https://guijiagi.com/categories/rag%E4%B8%8E%E5%BE%AE%E8%B0%83/Recent content in RAG与微调 on 硅基 AGI · 智能体学习与测评Hugozh-cn本站内容采用 CC BY-NC-SA 4.0 国际许可协议授权Thu, 16 Jul 2026 11:31:00 +0800https://guijiagi.com/posts/b1-c4dac6e8/Thu, 16 Jul 2026 11:31:00 +0800https://guijiagi.com/posts/b1-c4dac6e8/系统讲解Embedding模型的选型策略、微调方法和评估指标，助力构建高质量向量检索系统https://guijiagi.com/posts/b1-77165612/Thu, 16 Jul 2026 11:26:00 +0800https://guijiagi.com/posts/b1-77165612/深入解析知识蒸馏技术在大模型时代的应用，涵盖响应蒸馏、特征蒸馏与Agent蒸馏https://guijiagi.com/posts/b1-ca24c427/Thu, 16 Jul 2026 11:04:00 +0800https://guijiagi.com/posts/b1-ca24c427/对比分析LoRA、QLoRA和全参微调的适用场景、技术细节与工程实践指南https://guijiagi.com/posts/b1-f95dfde9/Thu, 16 Jul 2026 11:02:00 +0800https://guijiagi.com/posts/b1-f95dfde9/从基础RAG到高级RAG的完整技术路线，涵盖混合检索、重排序、查询改写、自反思等关键技术https://guijiagi.com/posts/b2-17d2f42e/Thu, 16 Jul 2026 10:22:00 +0800https://guijiagi.com/posts/b2-17d2f42e/系统梳理大模型知识蒸馏技术体系，涵盖logits蒸馏、特征蒸馏、关系蒸馏及多教师蒸馏的实践方法https://guijiagi.com/posts/b2-39472406/Thu, 16 Jul 2026 10:19:00 +0800https://guijiagi.com/posts/b2-39472406/深入分析自适应检索增强生成技术，探讨模型如何根据问题难度动态决定检索策略https://guijiagi.com/posts/b2-dae4b598/Thu, 16 Jul 2026 10:03:00 +0800https://guijiagi.com/posts/b2-dae4b598/从LoRA数学原理到工程实现，覆盖秩选择、学习率配置、QLoRA及常见微调陷阱的完整指南https://guijiagi.com/posts/b2-922f32c3/Thu, 16 Jul 2026 10:02:00 +0800https://guijiagi.com/posts/b2-922f32c3/从BM25到ColBERT，深入探讨RAG系统中混合检索策略、交叉编码器重排序及检索质量评估方法https://guijiagi.com/posts/article-85/Mon, 13 Jul 2026 07:00:00 +0800https://guijiagi.com/posts/article-85/系统梳理大模型对齐的完整流程，从数据标注到SFT到RLHF再到DPO的工程实践https://guijiagi.com/posts/article-67/Mon, 13 Jul 2026 04:00:00 +0800https://guijiagi.com/posts/article-67/深入分析大模型微调中灾难性遗忘的成因、影响和缓解策略，涵盖正则化、回放、参数隔离等方法https://guijiagi.com/posts/article-52/Mon, 13 Jul 2026 01:30:00 +0800https://guijiagi.com/posts/article-52/系统性介绍大模型训练数据质量评估的多维框架，包括准确性、多样性、时效性、安全性等核心维度https://guijiagi.com/posts/article-49/Mon, 13 Jul 2026 01:00:00 +0800https://guijiagi.com/posts/article-49/从logit蒸馏到agent蒸馏，知识蒸馏技术的最新进展与实践指南https://guijiagi.com/posts/b2-7b5c9e69/Mon, 13 Jul 2026 01:00:00 +0800https://guijiagi.com/posts/b2-7b5c9e69/从logit蒸馏到agent蒸馏，知识蒸馏技术的最新进展与实践指南https://guijiagi.com/posts/article-36/Sun, 12 Jul 2026 22:50:00 +0800https://guijiagi.com/posts/article-36/从数据采集到质量评估，大模型微调数据工程的完整实践指南https://guijiagi.com/posts/b2-e781991d/Sun, 12 Jul 2026 22:50:00 +0800https://guijiagi.com/posts/b2-e781991d/从数据采集到质量评估，大模型微调数据工程的完整实践指南https://guijiagi.com/posts/article-28/Sun, 12 Jul 2026 21:30:00 +0800https://guijiagi.com/posts/article-28/将知识图谱引入RAG系统，解决传统向量检索的语义鸿沟与多跳推理难题https://guijiagi.com/posts/b2-55210222/Sun, 12 Jul 2026 21:30:00 +0800https://guijiagi.com/posts/b2-55210222/将知识图谱引入RAG系统，解决传统向量检索的语义鸿沟与多跳推理难题https://guijiagi.com/posts/article-02/Sun, 12 Jul 2026 17:10:00 +0800https://guijiagi.com/posts/article-02/探讨GraphRAG如何通过知识图谱增强检索增强生成，解决传统向量检索的多跳推理瓶颈https://guijiagi.com/posts/b2-7138b571/Sun, 12 Jul 2026 17:10:00 +0800https://guijiagi.com/posts/b2-7138b571/探讨GraphRAG如何通过知识图谱增强检索增强生成，解决传统向量检索的多跳推理瓶颈https://guijiagi.com/posts/hermes4-finetuning-guide/Wed, 08 Jul 2026 13:00:00 +0800https://guijiagi.com/posts/hermes4-finetuning-guide/Nous Hermes 4微调完整指南：数据准备、LoRA训练、评估调优、量化部署的企业级实践https://guijiagi.com/posts/rag-from-scratch-2026/Thu, 02 Jul 2026 10:49:00 +0800https://guijiagi.com/posts/rag-from-scratch-2026/从零开始搭建生产级RAG系统的完整实战指南，覆盖架构、实现与优化https://guijiagi.com/posts/model-merge-techniques/Thu, 02 Jul 2026 10:48:00 +0800https://guijiagi.com/posts/model-merge-techniques/深入探讨LLM模型合并技术的原理、方法与工程实践，实现多模型能力融合https://guijiagi.com/posts/data-augmentation-llm/Thu, 02 Jul 2026 10:47:00 +0800https://guijiagi.com/posts/data-augmentation-llm/系统介绍大语言模型训练中的数据增强技术，从简单变换到AI生成增强https://guijiagi.com/posts/continual-learning-llm/Thu, 02 Jul 2026 10:46:00 +0800https://guijiagi.com/posts/continual-learning-llm/探讨大语言模型持续学习的方法、挑战与工程实践，防止灾难性遗忘https://guijiagi.com/posts/instruction-tuning-recipe/Thu, 02 Jul 2026 10:45:00 +0800https://guijiagi.com/posts/instruction-tuning-recipe/深入探讨LLM指令微调的数据构建、配方设计与训练最佳实践https://guijiagi.com/posts/rlhf-implementation-2026/Thu, 02 Jul 2026 10:44:00 +0800https://guijiagi.com/posts/rlhf-implementation-2026/系统介绍2026年RLHF的实现方法、工具链与最佳实践https://guijiagi.com/posts/dpo-training-practice/Thu, 02 Jul 2026 10:43:00 +0800https://guijiagi.com/posts/dpo-training-practice/系统介绍DPO（Direct Preference Optimization）的训练原理、数据准备与工程实践https://guijiagi.com/posts/lora-vs-qlora-2026/Thu, 02 Jul 2026 10:42:00 +0800https://guijiagi.com/posts/lora-vs-qlora-2026/系统对比LoRA与QLoRA在2026年的技术差异、性能表现与适用场景https://guijiagi.com/posts/self-rag-implementation/Thu, 02 Jul 2026 10:41:00 +0800https://guijiagi.com/posts/self-rag-implementation/深入探讨Self-RAG的原理、实现方法与生产实践，让模型自主控制检索行为https://guijiagi.com/posts/rag-reranking-2026/Thu, 02 Jul 2026 10:40:00 +0800https://guijiagi.com/posts/rag-reranking-2026/系统介绍2026年RAG重排序技术的原理、模型对比与工程实践https://guijiagi.com/posts/rag-chunking-strategies/Thu, 02 Jul 2026 10:39:00 +0800https://guijiagi.com/posts/rag-chunking-strategies/系统对比RAG系统中不同文档分块策略的原理、优劣势与最佳实践https://guijiagi.com/posts/rag-evaluation-metrics/Thu, 02 Jul 2026 10:38:00 +0800https://guijiagi.com/posts/rag-evaluation-metrics/构建RAG系统全维度评估指标体系，从检索质量到生成质量的科学度量方法https://guijiagi.com/posts/multi-modal-rag-2026/Thu, 02 Jul 2026 10:37:00 +0800https://guijiagi.com/posts/multi-modal-rag-2026/深入探讨多模态RAG系统的架构设计、跨模态检索与融合生成实践https://guijiagi.com/posts/hybrid-rag-graph-vector/Thu, 02 Jul 2026 10:36:00 +0800https://guijiagi.com/posts/hybrid-rag-graph-vector/深入探讨图检索与向量检索的混合架构，发挥结构化与非结构化检索的协同优势https://guijiagi.com/posts/advanced-rag-2026-patterns/Thu, 02 Jul 2026 10:35:00 +0800https://guijiagi.com/posts/advanced-rag-2026-patterns/系统解析2026年高级RAG架构的设计模式，包括多跳检索、自适应检索与推理增强检索https://guijiagi.com/posts/alignment-evolution-rlhf-to-orpo/Tue, 30 Jun 2026 17:10:00 +0800https://guijiagi.com/posts/alignment-evolution-rlhf-to-orpo/大模型对齐技术演进：RLHF、DPO、ORPO、GRPO等核心算法原理、实现要点与2026年最新进展https://guijiagi.com/posts/agentic-rag-evolution-2026/Tue, 30 Jun 2026 16:50:00 +0800https://guijiagi.com/posts/agentic-rag-evolution-2026/RAG 技术全景：从基础检索增强生成到 Agentic RAG 的架构演进，对比主流方案并给出选型指南https://guijiagi.com/posts/agentic-rag-architecture-revolution/Tue, 30 Jun 2026 