation-explained)
- [多模态架构与训练怎么改进?](https://www.wushixiongai.com/q/multimodal-architecture-training-challenges)
- [多模态大模型网络结构怎么搭?](https://www.wushixiongai.com/q/multimodal-vlm-network-structure)
- [位置编码怎么选? RoPE vs ALiBi](https://www.wushixiongai.com/q/arch-positional-encoding-methods)
- [GLoRA 比传统 LoRA 强在哪?](https://www.wushixiongai.com/q/train-glora-improvements-over-lora)
- [VQ-VAE 损失与码本更新](https://www.wushixiongai.com/q/arch-vqvae-discrete-representation-learning)
- [CPT 后灾难性遗忘怎么防?](https://www.wushixiongai.com/q/train-continual-pretraining-catastrophic-forgetting)
- [VLM 基本原理怎么理解?](https://www.wushixiongai.com/q/multimodal-vlm-principle-alignment)
- [多模态融合与对齐怎么改进?](https://www.wushixiongai.com/q/multimodal-fusion-alignment-improvements)
- [DDPM 为什么不先加噪再端到端去噪?](https://www.wushixiongai.com/q/train-ddpm-stepwise-denoising-rationale)
- [自我介绍怎么突出大模型?](https://www.wushixiongai.com/q/interview-self-introduction)
- [DDPM vs Flow Matching: 建模与训练差异?](https://www.wushixiongai.com/q/train-ddpm-flow-matching-formulation)
- [Long Horizon 训练环境怎么搭?](https://www.wushixiongai.com/q/rl-long-horizon-environment-design)
- [CPT数据清洗与时间序列一致性](https://www.wushixiongai.com/q/train-continual-pretraining-data-pipeline)
- [视频生成加速有哪些方法?](https://www.wushixiongai.com/q/infra-video-generation-acceleration)
- [大模型 vs 搜推小样本特征怎么融合?](https://www.wushixiongai.com/q/project-llm-search-recsys-adaptation)
- [LoRA 微调比全量微调更慢吗？](https://www.wushixiongai.com/q/train-lora-vs-full-compute-cost)
- [Stable Diffusion 各版本怎么选?](https://www.wushixiongai.com/q/multimodal-stable-diffusion-versions)
- [RMSNorm推理效率更高吗?](https://www.wushixiongai.com/q/arch-layernorm-rmsnorm-performance-inference)
- [Hive SQL 熟练度评估指标](https://www.wushixiongai.com/q/interview-hive-sql-bigdata-experience)
- [运筹学最优化 vs 机器学习最优化](https://www.wushixiongai.com/q/interview-optimization-operations-research)
- [Agent 中断恢复:状态持久化策略](https://www.wushixiongai.com/q/agent-interruption-recovery-persistence)
- [Agent 资源调度：多线程冲突避免](https://www.wushixiongai.com/q/agent-resource-scheduling-concurrency)
- [多智能体系统核心技术难点](https://www.wushixiongai.com/q/agent-multi-agent-project-challenges)
- [多轮 Agent RL 难在哪?](https://www.wushixiongai.com/q/agent-rl-training-strategy-challenges)
- [Agent 怎么处理代码库数据?](https://www.wushixiongai.com/q/agent-repository-level-data-handling)
- [Agent 强化学习怎么训练?](https://www.wushixiongai.com/q/agent-rl-training-approach)
- [Agent 实现 AGI 面临哪些挑战?](https://www.wushixiongai.com/q/agent-agi-challenges)
- [ReAct vs Plan-and-Execute 怎么选?](https://www.wushixiongai.com/q/agent-react-vs-plan-execute-context)
- [Agent 长短期记忆如何分层?](https://www.wushixiongai.com/q/agent-memory-system-architecture)
- [ReAct与Plan-and-Execute 上下文管理差异?](https://www.wushixiongai.com/q/agent-react-plan-execute-tradeoffs)
- [ReAct vs Plan-and-Execute 推理效率对比?](https://www.wushixiongai.com/q/agent-react-plan-execute-long-context-comparison)
- [Agent 大模型优化三方向](https://www.wushixiongai.com/q/agent-base-model-design-for-agents)
- [大模型记忆机制怎么改进?](https://www.wushixiongai.com/q/agent-memory-long-context-improvement)
- [Function Calling vs Toolformer 本质差异?](https://www.wushixiongai.com/q/agent-function-calling-vs-toolformer)
- [Attention 在 Agent 多轮对话有哪些局限?](https://www.wushixiongai.com/q/arch-standard-attention-agent-limitations)
- [架构 vs 模块 vs 工具集成怎么选?](https://www.wushixiongai.com/q/agent-framework-comparison-architecture)
- [Tool Calling 参数格式怎么保?](https://www.wushixiongai.com/q/agent-tool-call-parameter-validation)
- [主流 LLM 应用产品怎么选?](https://www.wushixiongai.com/q/project-llm-product-analysis)
- [海量历史记录怎么索引优化？](https://www.wushixiongai.com/q/agent-large-scale-memory-storage)
- [Agent 长期记忆存储结构](https://www.wushixiongai.com/q/agent-memory-retrieval-efficiency)
- [Agent 架构的取舍与适用条件](https://www.wushixiongai.com/q/agent-advantages-limitations-tradeoffs)
- [Agent 技术演进与挑战?](https://www.wushixiongai.com/q/agent-development-trends-challenges)
- [多轮对话上下文状态管理机制](https://www.wushixiongai.com/q/rag-conversation-state-management)
- [多 Agent 冲突怎么检测与协调?](https://www.wushixiongai.com/q/agent-multi-agent-conflict-detection-coordination)
- [Agent模块协作机制与失败场景](https://www.wushixiongai.com/q/agent-autonomous-architecture-modules)
- [Agent 组件如何协同完成复杂任务?](https://www.wushixiongai.com/q/agent-working-principle-with-example)
- [Agent 中 RAG 瓶颈怎么破?](https://www.wushixiongai.com/q/rag-agent-latency-accuracy-bottlenecks)
- [Agent 外部工具怎么集成?](https://www.wushixiongai.com/q/agent-external-tool-design-integration)
- [生成式推荐系统有哪些改进方向？](https://www.wushixiongai.com/q/recsys-generative-future-improvements)
- [冷启动新用户怎么推荐?](https://www.wushixiongai.com/q/recsys-generative-cold-start)
- [Agent记忆分层：摘要与检索协同](https://www.wushixiongai.com/q/agent-ultra-long-memory-scalability)
- [旅行Agent怎么拆解任务?](https://www.wushixiongai.com/q/agent-travel-planning-task-decomposition)
- [LLM 输出格式不规范怎么修?](https://www.wushixiongai.com/q/prompt-intent-output-format-correction)
- [Agent 规划记忆工具调用机制](https://www.wushixiongai.com/q/agent-from-scratch-components-workflow)
- [Agent记忆分层与RAG检索](https://www.wushixiongai.com/q/agent-long-history-context-window)
- [Agent 组件如何实现自主决策与学习?](https://www.wushixiongai.com/q/agent-working-mechanism-components)
- [模糊意图怎么补全?](https://www.wushixiongai.com/q/agent-incomplete-intent-clarification)
- [Agent 工具调用失败怎么处理?](https://www.wushixiongai.com/q/agent-tool-failure-feedback-strategy)
- [多 Agent 协同如何分工?](https://www.wushixiongai.com/q/agent-multi-agent-coordination-design)
- [Function Calling 消息流](https://www.wushixiongai.com/q/agent-function-calling-implementation)
- [Agent 核心模块怎么协同?](https://www.wushixiongai.com/q/agent-components-coordination-mechanism)
- [LangChain vs LlamaIndex 在 RAG 中怎么选?](https://www.wushixiongai.com/q/agent-frameworks-rag-agent-comparison)
- [Agent API容错：重试降级与澄清](https://www.wushixiongai.com/q/agent-tool-call-fault-tolerance)
- [全局规划+局部避障怎么融合?](https://www.wushixiongai.com/q/robotics-global-local-path-fusion)
- [最大重试次数怎么设?](https://www.wushixiongai.com/q/agent-retry-max-attempts-criteria)
- [多智能体系统怎么协同?](https://www.wushixiongai.com/q/agent-multi-agent-concept-collaboration)
- [ReAct 框架核心优势在哪?](https://www.wushixiongai.com/q/agent-react-design-vs-prompting)
- [多智能体系统 vs 单一 Agent 怎么选?](https://www.wushixiongai.com/q/agent-multi-agent-advantages-challenges)
- [Agent 推理链路延迟优化策略](https://www.wushixiongai.com/q/agent-reasoning-chain-latency-optimization)
- [Agent 架构怎么从零设计?](https://www.wushixiongai.com/q/agent-autonomous-design-key-choices)
- [RAG+Agent 延迟与准确率成因](https://www.wushixiongai.com/q/rag-agent-integration-latency-accuracy)
- [Agent长期记忆Embedding检索策略](https://www.wushixiongai.com/q/agent-memory-embedding-retrieval-design)
- [Agent 调用链路容错机制](https://www.wushixiongai.com/q/agent-call-chain-fault-tolerance)
- [Agent 系统哪个模块最不稳定?](https://www.wushixiongai.com/q/agent-module-reliability-stability)
- [Single-Agent vs Multi-Agent 架构取舍](https://www.wushixiongai.com/q/agent-single-vs-multi-design-patterns)
- [工具调用方法怎么选?](https://www.wushixiongai.com/q/agent-tool-calling-methods-overview)
- [Tool Use 实现方法有哪些?](https://www.wushixiongai.com/q/agent-tool-use-methods-comparison)
- [多 Agent 调度策略总览](https://www.wushixiongai.com/q/agent-multi-agent-scheduling-strategies)
- [多 Agent 调度模式选型](https://www.wushixiongai.com/q/agent-multi-agent-scheduling-scenarios)
- [多 Agent 协作一致性](https://www.wushixiongai.com/q/agent-async-execution-consistency)
- [异步 Agent 最终一致性](https://www.wushixiongai.com/q/agent-async-stability-result-consistency)
- [动态障碍物怎么检测与规避？](https://www.wushixiongai.com/q/robotics-dynamic-obstacle-avoidance)
- [多 Agent 协作复杂度](https://www.wushixiongai.com/q/agent-multi-agent-benefits-and-issues)
