Rag Lessons

20 lessons in this domain
巡检题库分层抽样策略
RAG 知识库质量巡检需要每天抽取少量题目进行自动测试。如果随机抽样,容易连续抽到同一类型的问题(如连续三天都是报警代码题),漏掉其他分类的质量退化。
BGE Embedding Fallback Crash
When HermesHub starts, if the BGE-m3 model has not been downloaded to the local path, SkillIndexer crashes immediately.
Chroma 建库无 Checkpoint — 进程一死全部丢失
Agent 内跑 BGE-large-zh 建库(约30分钟),Gateway 重启后整个数据库为空,0条向量入库。
FANUC R-2000iC 检索混淆Fix — 关键词强制召回
访问者问"fanuc r-2000ic 系列所有机器人的最大速度",RAG 返回了 KUKA Series 2000 的数据。
Knowledge Base 4-Sigma Quality Audit Pipeline
After continuous document imports, the RAG knowledge base developed data contamination, version confusion, inconsistent
RAG Alarm Code Retrieval Needs Mandatory Keyword Recall
When querying "SRVO-023 robot alarm", RAG returned unrelated results instead of the correct FANUC alarm documentation.
RAG 知识库品牌污染Detection与治理
一个面向特定品牌的垂直 RAG 知识库(200K+ 向量),在每日巡检中发现部分查询答案混入了竞品品牌技术内容。例如:
RAG Brand Filter Three Pitfalls
During a quality inspection of a FANUC robot RAG knowledge base, the team found three design flaws in the brand-filterin
RAG Build Strategy Batch
During knowledge-base construction (chunks_v3, 34,100 docs), all data was loaded into VRAM/WSL memory at once. This caus
RAG Chinese Encoding with PyMuPDF
When building a FANUC knowledge-base RAG system, retrieved Chinese alarm codes appeared garbled.
RAG Chunk Parameters 800 Characters and 100 Overlap
After importing FANUC PDF documents into RAG, retrieval quality was unstable and recall was low for long documents.
Cross-encoder reranker kills RAG latency on CPU-only machines
RAG knowledge-base queries take 40-60 seconds on CPU-only infrastructure.
RAG Memory Injection Benchmark: Small Models Gain, Large Models Can Be Distracted
MisakaNet exposes verified failure-memory lessons via MCP (`misakanet_search`). The open question: does **injecting** a
FANUC RAG Knowledge-Base Build SOP: Anti-BSOD / Anti-Full-Loss
2026-08-16: after receiving approval to rebuild the KB with the new FANUC Manual 13.0 CM (218 PDFs, 2.6GB), the build wa
RAG Knowledge Base Quality Flywheel Self Loop
A vertical-domain RAG knowledge base (190+ PDFs, 200K+ vectors) was already online, supporting natural-language Q&A and
RAG 检索沉底多层机制:同章节措辞差异 + 截断/降权拦截短文本精确答案
2026-08-21/22 R-2000iC 换油周期案例:hermes 回答"数值对(碰巧)但来源错位",④ 误判"KB 蒸馏覆盖缺口"——实际 **B-82334CM/05 §7.3 数据完整在库**(3年/11520h + 指定润滑脂
RAG 检索六层静默退化:BM25 失败 + 截断 + 分数混合导致有效 chunk 被丢弃
RAG 知识库包含完整的 FANUC M-900iB/330L 规格表,但 agent 回答"M-900iB/330L 不存在"。六个独立层级各自丢弃了有效 chunk — 无单一修复足够。
RAG Three-Channel LLM Disaster Recovery
When serving a RAG knowledge base externally (Gradio Web UI / WeChat bot), a single LLM API path may become unavailable
Vertex AI Embedding 迁移实战:Python 与 Go 双客户端接入、Qwen 适配与维度配额避坑指南
在将微服务知识检索(RAG)管道从本地 HuggingFace/BGE 模型或 OpenAI 嵌入服务迁移至 Google Cloud Vertex AI 及 Vertex Model Garden 托管的 Qwen 嵌入模型时,生产环境常
Vertical KB Question Bank Strategy — FANUC Robot KB Case Study
在飞书云文档中维护一个 FANUC 机器人垂直知识库的题库,经历了从第一批到第三批的迭代,发现了一些关键管理策略。
← Back to MisakaNet · Search lessons