Building Domain Agents from RAG to Agentic Research
This tutorial presents a progression from basic RAG to domain-specific agents and deeper research workflows. It covers document ingestion, chunking, embedding, hybrid retrieval, reranking, dynamic routing, ReAct loops, query planning, evidence gathering, memory, and answer verification, with a travel-policy example throughout. The article also compares RAG with fine-tuning and frames agentic search, integrated research, and full-stack AI research as progressively higher-complexity deliverables.
The article provides a practical end-to-end blueprint for designing retrieval and agent workflows that remain traceable to source material.
从 Rag 到领域 Agent
第一天的 Agent 会查日程、订机票、调用接口。同事接着问的往往是另一类问题：「上海住一晚上限多少」「未休年假怎么折算」「这张票超标了要不要审批」。这些话的依据在制度、公告和标准表里，不在模型参数里。
做法是 RAG：先从资料里找出相关段落，再让模型只根据这些段落回答。它接在第一天的公式上。检索是一个 Tool，召回的段落是送进下一轮的 Context。模型仍然只做一次决策；哪几段文字有资格进入上下文，由外层程序决定。
下面按笔记的两条流水线往下讲：资料怎么进库，一次提问怎么变成答案，向量模型怎么选，关键词、向量和混合检索各用在什么问法上，以及检索策略可以由规则选定，也可以交给 Agent。第 6 节处理问法、空库和多跳带来的偏差。第 7 节把 RAG 分成五级。第 8 节是再往后的 Deep Research。例子沿用同一次出差。
#RAG#Agent#ReAct#BM25#OpenAI text-embedding-3-small#Google Gemini Embedding#Sentence-BERT#BERT-base#bge-reranker-v2-m3#Cohere Rerank#WebSailor#MaskSearch