- [LangChain vs LlamaIndex vs AutoGPT 架构差异?](https://www.wushixiongai.com/q/agent-frameworks-execution-comparison)
- [多轮对话上下文遗忘怎么解?](https://www.wushixiongai.com/q/llm-multi-turn-forgetting-coreference)
- [大模型输出怎么做真实性验证?](https://www.wushixiongai.com/q/eval-output-factuality-verification)
- [高并发下状态一致性怎么保证？](https://www.wushixiongai.com/q/agent-context-state-high-concurrency)
- [架构设计如何保障低延迟访问？](https://www.wushixiongai.com/q/agent-context-state-slot-management)
- [Agent框架模块化与工具集成怎么比?](https://www.wushixiongai.com/q/agent-frameworks-modularity-comparison)
- [长历史行为怎么处理?](https://www.wushixiongai.com/q/agent-long-behavior-sequence)
- [Agent 工具选择：打分排序机制](https://www.wushixiongai.com/q/agent-tool-selection-decision-module)
- [旅行AI Agent 多智能体怎么搭?](https://www.wushixiongai.com/q/agent-travel-itinerary-subagents)
- [多智能体系统如何分解旅行任务](https://www.wushixiongai.com/q/agent-travel-multi-agent-decomposition)
- [多Agent协同机制与共享记忆原理](https://www.wushixiongai.com/q/agent-multi-agent-mechanism-example)
- [Tool Calling 决策机制与执行流程](https://www.wushixiongai.com/q/agent-tool-calling-architecture-design)
- [Agent 重试次数怎么定?](https://www.wushixiongai.com/q/agent-retry-attempts-determination)
- [传统 RAG 有哪些痛点?](https://www.wushixiongai.com/q/rag-traditional-pain-points)
- [多用户Memory隔离：LangChain改造方案](https://www.wushixiongai.com/q/agent-langchain-memory-isolation)
- [Agent 反思机制原理与实现](https://www.wushixiongai.com/q/agent-reflection-mechanism-purpose)
- [反思机制原理与失败场景](https://www.wushixiongai.com/q/agent-reflection-design-principle)
- [长序列行为数据怎么处理?](https://www.wushixiongai.com/q/recsys-long-interaction-history-handling)
- [长序列怎么压缩与摘要?](https://www.wushixiongai.com/q/recsys-long-sequence-compression)
- [LangGraph 状态机与场景](https://www.wushixiongai.com/q/agent-langgraph-directed-graph-orchestration)
- [LangGraph 优势场景有哪些?](https://www.wushixiongai.com/q/agent-langgraph-vs-sequential)
- [Agent 工具执行用 Workflow 吗?](https://www.wushixiongai.com/q/agent-tools-as-workflow-tradeoffs)
- [Workflow 编排复杂任务优劣？](https://www.wushixiongai.com/q/agent-tool-workflow-orchestration)
- [货物装载决策怎么定?](https://www.wushixiongai.com/q/robotics-cargo-loading-decision)
- [取餐流程中先打卡还是先取餐?](https://www.wushixiongai.com/q/robotics-task-ordering-efficiency)
- [Agent 核心定义与关键能力](https://www.wushixiongai.com/q/agent-definition-key-capabilities)
- [Function Calling 怎么解析用户指令?](https://www.wushixiongai.com/q/agent-function-calling-intent-mapping)
- [多Agent协同怎么提升推理正确率?](https://www.wushixiongai.com/q/agent-multi-agent-reasoning-accuracy)
- [LangGraph 核心特性与设计理念](https://www.wushixiongai.com/q/agent-langgraph-features-comparison)
- [MCP vs Function Calling 怎么选?](https://www.wushixiongai.com/q/agent-mcp-vs-function-calling)
- [Agent 核心机制怎么工作?](https://www.wushixiongai.com/q/agent-working-principle-core-mechanism)
- [多 Agent 任务拆解与分工](https://www.wushixiongai.com/q/agent-multi-agent-task-decomposition-benefits)
- [Agent 规划记忆工具分工](https://www.wushixiongai.com/q/agent-core-components-and-functions)
- [Agent 核心组件与工作流程?](https://www.wushixiongai.com/q/agent-concept-components-workflow)
- [Tool Calling 机制详解](https://www.wushixiongai.com/q/agent-tool-calling-mechanism)
- [Agent 核心组件架构](https://www.wushixiongai.com/q/agent-component-architecture-functions)
- [多 Agent 协同为什么必要?](https://www.wushixiongai.com/q/agent-why-decompose-into-multiple)
- [ReAct 与 CoT 怎么选？](https://www.wushixiongai.com/q/agent-understanding-behavior-patterns)
- [Agent 架构落地看哪些组件?](https://www.wushixiongai.com/q/agent-architecture-application-understanding)
- [Reflection vs Memory 怎么分工?](https://www.wushixiongai.com/q/agent-multi-agent-reflection-memory-collaboration)
- [多智能体协同关键策略](https://www.wushixiongai.com/q/agent-multi-agent-collaboration-strategies)
- [为何拆解任务给多Agent协作?](https://www.wushixiongai.com/q/agent-decompose-vs-single-rationale)
- [LangGraph 核心功能与设计理念](https://www.wushixiongai.com/q/agent-langgraph-features-applications)
- [Agent 工具类型怎么选?](https://www.wushixiongai.com/q/agent-tool-types-and-uses)
- [多智能体协同怎么保证?](https://www.wushixiongai.com/q/agent-multi-agent-communication-consistency)
- [多智能体 Memory 机制怎么用?](https://www.wushixiongai.com/q/agent-multi-agent-memory-role)
- [Agent 推理模式对比](https://www.wushixiongai.com/q/agent-frameworks-methods-overview)
- [Agent 核心组件有哪些?](https://www.wushixiongai.com/q/agent-core-components-detailed)
- [DeepResearch Agent 怎么工作?](https://www.wushixiongai.com/q/agent-deep-research-architecture)
- [Agent 系统可调用哪些工具?](https://www.wushixiongai.com/q/agent-tool-api-types)
- [Tool Calling 工具集设计原则](https://www.wushixiongai.com/q/agent-toolset-design-principles)
- [多Agent系统框架怎么搭?](https://www.wushixiongai.com/q/agent-multi-agent-framework-methodology)
- [Agent项目经验怎么突出亮点?](https://www.wushixiongai.com/q/project-agent-business-experience)
- [多Agent系统设计原则有哪些?](https://www.wushixiongai.com/q/agent-multi-agent-design-principles)
- [工具集设计关键因素?](https://www.wushixiongai.com/q/agent-toolset-design-factors)
- [Agent 主流架构怎么选?](https://www.wushixiongai.com/q/agent-mainstream-architecture-comparison)
- [Agent 开发框架对比](https://www.wushixiongai.com/q/agent-implementation-frameworks-overview)
- [Agent 框架 vs LLM 怎么选?](https://www.wushixiongai.com/q/agent-why-framework-needed)
- [LangChain vs AutoGPT vs BabyAGI 怎么选?](https://www.wushixiongai.com/q/agent-frameworks-langchain-autogpt-babyagi)
- [Tool Calling 失败处理](https://www.wushixiongai.com/q/agent-tool-calling-principle-value)
- [Agent 开发技术栈怎么选?](https://www.wushixiongai.com/q/project-agent-tech-stack-selection)
- [Agent 训练数据怎么获取?](https://www.wushixiongai.com/q/agent-training-data-acquisition)
- [Agent 系统实现流程与核心组件](https://www.wushixiongai.com/q/agent-implementation-flow-components)
- [Agent 项目经历完整复盘](https://www.wushixiongai.com/q/project-ai-agent-case-study)
- [工具集设计最佳实践?](https://www.wushixiongai.com/q/agent-toolset-best-practices)
- [LangGraph 构建 Agent 有几种方式?](https://www.wushixiongai.com/q/agent-langgraph-build-approaches)
- [Agent 框架核心组件怎么搭?](https://www.wushixiongai.com/q/agent-framework-design-core-ideas)
- [AI Agent 趋势与挑战怎么破?](https://www.wushixiongai.com/q/agent-trends-technical-challenges)
- [通信 vs 角色分工怎么选？](https://www.wushixiongai.com/q/agent-multi-agent-consensus-mechanism)
- [多 Agent 协同选通信还是共识?](https://www.wushixiongai.com/q/agent-multi-agent-message-passing-coordination)
- [大量工具怎么选?](https://www.wushixiongai.com/q/agent-large-toolset-selection-challenges)
- [Agent 长期记忆存哪里?](https://www.wushixiongai.com/q/agent-long-term-memory-implementations)
- [Agent 长期记忆架构与更新策略](https://www.wushixiongai.com/q/agent-long-term-memory-storage-scheme)
- [Agent 工具集成与工作流编排](https://www.wushixiongai.com/q/agent-tool-integration-workflow)
- [AutoGen vs LangGraph 怎么选?](https://www.wushixiongai.com/q/agent-autogen-vs-langgraph)
- [LangGraph外还有哪些流程编排框架?](https://www.wushixiongai.com/q/agent-orchestration-frameworks-alternatives)
- [Agent 记忆衰退：为何与如何实现](https://www.wushixiongai.com/q/agent-memory-decay-mechanism)
- [Agent 必备基础能力有哪些?](https://www.wushixiongai.com/q/agent-foundational-abilities-complex-tasks)
- [闲聊误判为工具调用怎么防?](https://www.wushixiongai.com/q/agent-chitchat-vs-function-intent)
- [哪类 Agent 最易大规模落地?](https://www.wushixiongai.com/q/agent-near-term-scalable-category)
- [工具 Agent 训练环境设计](https://www.wushixiongai.com/q/agent-rl-environment-design)
- [Agent工具系统：注册与权限控制](https://www.wushixiongai.com/q/agent-scalable-secure-tool-system)
- [Agent vs Workflow 怎么选?](https://www.wushixiongai.com/q/agent-vs-traditional-workflow)
- [Agent工具接口与错误处理](https://www.wushixiongai.com/q/agent-external-tool-implementation)
- [Agent 系统性能优化陷阱](https://www.wushixiongai.com/q/agent-slowness-error-accumulation-optimization)
- [Agent组件通信机制与架构模式对比](https://www.wushixiongai.com/q/agent-internal-component-communication)