09:40:00 +0800https://guijiagi.com/posts/agentic-rag-architecture-revolution/Agentic RAG将检索增强生成从被动管道升级为主动决策系统，本文深入解析其架构设计、核心模式与工程实践https://guijiagi.com/posts/dpo-preference-alignment-guide/Tue, 30 Jun 2026 09:40:00 +0800https://guijiagi.com/posts/dpo-preference-alignment-guide/DPO已成为偏好对齐的主流方案，本文详解DPO的数据构建、训练配置与超参调优全流程https://guijiagi.com/posts/embedding-models-2026-benchmark/Tue, 30 Jun 2026 09:40:00 +0800https://guijiagi.com/posts/embedding-models-2026-benchmark/深度测评2026年主流Embedding模型在中文检索场景的表现，包括BGE、GTE、text-embedding-3等10款模型https://guijiagi.com/posts/graphrag-2026-knowledge-graph-retrieval-guide/Tue, 30 Jun 2026 09:40:00 +0800https://guijiagi.com/posts/graphrag-2026-knowledge-graph-retrieval-guide/深入探讨GraphRAG在2026年的最新实践，从知识图谱构建到图谱增强检索的完整架构与代码实现https://guijiagi.com/posts/graphrag-2026-practice-guide/Tue, 30 Jun 2026 09:40:00 +0800https://guijiagi.com/posts/graphrag-2026-practice-guide/深入解析GraphRAG在2026年的最新实践，从知识图谱构建到社区检测检索，附完整代码示例与性能基准https://guijiagi.com/posts/lora-finetuning-2026-guide/Tue, 30 Jun 2026 09:40:00 +0800https://guijiagi.com/posts/lora-finetuning-2026-guide/2026年LoRA微调的最新实践指南，涵盖数据工程、训练配置、评估优化到生产部署的完整流程https://guijiagi.com/posts/qlora-quantized-finetuning-guide/Tue, 30 Jun 2026 09:40:00 +0800https://guijiagi.com/posts/qlora-quantized-finetuning-guide/QLoRA让7B模型微调只需8GB显存，70B模型只需24GB，本文详解QLoRA原理与实战配置https://guijiagi.com/posts/rag-vs-long-context-comparison/Tue, 30 Jun 2026 09:40:00 +0800https://guijiagi.com/posts/rag-vs-long-context-comparison/深度对比RAG与长上下文模型的优劣势，通过实测数据给出清晰的选型决策框架https://guijiagi.com/posts/rag-chunking-strategies-comparison/Tue, 30 Jun 2026 09:40:00 +0800https://guijiagi.com/posts/rag-chunking-strategies-comparison/分块是RAG系统中影响检索质量最关键的环节，本文深度对比六种分块策略的原理、效果与适用场景https://guijiagi.com/posts/rag-evaluation-framework-ragas/Tue, 30 Jun 2026 09:40:00 +0800https://guijiagi.com/posts/rag-evaluation-framework-ragas/系统介绍RAGAS评估框架的核心指标体系、自动化评估流程，以及如何构建自定义评估pipelinehttps://guijiagi.com/posts/rag-production-pitfalls-guide/Tue, 30 Jun 2026 09:40:00 +0800https://guijiagi.com/posts/rag-production-pitfalls-guide/RAG系统上线后最常见的三大难题的实战解决方案，来自生产环境的真实排坑经验https://guijiagi.com/posts/reranker-model-comparison/Tue, 30 Jun 2026 09:40:00 +0800https://guijiagi.com/posts/reranker-model-comparison/Reranker是RAG系统中提升检索精度的关键组件，本文深度对比三大主流Reranker方案https://guijiagi.com/posts/llm-distillation-gpt5-to-7b/Tue, 30 Jun 2026 09:40:00 +0800https://guijiagi.com/posts/llm-distillation-gpt5-to-7b/知识蒸馏是让小模型获得大模型能力的核心技术，本文详解白盒蒸馏、黑盒蒸馏与混合蒸馏三种方案https://guijiagi.com/posts/multimodal-rag-architecture/Tue, 30 Jun 2026 09:40:00 +0800https://guijiagi.com/posts/multimodal-rag-architecture/多模态RAG让AI能同时理解和检索图片与文字，本文详解四种架构设计与Claude 4/GPT-5的实践方案https://guijiagi.com/posts/vector-database-2026-comparison/Tue, 30 Jun 2026 09:40:00 +0800https://guijiagi.com/posts/vector-database-2026-comparison/从性能、功能、成本、生态四个维度深度对比四大向量数据库，附2026年最新基准测试数据https://guijiagi.com/posts/llm-distillation-2026/Tue, 30 Jun 2026 09:20:00 +0800https://guijiagi.com/posts/llm-distillation-2026/系统介绍2026年大模型知识蒸馏技术的最新进展，涵盖白盒蒸馏、黑盒蒸馏、思维链蒸馏等方法https://guijiagi.com/posts/rag-evolution-2026/Tue, 