What is the relationship between RAG, LangChain, and Agent?

Then summarize this picture in one sentence:

      You can use RAG technology in the LangChain framework to create an Agent that plays a specific role to specifically address the specific needs of users.

      Next, we will introduce these three nouns in order.

RAG (Retrieval Augmented Generation): Knowledge Augmentation Plugin

      RAG (Retrieval-Augmented Generation) is a technology that enables large language models to read external information when generating answers. This technology can be simply understood as retrieving relevant information from an external database as a reference before generating content.

      For example, if you complete a full-time high school course, we can say that you have completed pre-training and have the basic capabilities of a general large model. Then you apply for a computer science major and complete four years of professional courses. We can say that you have completed a round of fine-tuning , and then you are a qualified large model that can solve various basic computer-related downstream tasks .

      But after all, the four-year undergraduate study cycle is a bit long. Let's think of a quick way to learn it quickly, which is to use the knowledge you have mastered in high school, plus a bunch of computer science books around you. Then, with the buff + blessing of open-book exams, you can barely become a low-profile model for answering basic computer-related questions. The process of receiving questions, finding answers in professional books based on your high school knowledge, and organizing and feeding back to the questioner is Retrieval Enhanced Generation (RAG) .

      Since the cost of training and fine-tuning is still high and ultra-long context is still in the theoretical stage, plug-in technologies such as RAG currently have visible application prospects in the fields of news, technology, medical care, and internal think tanks of companies with teams of a certain size, where knowledge is frequently updated .

LangChain: Create AI applications quickly and easily

      With RAG technology, a new problem arises. How to integrate functions such as retrieval, generation, data storage, and API calls into an overall application system? At this time, LangChain comes in handy.

      Langchain is an open source framework for developing LLM applications, designed to help developers build applications driven by large language models more easily. RAG is a very important application field of large language models, and LangChain naturally has relatively sufficient support. LangChain can also help developers flexibly design multi-step workflows to make RAG results more controllable.

      Let's take a simple application example. If we want to build a legal consulting application, it needs to complete the following steps:

1. Receive specific cases from users;

2. Retrieve relevant legal terms from the vector database;

3. Go back and generate answers using a large language model;

4. Remind users of the compliance of their answers before generating answers.

      This specific RAG application example can be quickly implemented through LangChain. The modular components and interfaces provided by LangChain are convenient for developers to freely combine according to specific needs. For non-experienced developers, the LangChain framework greatly simplifies the development difficulty, allowing everyone to easily implement "building blocks" development.

Agent : The “agent” that AI uses to perform tasks

      After understanding RAG and LangChain, let’s take a look at the concept of AlAgent.

      Agent  is an intelligent system that can autonomously perceive the environment, make decisions and execute actions . As its English name suggests, Agent is equivalent to a task agent, which can "understand" the processes and rules set by the user and handle a series of tasks according to its own ideas.

      As shown in the figure, Agent is usually based on Large Language Model (LLM) , and the creator uses prompt template to specify its role and work content; Agent has "memory" , which allows it to remember not only the context of the conversation, but also the user's preferences and personalized requirements to better meet the needs of the user; Agent has "behavioral autonomy" (Action) , after receiving instructions, it can use the large language model to determine whether it needs to use the corresponding tools to complete the task independently.

      Because Agent has these characteristics, it is more suitable for applications such as automated tasks, digital assistants, and game characters .

Using RAG technology to create a dedicated Agent in the LangChain framework

      Let's go back to the picture at the beginning. The responsibilities and execution processes of these three terms are as follows:

      LangChain can provide a sufficiently complex workflow structure for tasks, and the Agent is responsible for executing each task link in these processes according to the settings of the PromptTemplate.

      The LangChain framework also provides various corresponding libraries to support RAG technology, allowing RAG technology to be used as a tool for Agents to acquire knowledge from the Knowledge Base.

      After the Agent obtains the corresponding knowledge, it is organized and understood by the LLM, and a judgment is made to return useful information to the client or to perform a specific operation, which is then completed by the Agent.