- [Agent工具定义、注册与安全控制有哪些？](https://www.wushixiongai.com/q/agent-tool-definition-registration-security)
- [长文档 Chunking 怎么选?](https://www.wushixiongai.com/q/rag-domain-document-chunking-strategies)
- [文本块怎么转向量并建索引?](https://www.wushixiongai.com/q/rag-chunk-embedding-vector-index-build)
- [Chunking 策略怎么选?](https://www.wushixiongai.com/q/rag-chunking-strategies-performance-impact)
- [RAG 向量索引构建流程详解](https://www.wushixiongai.com/q/rag-document-to-vector-index-pipeline)
- [RAG 数据处理与索引流程?](https://www.wushixiongai.com/q/rag-document-to-knowledge-base-pipeline)
- [Chunk 向量化与索引怎么选?](https://www.wushixiongai.com/q/rag-embedding-vector-database-selection)
- [Embedding+向量库怎么做语义匹配?](https://www.wushixiongai.com/q/rag-semantic-similarity-matching-embedding)
- [RAG 文本分块与向量索引流程](https://www.wushixiongai.com/q/rag-chunking-embedding-indexing-workflow)
- [RAG ANN 索引优化](https://www.wushixiongai.com/q/rag-vector-index-efficiency-recall-optimization)
- [Chunking 策略怎么平衡上下文与精度?](https://www.wushixiongai.com/q/rag-chunking-context-precision-tradeoff)
- [RAG 长文档怎么切块?](https://www.wushixiongai.com/q/rag-long-document-chunking-comparison)
- [RAG 数据清洗与 Chunking 怎么选?](https://www.wushixiongai.com/q/rag-data-cleaning-and-chunking-strategies)
- [Embedding 模型选型 vs 落地流程?](https://www.wushixiongai.com/q/rag-embedding-model-selection-vectorization)
- [RAG 知识库搭建流程?](https://www.wushixiongai.com/q/rag-knowledge-base-build-flow)
- [RAG知识库构建流程详解](https://www.wushixiongai.com/q/rag-knowledge-base-build-with-db-selection)
- [医学论文与法律条文切块](https://www.wushixiongai.com/q/rag-professional-document-chunking-scenarios)
- [医疗法律文档切块](https://www.wushixiongai.com/q/rag-medical-legal-document-chunking)
- [专业领域文档怎么清洗?](https://www.wushixiongai.com/q/rag-domain-document-cleaning-preprocessing)
- [RAG 文档 Chunk 划分策略怎么选?](https://www.wushixiongai.com/q/rag-chunking-methods-pros-cons)
- [RAG 分块策略选型](https://www.wushixiongai.com/q/rag-chunking-completeness-accuracy-balance)
- [RAG分块策略：粒度、重叠与写入](https://www.wushixiongai.com/q/rag-chunking-granularity-overlap-design)
- [RAG 分块参数调优](https://www.wushixiongai.com/q/rag-chunking-strategy-selection-guide)
- [分块策略怎么影响检索?](https://www.wushixiongai.com/q/rag-text-chunking-strategies-considerations)
- [常用分块方法有哪些?](https://www.wushixiongai.com/q/rag-chunking-methods-scenarios-comparison)
- [RAG 分块长度与重叠如何取舍?](https://www.wushixiongai.com/q/rag-chunk-size-overlap-semantic-optimization)
- [RAG向量化流程：预处理到归一化](https://www.wushixiongai.com/q/rag-document-query-vectorization-flow)
- [RAG向量化：分块与Embedding](https://www.wushixiongai.com/q/rag-vectorization-cosine-similarity-matching)
- [RAG 文档怎么存储和向量化?](https://www.wushixiongai.com/q/rag-document-storage-vectorization-pipeline)
- [RAG文本向量化：分块与嵌入陷阱](https://www.wushixiongai.com/q/rag-text-to-vector-encoding-pipeline)
- [RAG多轮对话怎么防知识遗忘?](https://www.wushixiongai.com/q/rag-multi-turn-context-drift)
- [高并发 Agent 延迟卡在哪?](https://www.wushixiongai.com/q/agent-high-concurrency-latency-optimization)
- [多Agent冲突怎么检测与缓解?](https://www.wushixiongai.com/q/agent-multi-agent-conflict-resolution)
- [子任务价值怎么归因?](https://www.wushixiongai.com/q/agent-subtask-contribution-evaluation)
- [Agent 训练链路怎么搭建?](https://www.wushixiongai.com/q/agent-training-pipeline-reward-design)
- [Agent 训练 vs RAG 怎么选?](https://www.wushixiongai.com/q/agent-training-environment-action-space)
- [Agent部署关键环节与陷阱](https://www.wushixiongai.com/q/agent-deployment-serving-monitoring)
- [RAG 工作原理与局限怎么分析?](https://www.wushixiongai.com/q/rag-understanding-principle-tradeoffs)
- [多路召回评估指标与RAG场景](https://www.wushixiongai.com/q/retrieval-multi-recall-quality-evaluation)
- [搜索排序A/B测试权重设计陷阱](https://www.wushixiongai.com/q/search-ranking-ab-test-feature-weights)
- [Agent 多工具如何评分选路?](https://www.wushixiongai.com/q/agent-tool-selection-scoring-mechanism)
- [Agent 如何防越权与隐私泄露?](https://www.wushixiongai.com/q/agent-security-privacy-access-control)
- [RAG 指标权衡与阈值](https://www.wushixiongai.com/q/rag-evaluation-key-metrics-challenges)
- [Agent 系统优缺点怎么分析?](https://www.wushixiongai.com/q/agent-advantages-challenges-limitations)
- [RAG 缓解幻觉的机制](https://www.wushixiongai.com/q/rag-hallucination-mitigation-mechanism)
- [Agent 系统哪些模块最易出问题?](https://www.wushixiongai.com/q/agent-monitoring-system-fault-localization)
- [RAG 为何要二次生成?](https://www.wushixiongai.com/q/rag-secondary-generation-necessity)
- [RAG 评测优化闭环](https://www.wushixiongai.com/q/rag-evaluation-system-and-optimization)
- [RAG 评估指标与迭代陷阱](https://www.wushixiongai.com/q/rag-evaluation-and-iteration-methods)
- [RAG 仍会幻觉的原因](https://www.wushixiongai.com/q/rag-hallucination-mitigation-effectiveness-bounds)
- [RAG 怎么减少幻觉?](https://www.wushixiongai.com/q/rag-generation-quality-improvement-mechanism)
- [RAG 错误怎么定位到检索或生成?](https://www.wushixiongai.com/q/rag-error-attribution-retriever-generator)
- [RAG 错误根源怎么定位?](https://www.wushixiongai.com/q/rag-error-source-retriever-vs-generator)
- [RAG 解决了 LLM 哪些局限?](https://www.wushixiongai.com/q/rag-concept-limitations-knowledge-intensive)
- [专业领域 RAG 链路怎么搭?](https://www.wushixiongai.com/q/rag-domain-pipeline-with-evaluation)
- [召回策略怎么选? Recall@k 怎么看?](https://www.wushixiongai.com/q/retrieval-recall-strategy-metrics)
- [RAG 召回评估指标对比](https://www.wushixiongai.com/q/retrieval-recall-method-evaluation)
- [RAG编码模型评估指标与对比](https://www.wushixiongai.com/q/rag-encoder-models-dpr-contriever)
- [RAG 三大瓶颈与迭代策略](https://www.wushixiongai.com/q/rag-bottlenecks-iterative-optimization)
- [LLM 幻觉治理方法对比](https://www.wushixiongai.com/q/eval-hallucination-methods-comparison)
- [RAG vs RLHF 幻觉治理](https://www.wushixiongai.com/q/eval-hallucination-rag-vs-rlhf)
- [医疗法律RAG：评估陷阱](https://www.wushixiongai.com/q/rag-medical-legal-pipeline-evaluation)
- [Single-Agent vs Multi-Agent 怎么选?](https://www.wushixiongai.com/q/agent-single-vs-multi-architecture)
- [BGE-M3 vs BM25 怎么选?](https://www.wushixiongai.com/q/retrieval-dense-vs-keyword-evaluation)
- [RAG 二次生成怎么减少幻觉?](https://www.wushixiongai.com/q/rag-secondary-generation-hallucination-reduction)
- [Agent 多工具怎么选?](https://www.wushixiongai.com/q/agent-tool-selection-cost-evaluation)
- [召回阶段怎么评估质量?](https://www.wushixiongai.com/q/retrieval-recall-offline-online-metrics)
- [粗排模型评估指标与离线在线方法](https://www.wushixiongai.com/q/recsys-coarse-rank-evaluation)
- [粗排模型评估指标与陷阱](https://www.wushixiongai.com/q/recsys-coarse-rank-online-evaluation)
- [Chain/Agent/Tool 怎么分工?](https://www.wushixiongai.com/q/agent-langchain-modular-components)
- [主流Agent框架优缺点与场景怎么选?](https://www.wushixiongai.com/q/agent-frameworks-comparison-overview)
- [多工具评分排序机制](https://www.wushixiongai.com/q/agent-tool-selection-confidence-decision)
- [LangChain 链式调用原理](https://www.wushixiongai.com/q/agent-langchain-chain-design-principle)
- [LangChain 关键组件与流程?](https://www.wushixiongai.com/q/agent-langchain-components-workflow)
- [RAG 缓解哪类幻觉?](https://www.wushixiongai.com/q/rag-hallucination-applicable-types)
- [RAG 自动与人工评估](https://www.wushixiongai.com/q/rag-evaluation-automatic-human-metrics)
- [搜索排序权重怎么A/B测试?](https://www.wushixiongai.com/q/search-ranking-ab-test-weight-tuning)
- [Agent 系统优缺点与发展前景](https://www.wushixiongai.com/q/agent-pros-cons-future-prospects)
- [粗排冷启动指标与路径](https://www.wushixiongai.com/q/recsys-coarse-rank-cold-start)
- [RAG向量数据库选型关键指标](https://www.wushixiongai.com/q/retrieval-vector-database-evaluation-selection)
- [知识型与行为型错误诊断](https://www.wushixiongai.com/q/eval-hallucination-rag-vs-rl-scenarios)
- [现代RAG架构有哪些改进?](https://www.wushixiongai.com/q/rag-modern-fusion-vs-pipeline)
- [用户兴趣建模：数据与特征陷阱](https://www.wushixiongai.com/q/recsys-user-interest-preference-modeling)
- [Attention 在多轮对话中有什么局限?](https://www.wushixiongai.com/q/arch-attention-limitations-multi-turn)
- [偏好感知系统架构怎么选?](https://www.wushixiongai.com/q/finetune-preference-dialogue-training-evaluation)