30 Jun 2026 09:10:00 +0800https://guijiagi.com/posts/rag-evolution-2026/2026年RAG技术全景：从Naive RAG到Agentic RAG，GraphRAG、Multi-modal RAG、Adaptive RAG等前沿方案的架构解析与实战对比https://guijiagi.com/posts/agentic-rag-when-rag-meets-agent/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/agentic-rag-when-rag-meets-agent/探索 Agentic RAG 的设计理念、架构模式和工程实现，理解为什么 Agent 是 RAG 的自然进化方向https://guijiagi.com/posts/dpo-training-practice-preference-alignment-data-engineering/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/dpo-training-practice-preference-alignment-data-engineering/深入解析 DPO（Direct Preference Optimization）训练流程，涵盖偏好数据构建、训练配置和效果评估https://guijiagi.com/posts/embedding-model-2026-chinese-retrieval-benchmark/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/embedding-model-2026-chinese-retrieval-benchmark/2026年主流Embedding模型在中文检索场景的全面排行与实测对比https://guijiagi.com/posts/graphrag-production-deployment-guide/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/graphrag-production-deployment-guide/从零到一搭建 GraphRAG 生产系统，涵盖知识图谱构建、社区检测、查询引擎和运维监控https://guijiagi.com/posts/grpo-algorithm-deepseek-reinforcement-learning/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/grpo-algorithm-deepseek-reinforcement-learning/深入解析 DeepSeek 提出的 GRPO（Group Relative Policy Optimization）算法，理解其原理、优势和实践方法https://guijiagi.com/posts/lora-finetuning-2026-data-to-deployment/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/lora-finetuning-2026-data-to-deployment/手把手教你用 LoRA 微调大语言模型，涵盖数据工程、训练配置、评估和部署的全流程https://guijiagi.com/posts/rag-troubleshooting-hallucination-miss-retrieval-latency/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/rag-troubleshooting-hallucination-miss-retrieval-latency/系统梳理 RAG 系统中的高频故障模式，提供从症状到根因的诊断框架和修复方案https://guijiagi.com/posts/rag-chunking-strategies-semantic-vs-document-aware-vs-hierarchical/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/rag-chunking-strategies-semantic-vs-document-aware-vs-hierarchical/深入对比 2026 年主流 RAG 分块策略，包含实验数据、代码实现和选型建议https://guijiagi.com/posts/rag-architecture-2026-naive-to-graphrag-agent/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/rag-architecture-2026-naive-to-graphrag-agent/系统梳理 RAG 架构演进，从最基础的 Naive RAG 到 2026 年最前沿的 GraphRAG+Agent 范式，附完整架构对比与代码示例https://guijiagi.com/posts/rag-evaluation-framework-2026-ragas-custom-metrics/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/rag-evaluation-framework-2026-ragas-custom-metrics/系统介绍 RAG 系统的评估方法论，涵盖 RAGAS 框架、自定义指标设计和端到端评测流水线https://guijiagi.com/posts/rag-data-pipeline-pdf-html-database-to-knowledge-base/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/rag-data-pipeline-pdf-html-database-to-knowledge-base/系统讲解 RAG 数据管道的工程设计，涵盖文档解析、清洗、增强、质量控制和自动化运维https://guijiagi.com/posts/rag-reranking-cohere-bge-jina-comparison/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/rag-reranking-cohere-bge-jina-comparison/深入对比三大主流重排序模型，包含性能基准、成本分析和工程选型建议https://guijiagi.com/posts/reranker-model-selection-2026-cohere-bge-jina/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/reranker-model-selection-2026-cohere-bge-jina/2026年主流Reranker模型深度对比：Cohere、BGE、Jina在中文与英文检索重排场景的全面评测https://guijiagi.com/posts/sft-data-quality-assessment-bad-data-destroys-finetuning/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/sft-data-quality-assessment-bad-data-destroys-finetuning/系统分析 