RAG knowledge base, AI agents, and remote large models

Recently, a friend told me that the website had too much content and was not user-friendly, so he asked me to consider creating a RAG knowledge base for the website. After that, I used Chatbase to create a RAG knowledge base.

Why is there such a demand? Many people want websites to offer "intelligent customer service" that can answer FAQs and even understand on-site articles.

Training a large model is prohibitively expensive, leading to the emergence of a "remote large model + knowledge base + packaging platform" combination.

I import my website link into Chatbase, which automatically builds a knowledge base for me. It then uses a remote large model (such as OpenAI's 4o Mini) to answer questions.

Concept Basics

** 1. Remote Large Model **: This refers to large language models (LLMs) running in the cloud, such as OpenAI GPT, Anthropic Claude, and Google Gemini.

Remote means the model isn't run locally, but rather accessed via an API. The large model is the "core brain," responsible for understanding and generating answers.

** 2. RAG Knowledge Base **: RAG stands for Retrieval-Augmented Generation.

It's not the model itself, but rather a methodology for using large models.

The idea is:

1) Website articles, FAQs, and PDFs are broken down into small segments → stored in a vector database.

2) Users ask questions → the system retrieves relevant content.

3) The retrieved content plus the user's question are fed into the remote master model.

4) The master model generates answers that are more accurate to the data.

Thus, the RAG knowledge base acts as a "plug-in memory" for the remote master model.

*3. AI Agent * : A "secretary" (execution layer) built on top of the big model and knowledge base, capable of making decisions and invoking tools.

It not only communicates with the system but also invokes tools, performs tasks, and makes decisions.

For example, to help optimize Swift code in Xcode, the AI agent:

1) Accesses the Swift code in Xcode through an API.

2) Accesses the RAG knowledge base → retrieves relevant development documentation.

3) Allows the big model to generate optimization recommendations based on the knowledge base content.

4) Calls a local API to write the optimized code back to Xcode.

Product Introduction

** Official Customization **

OpenAI GPTs: Upload documents and add model specifications to create a custom bot. Officially supported, but only usable within ChatGPT.

** RAG Knowledge Base SaaS Platforms **

Chatbase, Dify, and Claude Workflows: Upload documents/websites, and these platforms automatically build a RAG knowledge base, then pass user questions to a larger model.

No code required, quick to get started, and suitable for website FAQs and customer service scenarios. The downside is platform dependency.

** Open Source/Developer Tools **

Flowise: Visual drag-and-drop, self-deployment.

LangChain/LlamaIndex: Developer-oriented, using Python/JS to build code. High degree of flexibility, suitable for engineers, but requires programming skills.

** Chatbot Platforms **

BotPress and Typebot: These platforms focus on process design. Instead of starting with a large model, they use a "conversation flow designer" and integrate with Live Chat Manager. They're suitable for multi-step Q&A/guidance (for example: first ask the user's name → then ask for their needs → provide recommendations).