- [LLM 如何集成外部工具与记忆?](https://www.wushixiongai.com/q/agent-langchain-integration-architecture)
- [Agent 工具调用用 Workflow 吗？](https://www.wushixiongai.com/q/agent-tool-calling-workflow-tradeoff)
- [AI Agent 核心理解与未来趋势](https://www.wushixiongai.com/q/agent-understanding-and-outlook)
- [RAG 检索语义相似但无关怎么解?](https://www.wushixiongai.com/q/rag-high-similarity-low-relevance)
- [Reflection 和 Memory 怎么协同?](https://www.wushixiongai.com/q/agent-multi-agent-reflection-memory-implementation)
- [RAG 文档检索选型](https://www.wushixiongai.com/q/rag-document-retrieval-method-optimization)
- [MRR vs NDCG vs Precision 怎么选?](https://www.wushixiongai.com/q/eval-mrr-vs-ndcg-precision)
- [RAG 检索与生成评估](https://www.wushixiongai.com/q/rag-evaluation-retrieval-generation-dimensions)
- [RAG 有哪些主要缺陷?](https://www.wushixiongai.com/q/rag-defects-limitations-solutions)
- [RAG 上下文不准确怎么诊断?](https://www.wushixiongai.com/q/rag-context-issue-diagnosis-data-quality)
- [Reflection vs Memory 如何影响决策?](https://www.wushixiongai.com/q/agent-multi-agent-reflection-memory)
- [大模型怎么用于搜索检索?](https://www.wushixiongai.com/q/search-llm-application-techniques)
- [RAG 失败场景诊断](https://www.wushixiongai.com/q/rag-evaluation-methods-and-metrics)
- [单 Agent 与多 Agent 选型](https://www.wushixiongai.com/q/agent-single-vs-multi-by-complexity)
- [RAG 线上效果评估](https://www.wushixiongai.com/q/rag-evaluation-method-metric-system)
- [Agent项目背景与技术方案怎么写?](https://www.wushixiongai.com/q/project-agent-development-experience)
- [大模型客服评估：伪代码与指标](https://www.wushixiongai.com/q/eval-customer-service-quality-system)
- [多Agent组件优化策略与架构权衡](https://www.wushixiongai.com/q/agent-multi-agent-component-optimization)
- [RAG 解决了哪些问题?](https://www.wushixiongai.com/q/project-rag-problem-solving-impact)
- [Agent项目挑战与解决方案怎么讲?](https://www.wushixiongai.com/q/project-agent-practical-experience)
- [Agent 技术实际有效性怎么评?](https://www.wushixiongai.com/q/agent-effectiveness-vs-frameworks)
- [意图识别在 RAG 中怎么用?](https://www.wushixiongai.com/q/rag-intent-recognition-importance)
- [RAG 核心指标优先级](https://www.wushixiongai.com/q/rag-evaluation-most-critical-metrics)
- [多Agent系统优化策略](https://www.wushixiongai.com/q/agent-multi-agent-optimization-strategies)
- [Agent 项目踩坑与恢复](https://www.wushixiongai.com/q/project-agent-implementation-experience)
- [LangGraph 适用边界与替代](https://www.wushixiongai.com/q/agent-langgraph-use-cases)
- [ReAct vs 传统规划怎么选?](https://www.wushixiongai.com/q/agent-react-vs-planning-architecture)
- [RG 系统效果差怎么排查?](https://www.wushixiongai.com/q/rag-effect-issue-localization)
- [训练/推理/知识增强怎么选?](https://www.wushixiongai.com/q/eval-hallucination-multi-layer-mitigation)
- [Agent 中规划与执行怎么协同?](https://www.wushixiongai.com/q/agent-planning-and-execution)
- [RAG 效果差怎么排查？](https://www.wushixiongai.com/q/rag-generation-issue-diagnosis-steps)
- [RAG效果差：检索vs生成怎么诊断？](https://www.wushixiongai.com/q/rag-poor-output-systematic-localization)
- [检索准但生成差，怎么调优？](https://www.wushixiongai.com/q/rag-localization-and-generation-fix)
- [LangChain vs AutoGPT vs MetaGPT 怎么选?](https://www.wushixiongai.com/q/agent-frameworks-langchain-autogpt-metagpt)
- [Agent 人工干预 vs 自动处理怎么切换?](https://www.wushixiongai.com/q/agent-human-auto-switch-mechanism)
- [RAG 架构怎么工作?](https://www.wushixiongai.com/q/rag-architecture-knowledge-intensive-tradeoffs)
- [Agent 建模方法怎么选?](https://www.wushixiongai.com/q/agent-modeling-approaches)
- [大模型 Agent 优缺点怎么分析?](https://www.wushixiongai.com/q/agent-tradeoffs-and-trends)
- [共识算法如何落地多Agent？](https://www.wushixiongai.com/q/agent-multi-agent-collaboration-mechanisms)
- [ReAct vs Plan-and-Execute 错误恢复谁强?](https://www.wushixiongai.com/q/agent-react-vs-plan-execute-long-context)
- [Agent 哪个方向最易落地?](https://www.wushixiongai.com/q/agent-commercialization-direction-outlook)
- [Agent 基础能力怎么定义?](https://www.wushixiongai.com/q/agent-foundational-capability-definition)
- [Agent/RAG 评估指标与挑战](https://www.wushixiongai.com/q/rag-agent-evaluation-dimensions)
- [Agent 外部工具怎么封装?](https://www.wushixiongai.com/q/agent-external-tool-integration-design)
- [Agent三大能力评估陷阱](https://www.wushixiongai.com/q/agent-capability-coding-gui-search)
- [Query改写怎么训练与评估?](https://www.wushixiongai.com/q/agent-multi-agent-query-rewriting-training)
- [RAG 问答金标构建](https://www.wushixiongai.com/q/rag-evaluation-dataset-construction)
- [Agent/RAG 评估体系怎么搭?](https://www.wushixiongai.com/q/rag-agent-evaluation-system-design)
- [Agent/RAG 评估集构建](https://www.wushixiongai.com/q/rag-agent-eval-dataset-design)
- [Coding vs GUI vs Search 谁先落地?](https://www.wushixiongai.com/q/agent-coding-gui-search-analysis)
- [Agent 功能模块怎么分工?](https://www.wushixiongai.com/q/agent-coding-gui-search-module-roles)
- [评估集防泄漏与复现](https://www.wushixiongai.com/q/rag-eval-dataset-quality-reproducibility)
- [知识库更新策略怎么选?](https://www.wushixiongai.com/q/rag-knowledge-base-update-strategies)
- [AI Agent 四大瓶颈怎么破?](https://www.wushixiongai.com/q/agent-reliability-interpretability-bottlenecks)
- [AI Agent 核心能力现状与差距](https://www.wushixiongai.com/q/agent-ideal-capabilities-assessment)
- [Multi-Agent 训练流程关键阶段](https://www.wushixiongai.com/q/agent-multi-agent-group-training)
- [MRR 在 RAG 中怎么定义?](https://www.wushixiongai.com/q/eval-mrr-definition-rag-significance)
- [300-500意图怎么召回?](https://www.wushixiongai.com/q/rag-large-scale-intent-recognition-recall)
- [Agent 规划能力怎么训练?](https://www.wushixiongai.com/q/agent-planning-ability-training)
- [架构与训练怎么提升多轮一致性?](https://www.wushixiongai.com/q/finetune-multi-turn-dialogue-consistency)
- [大模型落地技术范式怎么选?](https://www.wushixiongai.com/q/llm-application-paradigms-comparison)
- [RAG检索偏差怎么缓解?](https://www.wushixiongai.com/q/rag-document-distribution-shift-robustness)
- [Tool Calling 技术路径怎么选?](https://www.wushixiongai.com/q/agent-tool-use-function-calling-approaches)
- [Reranker 训练数据怎么构造?](https://www.wushixiongai.com/q/finetune-reranker-training-data-construction)
- [Prompt 怎么设计和优化?](https://www.wushixiongai.com/q/prompt-design-iteration-principles)
- [RAG/蒸馏/微调怎么选?](https://www.wushixiongai.com/q/eval-hallucination-mitigation-techniques-overview)
- [Agent 记忆怎么用向量库?](https://www.wushixiongai.com/q/agent-memory-vector-store-embedding-design)
- [RAG vs SFT 场景怎么选?](https://www.wushixiongai.com/q/rag-vs-sft-when-to-use)
- [RAG 为何比 SFT 更优?](https://www.wushixiongai.com/q/rag-vs-sft-knowledge-freshness-accuracy)
- [意图识别结构化输出失败如何修复](https://www.wushixiongai.com/q/prompt-intent-recognition-structured-output)
- [Few-shot vs Zero-shot vs Fine-tuning 怎么选?](https://www.wushixiongai.com/q/prompt-few-shot-vs-finetune-tradeoff)
- [向量检索怎么保准又保真?](https://www.wushixiongai.com/q/retrieval-result-accuracy-relevance)
- [RAG 检索精度提升的三大陷阱](https://www.wushixiongai.com/q/rag-optimization-retriever-generator-architecture)
- [RAG Embedding 模型怎么选?](https://www.wushixiongai.com/q/rag-embedding-selection-key-factors)
- [RAG 工作原理与架构怎么搭?](https://www.wushixiongai.com/q/rag-principle-architecture-optimization)
- [RAG 端到端工作流程](https://www.wushixiongai.com/q/rag-principle-architecture-vs-pure-generation)
- [RAG 技术原理与架构怎么搭?](https://www.wushixiongai.com/q/rag-principle-best-practices-challenges)
- [领域 RAG 知识库构建](https://www.wushixiongai.com/q/rag-domain-knowledge-system-pipeline)
- [RAG 技术流程怎么跑?](https://www.wushixiongai.com/q/rag-technical-workflow-principle)
- [领域RAG应用链路怎么搭?](https://www.wushixiongai.com/q/rag-vertical-domain-application-pipeline)
- [大模型知识库怎么构建?](https://www.wushixiongai.com/q/rag-knowledge-base-construction-pipeline)
- [文本嵌入模型怎么选?](https://www.wushixiongai.com/q/retrieval-text-embedding-models-comparison)
- [Embedding 模型选型考虑哪些因素?](https://www.wushixiongai.com/q/retrieval-embedding-model-selection-factors)
- [RAG 向量模型怎么选?](https://www.wushixiongai.com/q/rag-embedding-model-performance-metrics)