SFT 数据质量问题对微调效果的影响，提供数据清洗、质量评估和持续监控的工程方案https://guijiagi.com/posts/continual-pretraining-domain-knowledge-injection/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/continual-pretraining-domain-knowledge-injection/深入讲解大模型持续预训练（CPT）的工程实践，涵盖数据配比、训练策略、灾难性遗忘的预防和评估方法https://guijiagi.com/posts/llm-evaluation-pipeline-benchmark-to-custom/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/llm-evaluation-pipeline-benchmark-to-custom/系统讲解大模型评估流水线的搭建方法，涵盖主流 Benchmark 集成、自定义评测设计和自动化 CI/CDhttps://guijiagi.com/posts/finetuning-cost-analysis-lora-qlora-full-parameter/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/finetuning-cost-analysis-lora-qlora-full-parameter/深入分析三种主流微调方法的成本结构，涵盖显存、计算、存储和部署的全面费用对比https://guijiagi.com/posts/finetuning-toolchain-2026-llamafactory-axolotl-unsloth/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/finetuning-toolchain-2026-llamafactory-axolotl-unsloth/三大微调工具链全面对比：LLaMA-Factory、Axolotl、Unsloth的功能、性能与易用性评测https://guijiagi.com/posts/llm-distillation-2026-gpt5-to-7b-capability-transfer/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/llm-distillation-2026-gpt5-to-7b-capability-transfer/系统介绍大模型知识蒸馏的最新技术，涵盖响应蒸馏、特征蒸馏、Agent 蒸馏和实际部署案例https://guijiagi.com/posts/multimodal-rag-image-text-hybrid-retrieval/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/multimodal-rag-image-text-hybrid-retrieval/深入多模态 RAG 的工程实现，涵盖图文统一编码、混合检索策略、跨模态重排序等核心技术https://guijiagi.com/posts/domain-llm-finetuning-medical-legal-finance/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/domain-llm-finetuning-medical-legal-finance/深入探讨医疗、法律、金融三大垂直领域的大模型微调实践，涵盖数据策略、训练方案和合规要求https://guijiagi.com/posts/vector-database-2026-milvus-pinecone-weaviate-qdrant/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/vector-database-2026-milvus-pinecone-weaviate-qdrant/2026 年四大主流向量数据库深度横评，涵盖性能基准、功能对比、成本分析和选型建议https://guijiagi.com/posts/graphrag-graph-retrieval/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/graphrag-graph-retrieval/GraphRAG图检索增强生成https://guijiagi.com/posts/graphrag-knowledge-graph-rag/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/graphrag-knowledge-graph-rag/GraphRAG图检索增强生成的原理、架构与实践，突破传统向量检索的局限性https://guijiagi.com/posts/lora-finetune-tuning/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/lora-finetune-tuning/LoRA微调参数调优指南https://guijiagi.com/posts/lora-finetuning-tuning-guide/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/lora-finetuning-tuning-guide/LoRA微调的核心参数调优指南，从rank选择到学习率设置的全面实践https://guijiagi.com/posts/qlora-finetune-guide/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/qlora-finetune-guide/QLoRA量化微调指南https://guijiagi.com/posts/rag-chunking-strategy-evaluation/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/rag-chunking-strategy-evaluation/固定分块、语义分块、递归分块等RAG分块策略的