- [RAG 检索噪声 3 大策略](https://www.wushixiongai.com/q/rag-noisy-context-mitigation-strategies)
- [RAG 深度与局限怎么破?](https://www.wushixiongai.com/q/rag-limitations-and-optimization-directions)
- [RAG 检索与生成流程详解](https://www.wushixiongai.com/q/rag-implementation-architecture-scenarios)
- [Agent 系统怎么调优优化?](https://www.wushixiongai.com/q/agent-post-training-tuning-cases)
- [大模型输出JSON怎么保正确?](https://www.wushixiongai.com/q/prompt-structured-json-output-reliability)
- [LLM 标签约束解码](https://www.wushixiongai.com/q/prompt-constrained-label-set-output)
- [Prompt Engineering 关键原则有哪些?](https://www.wushixiongai.com/q/prompt-engineering-principles-techniques)
- [SFT vs RAG 知识更新](https://www.wushixiongai.com/q/rag-vs-sft-multidimensional-comparison)
- [RAG检索相关但生成差原因](https://www.wushixiongai.com/q/rag-good-retrieval-poor-generation-tuning)
- [LLM 封闭标签如何强约束?](https://www.wushixiongai.com/q/prompt-closed-label-set-constrained-output)
- [生成质量差怎么调优?](https://www.wushixiongai.com/q/rag-generation-incoherence-off-topic-tuning)
- [生成模型回答不准怎么调?](https://www.wushixiongai.com/q/rag-generation-inaccuracy-tuning-strategies)
- [怎么提升对话模型事实一致性?](https://www.wushixiongai.com/q/finetune-dialogue-accuracy-consistency-methods)
- [大模型准确性瓶颈在哪?](https://www.wushixiongai.com/q/finetune-llm-output-accuracy-improvement)
- [动态Prompt训练数据怎么构建?](https://www.wushixiongai.com/q/finetune-dynamic-prompt-training-data)
- [RAG 基本工作流程是什么?](https://www.wushixiongai.com/q/rag-basic-workflow-description)
- [RAG 解决了哪些核心问题?](https://www.wushixiongai.com/q/rag-core-problems-and-value)
- [什么是 RAG？一句话讲清检索增强生成](https://www.wushixiongai.com/q/intro-what-is-rag)
- [RAG 和微调怎么选？](https://www.wushixiongai.com/q/intro-rag-vs-finetune)
- [RAG 完整流程分几步？](https://www.wushixiongai.com/q/intro-rag-pipeline-steps)
- [大模型幻觉是什么？怎么缓解？](https://www.wushixiongai.com/q/intro-llm-hallucination-rag)
- [什么是 Embedding 向量嵌入？](https://www.wushixiongai.com/q/intro-what-is-embedding)
- [向量数据库和普通数据库有什么区别?](https://www.wushixiongai.com/q/intro-vector-database-basics)
- [什么是 Prompt 工程?](https://www.wushixiongai.com/q/intro-what-is-prompt-engineering)
- [System Prompt 和 User Prompt 有什么区别?](https://www.wushixiongai.com/q/intro-system-vs-user-prompt)
- [Few-shot 和 Zero-shot 是什么意思?](https://www.wushixiongai.com/q/intro-few-shot-zero-shot)
- [什么是 AI Agent?和聊天机器人有什么区别?](https://www.wushixiongai.com/q/intro-what-is-agent)
- [Function Calling 是什么?](https://www.wushixiongai.com/q/intro-what-is-function-calling)
- [ReAct 框架是什么?](https://www.wushixiongai.com/q/intro-what-is-react-framework)
- [MCP 模型上下文协议是什么?](https://www.wushixiongai.com/q/intro-what-is-mcp)
- [什么是微调(SFT)?和预训练有什么区别?](https://www.wushixiongai.com/q/intro-what-is-sft)
- [LoRA 是什么?为什么能省这么多资源?](https://www.wushixiongai.com/q/intro-what-is-lora)
- [全量微调和 LoRA 微调怎么选?](https://www.wushixiongai.com/q/intro-full-vs-lora-finetune)
- [Transformer 是什么?为什么是大模型的基石?](https://www.wushixiongai.com/q/intro-what-is-transformer)
- [Token 是什么?为什么大模型按 Token 计费?](https://www.wushixiongai.com/q/intro-what-is-token)
- [上下文窗口是什么?超了怎么办?](https://www.wushixiongai.com/q/intro-context-window-basics)
- [KV Cache 是什么?为什么能加速推理?](https://www.wushixiongai.com/q/intro-what-is-kv-cache)
- [LLM 事实准确性怎么保证?](https://www.wushixiongai.com/q/rag-financial-factual-accuracy)
- [医疗法律 RAG 合规追溯](https://www.wushixiongai.com/q/rag-vertical-domain-system-design)
- [GraphRAG 原理与构建方法](https://www.wushixiongai.com/q/knowledge-graph-graphrag-principle-vs-vector)
- [知识图谱更新：增量 vs 流式](https://www.wushixiongai.com/q/knowledge-graph-large-scale-update-consistency)
- [GraphRAG原理 vs 传统RAG 怎么选?](https://www.wushixiongai.com/q/knowledge-graph-graphrag-vs-traditional-rag)
- [GraphRAG 增量更新怎么保实时?](https://www.wushixiongai.com/q/knowledge-graph-graphrag-incremental-update-strategy)
- [GraphRAG 最大技术挑战?](https://www.wushixiongai.com/q/knowledge-graph-graphrag-main-challenges)
- [GraphRAG 怎么实现复杂查询?](https://www.wushixiongai.com/q/knowledge-graph-graphrag-entity-relation-retrieval)
- [GraphRAG vs 传统 RAG 优势在哪?](https://www.wushixiongai.com/q/knowledge-graph-graphrag-structured-reasoning-advantages)
- [GraphRAG 架构原理与优势](https://www.wushixiongai.com/q/knowledge-graph-graphrag-architecture-advantages-challenges)
- [伪多模态与原生多模态差在哪?](https://www.wushixiongai.com/q/multimodal-rag-pseudo-vs-true-analysis)
- [GraphRAG 技术难点有哪些?](https://www.wushixiongai.com/q/knowledge-graph-graphrag-entity-linking-challenges)
- [GraphRAG 增量更新策略与挑战？](https://www.wushixiongai.com/q/knowledge-graph-graphrag-incremental-index-maintenance)
- [GraphRAG 知识图谱构建有哪些挑战?](https://www.wushixiongai.com/q/knowledge-graph-graphrag-entity-alignment-challenges)
- [GraphRAG 为何召回更精准?](https://www.wushixiongai.com/q/knowledge-graph-graphrag-structured-recall-precision)
- [GraphRAG 精准召回三大依赖](https://www.wushixiongai.com/q/knowledge-graph-graphrag-recall-key-dependencies)
- [GraphRAG适用场景 vs 传统RAG 有哪些?](https://www.wushixiongai.com/q/knowledge-graph-graphrag-when-to-use)
- [RAG 系统痛点怎么破?](https://www.wushixiongai.com/q/rag-pain-points-and-improvements)
- [GraphRAG 如何保证时效性与一致性？](https://www.wushixiongai.com/q/knowledge-graph-graphrag-dynamic-update-timeliness)
- [RAG+KG Agent 动态更新机制](https://www.wushixiongai.com/q/knowledge-graph-agent-dynamic-update-steps)
- [知识图谱怎么动态更新?](https://www.wushixiongai.com/q/knowledge-graph-agent-update-event-version)
- [不用图数据库能做 GraphRAG 吗?](https://www.wushixiongai.com/q/knowledge-graph-graphrag-without-graph-database)
- [GraphRAG 必须用图数据库吗?](https://www.wushixiongai.com/q/knowledge-graph-graphrag-graph-database-necessity)
- [GraphRAG创新点 vs 传统RAG 是什么?](https://www.wushixiongai.com/q/knowledge-graph-graphrag-core-ideas-innovations)
- [RAG 技术原理和实现流程](https://www.wushixiongai.com/q/rag-principle-and-workflow)
- [GraphRAG 怎么提升检索质量?](https://www.wushixiongai.com/q/knowledge-graph-graphrag-graph-structure-boost)
- [GraphRAG 怎么提升检索效果?](https://www.wushixiongai.com/q/knowledge-graph-graphrag-technical-scheme)
- [知识图谱更新怎么保证实时性与一致性?](https://www.wushixiongai.com/q/knowledge-graph-update-realtime-consistency)
- [知识图谱实时性怎么保证?](https://www.wushixiongai.com/q/knowledge-graph-agent-freshness-realtime)
- [GraphRAG 原理与实现方法](https://www.wushixiongai.com/q/knowledge-graph-graphrag-principle-use-cases)
- [GraphRAG 工作原理是什么?](https://www.wushixiongai.com/q/knowledge-graph-graphrag-principle-difference)
- [向量索引 vs 分层存储怎么选？](https://www.wushixiongai.com/q/agent-long-term-memory-storage-optimization)
- [向量库 vs 知识库 vs 符号记忆？](https://www.wushixiongai.com/q/agent-long-term-memory-storage-comparison)
- [数据/架构/推理/工具怎么选?](https://www.wushixiongai.com/q/eval-hallucination-mitigation-data-model-tools)
- [训练/推理/外部知识怎么选?](https://www.wushixiongai.com/q/eval-hallucination-mitigation-training-retrieval)
- [Multi-modal RAG 原理与挑战](https://www.wushixiongai.com/q/multimodal-rag-principle-architecture)
- [AI Agent 自主智能体实现陷阱](https://www.wushixiongai.com/q/agent-evolution-challenges-future)
- [截图界面怎么用 RAG 解析?](https://www.wushixiongai.com/q/multimodal-screenshot-rag-system-design)
- [多模态 RAG 与传统 RAG 差异在哪?](https://www.wushixiongai.com/q/multimodal-rag-vs-text-rag)
- [弱多模态+强文本模型怎么协同?](https://www.wushixiongai.com/q/multimodal-vlm-llm-collaborative-reasoning)
- [多模态 RAG 架构设计](https://www.wushixiongai.com/q/multimodal-rag-frameworks-overview)
- [截图 RAG 怎么理解界面组件?](https://www.wushixiongai.com/q/multimodal-screenshot-component-retrieval-flow)
- [截图组件功能怎么用 RAG 解释?](https://www.wushixiongai.com/q/multimodal-screenshot-visual-knowledge-explanation)
- [多模态 RAG 实现方案](https://www.wushixiongai.com/q/multimodal-rag-frameworks-components-datatypes)
- [伪 vs 真多模态 RAG 优劣在哪？](https://www.wushixiongai.com/q/multimodal-rag-pseudo-vs-true)
- [Multi-modal RAG 是什么?](https://www.wushixiongai.com/q/multimodal-rag-concept-components-scenarios)
- [RAG 中表格图片怎么处理?](https://www.wushixiongai.com/q/rag-tables-images-unstructured-documents)
- [Agent 优势与局限怎么分析?](https://www.wushixiongai.com/q/agent-advantages-limitations-trends)
- [RAG意图识别原理与检索协同](https://www.wushixiongai.com/q/rag-intent-recognition-module)
- [Agent 发展路径分哪几个阶段?](https://www.wushixiongai.com/q/agent-development-stages-roadmap)
- [Agent系统架构怎么选？](https://www.wushixiongai.com/q/agent-system-design-rationale)
- [大模型智能体系统架构设计思路](https://www.wushixiongai.com/q/agent-system-from-scratch-design)
- [Agent架构设计原则？](https://www.wushixiongai.com/q/agent-architecture-internal-prompt-principles)
- [Agent 能力怎么系统提升?](https://www.wushixiongai.com/q/agent-capability-improvement-paths)
- [动态Prompt怎么保证策略对齐?](https://www.wushixiongai.com/q/prompt-dynamic-policy-safety-alignment)
- [无标注时怎么训练 CoT?](https://www.wushixiongai.com/q/prompt-concise-cot-without-annotation)
- [JSON 输出怎么保证格式稳定?](https://www.wushixiongai.com/q/prompt-reliable-json-output)
- [CoT vs ToT vs GoT 怎么选?](https://www.wushixiongai.com/q/agent-planning-cot-tot-methods)
- [JSON 输出怎么保证稳定可靠?](https://www.wushixiongai.com/q/prompt-io-interface-structured-output)
- [Prompt 怎么系统设计与优化?](https://www.wushixiongai.com/q/prompt-systematic-design-optimization)
- [RAG Generator Prompt 整合原理](https://www.wushixiongai.com/q/rag-generator-prompt-template-design)
- [Prompt Engineering 技巧怎么选?](https://www.wushixiongai.com/q/prompt-engineering-practical-techniques)
- [Few-shot 提示怎么用?](https://www.wushixiongai.com/q/prompt-few-shot-when-and-why)
- [RAG Generator 怎么构造 Prompt?](https://www.wushixiongai.com/q/rag-generator-context-prompt-building)
- [RAG生成阶段Prompt整合策略](https://www.wushixiongai.com/q/rag-generation-context-prompt-integration)
- [Prompt 设计原则有哪些?](https://www.wushixiongai.com/q/prompt-design-construction-principles)
- [ReAct vs 纯提示：为何更优？](https://www.wushixiongai.com/q/agent-react-vs-prompting)
- [ReAct 核心思想与多步推理优势](https://www.wushixiongai.com/q/agent-react-multi-step-interaction)
- [ReAct 如何结合推理与工具使用？](https://www.wushixiongai.com/q/agent-react-advantage-over-prompting)
- [ReAct 框架为什么优于纯提示?](https://www.wushixiongai.com/q/agent-react-qa-decision-advantage)
- [Agent Prompt优化评估指标](https://www.wushixiongai.com/q/prompt-optimization-evaluation-metrics)
- [思维链(CoT)什么时候用?](https://www.wushixiongai.com/q/prompt-when-to-use-cot)
- [RAG 生成阶段 Prompt 怎么构建?](https://www.wushixiongai.com/q/rag-generation-prompt-templates)
- [Prompt 系统设计原则与技巧](https://www.wushixiongai.com/q/prompt-design-fewshot-cot-techniques)
- [LangGraph 多轮对话优势在哪?](https://www.wushixiongai.com/q/agent-langgraph-vs-manual-prompt-flow)
- [Agent推理链路Prompt优化评估指标](https://www.wushixiongai.com/q/prompt-optimization-impact-assessment)
- [Prompt优化评估指标有哪些？](https://www.wushixiongai.com/q/prompt-optimization-reasoning-chain-metrics)
- [Prompt优化怎么评估准确性与效率？](https://www.wushixiongai.com/q/prompt-optimization-accuracy-stability-efficiency)
- [怎么让模型稳定输出 JSON?](https://www.wushixiongai.com/q/prompt-json-output-prevention-retry)
- [Agent任务规划+工具调用原理](https://www.wushixiongai.com/q/agent-simple-planning-tool-reflection)
- [LLM 输出格式如何强校验?](https://www.wushixiongai.com/q/prompt-output-filtering-format-compliance)
- [RAG 改写与 Prompt 链路](https://www.wushixiongai.com/q/rag-query-rewrite-and-prompt-building)
- [电商Prompt设计陷阱](https://www.wushixiongai.com/q/prompt-ecommerce-application-design)
- [Prompt 工程最佳实践有哪些?](https://www.wushixiongai.com/q/prompt-engineering-patterns-best-practices)
- [提示工程有哪些核心方法?](https://www.wushixiongai.com/q/prompt-engineering-methods-pros-cons)
- [LLM Agent组件有哪些功能?](https://www.wushixiongai.com/q/agent-core-components-overview)
- [Agent 核心组件怎么协作?](https://www.wushixiongai.com/q/agent-working-principle-components)
- [AI Agent 架构与框架怎么选?](https://www.wushixiongai.com/q/agent-frameworks-react-cot-autogpt-comparison)
- [Agent 行为模式有哪些？](https://www.wushixiongai.com/q/agent-behavior-patterns-overview)
- [AI Agent组件如何协同工作?](https://www.wushixiongai.com/q/agent-core-components-functions)
- [Reflection 机制怎么提升协作?](https://www.wushixiongai.com/q/agent-multi-agent-reflection-mechanism)
- [意图识别 Prompt 设计要素?](https://www.wushixiongai.com/q/prompt-intent-recognition-template-best-practices)
- [RAG Prompt 模板设计陷阱](https://www.wushixiongai.com/q/rag-prompt-template-topk-integration)
- [ReAct 范式核心思想与工作流程](https://www.wushixiongai.com/q/agent-react-paradigm-workflow-advantages)
- [Agent系统Prompt设计：架构与机制](https://www.wushixiongai.com/q/agent-system-architecture-prompt-design)
- [Agent 系统架构与 Prompt 设计原则](https://www.wushixiongai.com/q/agent-architecture-to-prompt-principles)
- [Prompt 设计方法论怎么用?](https://www.wushixiongai.com/q/prompt-methodology-across-nlp-tasks)
- [AI Agent 项目实战经验](https://www.wushixiongai.com/q/project-ai-agent-experience)
- [ReAct 之外还有哪些规划方法?](https://www.wushixiongai.com/q/agent-planning-methods-beyond-react)
- [AI Agent 核心概念与架构](https://www.wushixiongai.com/q/agent-concept-architecture-scenarios)
- [意图识别 Prompt 最佳实践?](https://www.wushixiongai.com/q/prompt-intent-recognition-template-elements)
- [ReAct 架构原理与失败场景](https://www.wushixiongai.com/q/agent-react-design-workflow-benefits)
- [ReAct 框架怎么实现推理+行动?](https://www.wushixiongai.com/q/agent-react-reasoning-external-action)
- [ReAct 框架怎么工作？](https://www.wushixiongai.com/q/agent-react-qa-tool-application)
- [Coding Agent 表现差怎么排查?](https://www.wushixiongai.com/q/agent-coding-agent-failure-analysis)
- [CoT 提示怎么写?](https://www.wushixiongai.com/q/prompt-cot-writing-template-principles)
- [Agent任务规划:分解与ReAct机制](https://www.wushixiongai.com/q/agent-task-planning-decomposition-monitoring)
- [ReAct 框架核心思想与流程](https://www.wushixiongai.com/q/agent-react-core-workflow-advantage)
- [ReAct 框架怎么用?](https://www.wushixiongai.com/q/agent-react-vs-pure-reasoning)
- [ReAct 如何提升复杂任务？](https://www.wushixiongai.com/q/agent-react-interleaved-reasoning-action)
- [CoT 思维链由哪些部分构成?](https://www.wushixiongai.com/q/prompt-cot-components-structure)
- [AI Agent 核心特点与局限](https://www.wushixiongai.com/q/agent-characteristics-advantages-future)
- [封闭标签分类如何约束解码?](https://www.wushixiongai.com/q/prompt-closed-set-classification-methods)
- [Prompt 工程方法怎么选?](https://www.wushixiongai.com/q/prompt-manual-and-automated-optimization)
- [AI Agent 三大组件原理](https://www.wushixiongai.com/q/agent-planning-tool-memory-components)
- [ReAct 各模块输入输出是什么？](https://www.wushixiongai.com/q/agent-react-plan-observation-explained)
- [Agent 工具调用链路怎么搭?](https://www.wushixiongai.com/q/agent-tool-calling-pipeline)
- [判别式转生成式模板怎么选?](https://www.wushixiongai.com/q/prompt-discriminative-to-generative-templates)
- [Prompt 性能上限评估指标](https://www.wushixiongai.com/q/prompt-performance-ceiling-evaluation)
- [Prompt 优化方法有哪些?](https://www.wushixiongai.com/q/prompt-optimization-methods-overview)
- [Agent 四大基础能力详解](https://www.wushixiongai.com/q/agent-foundational-capabilities)
- [Prompt 优化策略有哪些?](https://www.wushixiongai.com/q/prompt-optimization-cot-fewshot-strategies)
- [Policy 模型错误决策怎么防?](https://www.wushixiongai.com/q/prompt-dynamic-policy-model-safeguards)
- [Agent 性能优化：架构与推理陷阱](https://www.wushixiongai.com/q/agent-performance-execution-optimization)
- [增量式向量索引怎么更新?](https://www.wushixiongai.com/q/rag-incremental-vector-index-update)
- [对话模型准确性怎么提?](https://www.wushixiongai.com/q/llm-chat-output-accuracy-techniques)
- [RAG 核心架构与工作流程](https://www.wushixiongai.com/q/rag-concept-architecture-workflow)
- [分布漂移下外接数据库怎么增强模型?](https://www.wushixiongai.com/q/rag-distribution-drift-external-database)
- [RAG vs 传统检索生成怎么选?](https://www.wushixiongai.com/q/rag-vs-retrieve-then-generate-pipeline)
- [垂直领域 RAG 完整架构](https://www.wushixiongai.com/q/rag-domain-assistant-system-design)