深度对比与评估https://guijiagi.com/posts/rag-chunking-strategy/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/rag-chunking-strategy/RAG分块策略对比评估https://guijiagi.com/posts/rag-hybrid-search/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/rag-hybrid-search/RAG混合检索方案设计https://guijiagi.com/posts/rag-eval-optimization/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/rag-eval-optimization/RAG效果评估与优化闭环https://guijiagi.com/posts/rag-rerank-strategy/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/rag-rerank-strategy/RAG系统中重排序策略的原理、模型选择与工程实践，提升检索精度的关键环节https://guijiagi.com/posts/continual-learning/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/continual-learning/大模型持续学习的挑战与方案，在学新不忘旧的道路上探索实践路径https://guijiagi.com/posts/incremental-training-practice/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/incremental-training-practice/大模型增量训练实践https://guijiagi.com/posts/llm-incremental-training/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/llm-incremental-training/大模型增量训练的工程实践，从数据准备到训练优化的全链路指南https://guijiagi.com/posts/multimodal-rag-practice/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/multimodal-rag-practice/多模态RAG架构实践https://guijiagi.com/posts/domain-adaptation-finetune/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/domain-adaptation-finetune/领域适配微调方案https://guijiagi.com/posts/domain-adaptation-finetuning/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/domain-adaptation-finetuning/从通用大模型到领域专家的微调方案，数据构建、训练策略与效果评估全流程https://guijiagi.com/posts/embedding-finetune-practice/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/embedding-finetune-practice/嵌入模型微调实战https://guijiagi.com/posts/embedding-model-finetuning/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/embedding-model-finetuning/嵌入模型领域微调的完整流程，从数据准备到训练评估的实战指南https://guijiagi.com/posts/agentic-rag-architecture/Fri, 26 Jun 2026 14:00:00 +0800https://guijiagi.com/posts/agentic-rag-architecture/深入探讨 Agentic RAG 架构如何将传统 RAG 与智能体能力融合，实现自适应检索、多步推理与动态知识更新https://guijiagi.com/posts/rag-vs-fine-tuning-2026/Fri, 26 Jun 2026 14:00:00 +0800https://guijiagi.com/posts/rag-vs-fine-tuning-2026/RAG 和微调不是非此即彼的选择，而是互补的知识注入策略。本文从 2026 年的实践视角，给出全场景决策框架。https://guijiagi.com/posts/graph-rag-explained/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/graph-rag-explained/深入解析 GraphRAG 的完整架构：实体抽取、关系构建、社区检测、层级摘要与多跳检索，对比传统 RAG 的优劣，附 Neo4j + LLM 的完整代码实现。https://guijiagi.com/posts/llm-knowledge-distillation/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/llm-knowledge-distillation/深入讲解 LLM 知识蒸馏的完整体系：黑盒蒸馏、白盒蒸馏、在线/离线蒸馏，涵盖 Logit 蒸馏、中间层蒸馏、指令蒸馏、响应蒸馏等方法，附 PyTorch 代码实现与效果对比。https://guijiagi.com/posts/lora-vs-dora-vs-qlora/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/lora-vs-dora-vs-qlora/深入对比 LoRA、DoRA、QLoRA 三种参数高效微调方法的原理、数学公式、代码实现与效果差异，给出不同场景下的最佳选择建议和完整的实验数据对比。https://guijiagi.com/posts/fine-tuning-qlora-guide/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/fine-tuning-qlora-guide/深入解析 QLoRA 的 NF4 量化、双重量化、分页优化器原理，提供完整的 LLaMA-3 QLoRA 微调代码实战，涵盖数据准备、训练配置、显存优化技巧与常见问题排查。