- [RAG 怎么提升生成准确性?](https://www.wushixiongai.com/q/rag-concept-external-knowledge-accuracy)
- [RAG 怎么提升生成质量?](https://www.wushixiongai.com/q/rag-concept-quality-factual-accuracy)
- [RAG 检索模块怎么提升相关性?](https://www.wushixiongai.com/q/rag-retriever-relevance-improvement-methods)
- [多目标优化权重怎么定?](https://www.wushixiongai.com/q/recsys-ad-multi-objective-weighting)
- [地图避障方法有哪些局限?](https://www.wushixiongai.com/q/robotics-navigation-obstacle-avoidance)
- [RAG 索引构建管线](https://www.wushixiongai.com/q/rag-index-optimization-strategies)
- [LangChain vs LlamaIndex 架构差异？](https://www.wushixiongai.com/q/rag-langchain-vs-llamaindex-comparison)
- [RAG 查不到文档怎么降级?](https://www.wushixiongai.com/q/rag-no-results-graceful-fallback)
- [高精度RAG系统设计陷阱](https://www.wushixiongai.com/q/rag-domain-assistant-key-modules)
- [医疗法律RAG架构怎么搭？](https://www.wushixiongai.com/q/rag-medical-legal-assistant-architecture)
- [RAG 检索增强技术有哪些？](https://www.wushixiongai.com/q/rag-retrieval-quality-enhancement-techniques)
- [医疗 RAG 索引如何保合规?](https://www.wushixiongai.com/q/rag-domain-data-pipeline-compliance)
- [RAG 怎么提升事实准确性?](https://www.wushixiongai.com/q/rag-concept-principle-knowledge-coverage)
- [LangChain vs LlamaIndex 场景怎么选？](https://www.wushixiongai.com/q/rag-langchain-vs-llamaindex-design-philosophy)
- [CTR/CVR/ROI 冲突怎么协调?](https://www.wushixiongai.com/q/recsys-ad-ctr-cvr-roi-tradeoff)
- [RAG 检索噪声怎么处理?](https://www.wushixiongai.com/q/rag-noisy-retrieval-generation-robustness)
- [豆包等产品多轮一致性挑战与方案?](https://www.wushixiongai.com/q/llm-multi-turn-dialogue-consistency-challenges)
- [新用户冷启动：生成式推荐与Embedding](https://www.wushixiongai.com/q/recsys-cold-start-generative-recommendation)
- [RAG 知识库怎么增量更新?](https://www.wushixiongai.com/q/rag-incremental-document-index-update)
- [搜索排序权重怎么动态调?](https://www.wushixiongai.com/q/search-ranking-dynamic-feature-weighting)
- [查询文档相关性怎么算?](https://www.wushixiongai.com/q/search-ranking-query-document-relevance)
- [LLM怎么融入推荐系统？意图理解与生成](https://www.wushixiongai.com/q/recsys-llm-generative-recommendation-paths)
- [外挂数据库怎么提升大模型准确性?](https://www.wushixiongai.com/q/rag-distribution-shift-external-augmentation)
- [多信号评分函数原理与陷阱](https://www.wushixiongai.com/q/search-ranking-scoring-function-design)
- [推荐不准的分层诊断](https://www.wushixiongai.com/q/recsys-imprecise-recommendation-improvement)
- [搜索评分如何融合点击与停留?](https://www.wushixiongai.com/q/search-engine-scoring-function-features)
- [RAG 模型分类有哪些?](https://www.wushixiongai.com/q/rag-model-taxonomy-architectures)
- [现代 RAG 架构怎么变?](https://www.wushixiongai.com/q/rag-modern-vs-classic-end-to-end)
- [时间窗口对话系统：约束与取舍](https://www.wushixiongai.com/q/llm-time-window-dialogue-understanding)
- [RAG 二次生成怎么提升安全性?](https://www.wushixiongai.com/q/rag-secondary-generation-safety-filtering)
- [偏好数据如何收集与建模?](https://www.wushixiongai.com/q/finetune-preference-aware-dialogue-training)
- [生成式推荐的序列原理](https://www.wushixiongai.com/q/recsys-generative-recommendation-principle)
- [生成式推荐系统未来怎么发展？](https://www.wushixiongai.com/q/recsys-generative-future-directions)
- [生成式与判别式推荐选型](https://www.wushixiongai.com/q/recsys-generative-recommendation-architecture)
- [用户兴趣模型：数据与特征取舍](https://www.wushixiongai.com/q/recsys-user-interest-modeling-design)
- [ROI 问题怎么影响推荐效果?](https://www.wushixiongai.com/q/recsys-generative-repetition-overfitting-diversity)
- [Query 改写有哪些创新方法?](https://www.wushixiongai.com/q/rag-query-rewriting-innovative-methods)
- [动态增量更新怎么防检索偏差?](https://www.wushixiongai.com/q/rag-incremental-update-distribution-bias)
- [搜索排序中时效性特征量化方法](https://www.wushixiongai.com/q/search-ranking-recency-feature-quantification)
- [内容热门度怎么量化?](https://www.wushixiongai.com/q/search-ranking-popularity-quantification)
- [推荐不精准怎么诊断?](https://www.wushixiongai.com/q/recsys-ecommerce-imprecision-diagnosis)
- [时间窗口对话管理机制](https://www.wushixiongai.com/q/llm-time-window-dialogue-management)
- [逻辑引导怎么用？](https://www.wushixiongai.com/q/prompt-logical-reasoning-guidance)
- [人工干预怎么控制 LLM 生成?](https://www.wushixiongai.com/q/llm-human-intervention-generation-control)
- [排序模型初始权重怎么设?](https://www.wushixiongai.com/q/search-ranking-initial-weight-setting)
- [大模型输出怎么检测和过滤?](https://www.wushixiongai.com/q/llm-output-validation-filtering)
- [RAG 技术原理与核心组件详解](https://www.wushixiongai.com/q/rag-principle-components-workflow-overview)
- [上下文错误时怎么保证输出准确?](https://www.wushixiongai.com/q/rag-incorrect-context-output-accuracy)
- [RAG 项目指标化复盘](https://www.wushixiongai.com/q/project-rag-system-experience)
- [RAG 检索准但回答差怎么调?](https://www.wushixiongai.com/q/rag-generation-quality-systematic-tuning)
- [LLM 封闭标签方法对比](https://www.wushixiongai.com/q/prompt-closed-set-label-classification)
- [Deep Research vs RAG 怎么选?](https://www.wushixiongai.com/q/rag-deep-research-vs-rag-comparison)
- [医疗法律 RAG 架构关键环节](https://www.wushixiongai.com/q/rag-domain-system-architecture-design)
- [RAG检索质量提升5法](https://www.wushixiongai.com/q/rag-retrieval-relevance-boosting-methods)
- [RAG Retriever 工作流程详解](https://www.wushixiongai.com/q/rag-retriever-full-workflow-detail)
- [专业领域RAG链路优化陷阱](https://www.wushixiongai.com/q/rag-domain-assistant-pipeline-optimization)
- [向量数据库怎么支撑 Agent 记忆?](https://www.wushixiongai.com/q/agent-memory-vector-db-embedding)
- [RAG 瓶颈在哪一步](https://www.wushixiongai.com/q/rag-workflow-performance-bottleneck-analysis)
- [RAG 检索偏差怎么系统性解决?](https://www.wushixiongai.com/q/rag-irrelevant-retrieval-systematic-fix)
- [RAG 为什么需要 Re-ranking?](https://www.wushixiongai.com/q/rerank-why-needed-mechanism)
- [RAG 原理与核心架构](https://www.wushixiongai.com/q/rag-architecture-factual-accuracy-hallucination)
- [RAG流程：检索到生成各阶段陷阱](https://www.wushixiongai.com/q/rag-query-to-answer-workflow)
- [RAG检索不准的失败场景与优化](https://www.wushixiongai.com/q/rag-retrieval-miss-causes-optimization)
- [RAG 系统迭代优化方向有哪些?](https://www.wushixiongai.com/q/rag-iterative-optimization-directions)
- [精排精度高，为何不跳过召回粗排？](https://www.wushixiongai.com/q/rerank-recall-coarse-fine-funnel)
- [直接精排所有候选，为何不可行？](https://www.wushixiongai.com/q/recsys-recall-coarse-fine-necessity)
- [RAG 流程各阶段怎么拆?](https://www.wushixiongai.com/q/rag-end-to-end-retrieval-rerank-generation)
- [RAG 完整工作流程怎么跑?](https://www.wushixiongai.com/q/rag-full-pipeline-context-construction)
- [RAG 检索优化:重排序怎么用?](https://www.wushixiongai.com/q/rag-retrieval-relevance-methods)
- [RAG 为何还要重排?](https://www.wushixiongai.com/q/rerank-after-initial-retrieval-impact)
- [记忆型Agent为何用向量数据库?](https://www.wushixiongai.com/q/agent-memory-vector-db-design)
- [RAG 重排序怎么选?](https://www.wushixiongai.com/q/rerank-candidate-chunks-methods)
- [RAG 中 Re-ranking 怎么用?](https://www.wushixiongai.com/q/rerank-role-and-common-models)
- [召回和粗排为什么不能跳过?](https://www.wushixiongai.com/q/recsys-recall-coarse-stage-roles)
- [RAG 重排收益与成本](https://www.wushixiongai.com/q/rerank-quality-improvement-after-retrieval)
- [RAG 系统怎么持续迭代?](https://www.wushixiongai.com/q/rag-continuous-iteration-optimization)
- [RAG 重排作用与实现](https://www.wushixiongai.com/q/rerank-module-purpose-implementation)
- [Naive RAG vs Advanced RAG 区别?](https://www.wushixiongai.com/q/rag-naive-vs-advanced-taxonomy)
- [RAG 流程各环节优化陷阱](https://www.wushixiongai.com/q/rag-workflow-with-stage-optimizations)
- [RAG 流程：检索、融合、生成](https://www.wushixiongai.com/q/rag-complete-technical-process-functions)
- [RAG 延迟与准确率优化](https://www.wushixiongai.com/q/rag-agent-latency-accuracy-optimization)
- [RAG流程优化：三大层面策略](https://www.wushixiongai.com/q/rag-pipeline-optimization-methods)
- [RAG 为什么需要 Rerank?](https://www.wushixiongai.com/q/rerank-purpose-advantages)
- [两阶段检索 vs 直接重排?](https://www.wushixiongai.com/q/rag-two-stage-recall-rerank-rationale)
- [RAG 延迟准确率根因](https://www.wushixiongai.com/q/rag-agent-latency-accuracy-root-causes)
- [RAG Query 改写方法](https://www.wushixiongai.com/q/rag-query-rewrite-prompt-construction)
- [RAG增强与生成如何衔接？](https://www.wushixiongai.com/q/rag-three-stages-implementation-detail)
- [RAG 工作原理与核心组件](https://www.wushixiongai.com/q/rag-principle-core-components)
- [RAG 技术实现流程怎么走?](https://www.wushixiongai.com/q/rag-implementation-document-to-generation)
- [RAG 工作流程与关键技术组件](https://www.wushixiongai.com/q/rag-basic-workflow-components-challenges)
- [RAG 文档处理与检索策略怎么搭？](https://www.wushixiongai.com/q/rag-typical-architecture-implementation)
- [RAG 组件功能与数据流向?](https://www.wushixiongai.com/q/rag-workflow-components-dataflow)
- [RAG检索方法选型原因？](https://www.wushixiongai.com/q/rag-retrieval-method-choice-rationale)
- [RAG 核心实现原理是什么?](https://www.wushixiongai.com/q/rag-core-implementation-architecture)
- [BM25/向量/混合检索选型](https://www.wushixiongai.com/q/rag-retrieval-method-selection-scenarios)
- [专业领域RAG数据与检索优化？](https://www.wushixiongai.com/q/rag-domain-architecture-retrieval-generation)
- [RAG检索方法如何优化？](https://www.wushixiongai.com/q/rag-retrieval-methods-selection-optimization)
- [RAG 架构组件与失败场景](https://www.wushixiongai.com/q/rag-architecture-component-functions-flow)
- [RAG 流程优化：检索与重排陷阱](https://www.wushixiongai.com/q/rag-pipeline-per-stage-optimization)
- [RAG 怎么解决大模型幻觉?](https://www.wushixiongai.com/q/rag-principle-components-hallucination-advantage)
- [RAG 架构：检索到生成全链路](https://www.wushixiongai.com/q/rag-implementation-architecture-dataflow)
- [RAG 局限性怎么破?](https://www.wushixiongai.com/q/rag-limitations-challenges-improvements)
- [RAG 系统实现：组件与流程](https://www.wushixiongai.com/q/rag-implementation-components-workflow)
- [领域 RAG 架构如何控幻觉?](https://www.wushixiongai.com/q/rag-domain-assistant-implementation-plan)
- [RAG架构：检索器与重排序策略](https://www.wushixiongai.com/q/rag-architecture-retriever-optimization)
- [RAG 检索器与知识库怎么搭?](https://www.wushixiongai.com/q/project-rag-technical-details)
- [RAG 两阶段架构：模块与数据流](https://www.wushixiongai.com/q/rag-two-stage-document-module-dataflow)
- [RAG 架构怎么搭?](https://www.wushixiongai.com/q/rag-retriever-generator-collaboration)
- [RAG 系统核心组件怎么搭?](https://www.wushixiongai.com/q/rag-components-document-retriever-generator)
- [BM25+向量+Cross-Encoder 怎么组合?](https://www.wushixiongai.com/q/rag-hybrid-sparse-dense-rerank)
- [RAG检索优化：向量检索与重排序](https://www.wushixiongai.com/q/rag-retrieval-stage-optimization)
- [Context vs Prompt Engineering 区别?](https://www.wushixiongai.com/q/prompt-context-engineering-vs-prompt-engineering)
- [RAG 检索器与生成器如何协同？](https://www.wushixiongai.com/q/rag-typical-architecture-retriever-generator)
- [RAG 系统有哪些常见陷阱?](https://www.wushixiongai.com/q/rag-implementation-pitfalls-best-practices)
- [推荐结果多样性怎么控制?](https://www.wushixiongai.com/q/recsys-generative-output-diversity-control)
- [RAG 系统实现流程怎么搭?](https://www.wushixiongai.com/q/rag-workflow-with-result-fusion)
- [语义检索向量数据库构建流程](https://www.wushixiongai.com/q/retrieval-semantic-vector-database-build)
- [两阶段 vs 端到端怎么选?](https://www.wushixiongai.com/q/rag-two-stage-vs-end-to-end)
- [RAG 召回准确性：三大失败场景](https://www.wushixiongai.com/q/rag-recall-relevance-evaluation-improvement)
- [RAG 为何需要重排序?](https://www.wushixiongai.com/q/rerank-beyond-vector-similarity-necessity)
- [向量数据库 vs 传统数据库怎么选？](https://www.wushixiongai.com/q/rag-vector-vs-traditional-database-storage)
- [RAG 三阶段原理与失败场景](https://www.wushixiongai.com/q/rag-three-stage-workflow-overview)
- [混合检索怎么提升 RAG 召回?](https://www.wushixiongai.com/q/retrieval-hybrid-sparse-dense-design)
- [RAG Retriever 工作机制详解](https://www.wushixiongai.com/q/rag-retriever-module-mechanism)
- [Embedding 怎么实现语义召回?](https://www.wushixiongai.com/q/retrieval-embedding-cosine-semantic-recall)
- [BM25 原理与 TF-IDF 改进点](https://www.wushixiongai.com/q/retrieval-okapi-bm-saturation-normalization)
- [混合检索为什么优于单一检索?](https://www.wushixiongai.com/q/retrieval-hybrid-sparse-dense-rationale)
- [BM25 vs 向量检索 vs 混合检索怎么选?](https://www.wushixiongai.com/q/retrieval-keyword-vector-hybrid-comparison)
- [召回 vs 粗排怎么分工?](https://www.wushixiongai.com/q/recsys-recall-pre-ranking-roles)
- [召回阶段怎么选方法?](https://www.wushixiongai.com/q/recsys-recall-methods-comparison)
- [BM25公式参数含义与TF-IDF对比](https://www.wushixiongai.com/q/retrieval-okapi-bm-formula-parameters)
- [RAG 召回又快又准靠什么?](https://www.wushixiongai.com/q/rag-recall-pipeline-optimization-efficiency)
- [RAG检索器与索引结构怎么选？](https://www.wushixiongai.com/q/rag-recall-pipeline-retriever-index-rewrite)
- [召回模型怎么搭建?](https://www.wushixiongai.com/q/recsys-recall-model-build-pipeline)
- [稀疏稠密与混合检索](https://www.wushixiongai.com/q/retrieval-sparse-dense-similarity-hybrid)
- [混合检索为什么用 Sparse+Dense?](https://www.wushixiongai.com/q/retrieval-hybrid-why-similarity-measures)
- [多路召回 vs 排序模型怎么分工?](https://www.wushixiongai.com/q/recsys-multi-channel-recall-ranking)
- [RAG 工作流怎么应对时效衰减?](https://www.wushixiongai.com/q/rag-workflow-information-freshness-decay)
- [RAG 向量库怎么处理时间衰减?](https://www.wushixiongai.com/q/rag-workflow-time-decay-handling)
- [Embedding升级后向量对齐方案](https://www.wushixiongai.com/q/retrieval-embedding-version-migration-consistency)
- [LLM怎么融入推荐系统？特征提取与召回](https://www.wushixiongai.com/q/recsys-llm-integration-paths)
- [LLM怎么融入推荐系统？序列建模与排序](https://www.wushixiongai.com/q/recsys-llm-sequence-modeling-recall)
- [向量数据库选型关键因素有哪些？](https://www.wushixiongai.com/q/retrieval-vector-database-selection-factors)
- [RAG 向量检索怎么高效召回?](https://www.wushixiongai.com/q/rag-retriever-efficient-recall-factors)
- [BM25 公式与 TF-IDF 对比](https://www.wushixiongai.com/q/retrieval-okapi-bm-vs-tfidf-improvements)
- [Embedding + 向量数据库怎么搭?](https://www.wushixiongai.com/q/retrieval-embedding-vector-db-semantic-search)
- [BM25 参数 k1/b 调优](https://www.wushixiongai.com/q/retrieval-okapi-bm-principle-advantages)
- [中文知识库检索选型](https://www.wushixiongai.com/q/retrieval-dense-sparse-methods-pros-cons)
- [推荐不准的算法改进](https://www.wushixiongai.com/q/recsys-recommendation-accuracy-improvement)
- [Embedding 模型结构选型](https://www.wushixiongai.com/q/retrieval-embedding-model-architectures-dimensions)
- [Hard Search vs Soft Search 怎么选?](https://www.wushixiongai.com/q/recsys-hard-vs-soft-search)
- [BM25原理、应用场景及与TF-IDF区别](https://www.wushixiongai.com/q/retrieval-okapi-bm-principle-comparison)
- [Embedding 相似性检索原理与索引构建](https://www.wushixiongai.com/q/retrieval-embedding-similarity-principle)
- [检索失败时 RAG 怎么兜底?](https://www.wushixiongai.com/q/rag-no-results-robustness-strategy)
- [Embedding 向量维度选型](https://www.wushixiongai.com/q/retrieval-embedding-architectures-dimension-impact)
- [向量召回实现方式有哪些?](https://www.wushixiongai.com/q/retrieval-vector-recall-implementations)
- [向量召回原理与实现?](https://www.wushixiongai.com/q/retrieval-vector-recall-large-scale)
- [向量检索加速：有哪些陷阱](https://www.wushixiongai.com/q/rag-retriever-performance-optimization)
- [向量检索 4 种范式怎么选?](https://www.wushixiongai.com/q/retrieval-ann-tree-hash-graph-methods)
- [RAG检索质量差：Embedding与Rerank调优](https://www.wushixiongai.com/q/rag-poor-retrieval-context-strategies)
- [RAG 技术实现怎么落地?](https://www.wushixiongai.com/q/rag-implementation-scenarios-advantages)
- [RAG文本分段策略与重叠机制](https://www.wushixiongai.com/q/rag-chunk-segmentation-retrieval-tradeoff)
- [ANN 索引怎么构建?](https://www.wushixiongai.com/q/retrieval-ann-index-build-pipeline)
- [RAG 知识库怎么存?](https://www.wushixiongai.com/q/rag-knowledge-base-storage-metadata-index)
- [RAG向量匹配原理与ANN算法取舍](https://www.wushixiongai.com/q/retrieval-ann-hnsw-ivf-precision-efficiency)
- [RAG 向量匹配怎么高效实现?](https://www.wushixiongai.com/q/retrieval-similarity-matching-ann-tradeoffs)
- [RAG 知识库存储怎么选?](https://www.wushixiongai.com/q/rag-knowledge-base-storage-design)
- [Embedding 相似性检索怎么工作?](https://www.wushixiongai.com/q/retrieval-embedding-similarity-basics)
- [向量检索基本原理是什么?](https://www.wushixiongai.com/q/retrieval-vector-search-fundamentals)