https://guijiagi.com/posts/rag-reranking-guide/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/rag-reranking-guide/深入对比三种主流 RAG 重排序方案：Cohere Rerank API、BAAI bge-reranker、自训练 Cross-Encoder，涵盖原理、部署方式、代码实现、延迟测试与效果对比，给出不同场景的最佳选择。https://guijiagi.com/posts/continuous-pretraining/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/continuous-pretraining/系统讲解大模型持续预训练（CPT）的完整方法论：领域数据构建、训练策略、灾难性遗忘缓解、学习率调度、数据混合比例、评估体系，附完整训练代码与企业级实践指南。https://guijiagi.com/posts/advanced-rag-patterns/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/advanced-rag-patterns/深入解析四种高级 RAG 模式：HyDE 假设文档嵌入、CRAG 纠正性检索增强生成、Self-RAG 自反思检索增强生成、FLARE 主动检索增强生成，附完整代码实现与性能对比。https://guijiagi.com/posts/dpo-vs-rlhf/Thu, 25 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/dpo-vs-rlhf/深入对比 RLHF 三阶段流程与 DPO 直接偏好优化的原理、优劣势和工程实践https://guijiagi.com/posts/graphrag-explained/Thu, 25 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/graphrag-explained/深入解析微软 GraphRAG 的原理、架构和实现，对比 Vector RAG 的差异和适用场景https://guijiagi.com/posts/lora-qlora-finetune-guide/Thu, 25 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/lora-qlora-finetune-guide/从原理到实战，全面解析 LoRA 和 QLoRA 参数高效微调技术，包含 Rank 选择、模块配置和 Unsloth 加速方案https://guijiagi.com/posts/rag-vs-finetune-decision/Thu, 25 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/rag-vs-finetune-decision/深入对比 RAG 与微调的技术特征、成本结构和适用场景，提供决策矩阵和混合方案建议https://guijiagi.com/posts/rag-chunking-strategy-deep/Thu, 25 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/rag-chunking-strategy-deep/系统对比四种 RAG 分块策略的原理、实现和适用场景，包含代码示例和性能基准https://guijiagi.com/posts/rag-vs-long-context/Wed, 24 Jun 2026 16:00:00 +0800https://guijiagi.com/posts/rag-vs-long-context/系统对比 RAG 与长上下文模型方案，提供决策矩阵与混合架构建议https://guijiagi.com/posts/instruction-tuning-guide/Wed, 24 Jun 2026 16:00:00 +0800https://guijiagi.com/posts/instruction-tuning-guide/系统讲解 LLM 指令微调全流程，从 SFT 数据构造到 DPO 直接偏好优化与 RLHF 对比https://guijiagi.com/posts/lora-qlora-practice/Wed, 24 Jun 2026 14:00:00 +0800https://guijiagi.com/posts/lora-qlora-practice/深入讲解 LoRA 低秩分解原理、QLoRA 量化、PEFT 库使用及从数据准备到部署的完整实践https://guijiagi.com/posts/rag-pipeline-optimization/Wed, 24 Jun 2026 14:00:00 +0800https://guijiagi.com/posts/rag-pipeline-optimization/系统性讲解 RAG 流水线每个环节的优化策略：分块、召回、重排、查询改写、上下文压缩、缓存https://guijiagi.com/posts/rag-evaluation-guide/Wed, 24 Jun 2026 14:00:00 +0800https://guijiagi.com/posts/rag-evaluation-guide/系统介绍 RAG 评估维度、RAGAS 框架、核心指标及自动化评估工具实践https://guijiagi.com/posts/fine-tuning-vs-prompt/Wed, 24 Jun 2026 14:00:00 +0800https://guijiagi.com/posts/fine-tuning-vs-prompt/系统对比 Prompt 工程与微调的适用场景、成本、效果，提供决策矩阵和混合方案https://guijiagi.com/posts/rag-practical-guide/Mon, 08 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/rag-practical-guide/RAG 是 Agent 最实用的知识增强方案。本文从架构设计到代码实现，完整拆解生产级 RAG 系统的每个环节。https://guijiagi.com/posts/rag-vs-finetuning/Sat, 06 Jun 2026 16:00:00 +0800https://guijiagi.com/posts/rag-vs-finetuning/RAG 和 Fine-tuning 不是竞争关系，而是互补关系。本文用决策树帮你选对方案。https://guijiagi.com/posts/vector-db-comparison/Wed, 03 Jun 2026 14:00:00 +0800https://guijiagi.com/posts/vector-db-comparison/RAG 系统的心脏是向量数据库。本文实测 5 大向量数据库，帮你选出最适合的那个。