Are AI Agents Nothing But Multiple RAGs?
In this blog, we will explore the fundamental concept of AI Agents. Along the way, we’ll answer the intriguing question: “Are AI Agents Nothing But Multiple RAGs?”
Understanding LLMs
To start, let’s establish two key characteristics of Large Language Models (LLMs):
Stateless Nature: LLMs do not store information. They simply process inputs through a series of computations (essentially multiplications and additions).
Linguistic Capabilities: LLMs excel at understanding and generating language, hence the "Language" in their name. For example, they perform well in tasks akin to reading comprehension exams, where a paragraph is given, and questions are answered based on it.
However, note that there are other types of models like Large Quantitative Models (LQMs) and Large Concept Models (LCMs):
The Power of RAG
The value of LLMs became evident to us when Retrieval-Augmented Generation (RAG) emerged. Here’s one of our blogs that explains RAG in detail.
RAG addresses a significant limitation of LLMs: their lack of up-to-date and domain-specific knowledge. By augmenting stateless LLMs with external knowledge sources or data, RAG empowers them to deliver accurate and context-relevant responses. In essence, RAG combines a stateless component (LLM) with knowledge/data tailored to a specific use case or context.
AI Agents vs. RAG
When AI Agents gained popularity, we wondered what they introduced beyond RAG. To define AI Agents simply: they are systems capable of decision-making and action-taking using various tools. Let’s break this down further:
Action-Taking: LLMs don’t inherently take actions. However, they can generate language-based responses that specify actions to perform (e.g., API calls or function executions). The LLM consumer is responsible for executing these actions and returning the results to the LLM. In this context, the external data includes API schemas, function specifications, or tool schemas.
In the first image below, no tools are being utilized, and the input tokens stand at 15. In the second image, we observe an increase in the input tokens when the tool is activated. This indicates that the LLM is augmented with the tool schema in its input.
Decision-Making: Humans rely on reasoning, which involves analyzing past or present information, to make decisions. Similarly, AI Agents require reasoning capabilities. One popular framework enabling this is ReAct.
In ReAct, the LLM first generates a “thought” (a textual description) suggesting an action.
The action is executed by the LLM consumer, and the response, along with the previous thought, is fed back to the LLM.
The LLM then generates an “observation” based on the response and prior thoughts.
This cycle continues until the LLM produces a final, satisfactory response.
Here, the external data includes the action’s response and the LLM’s own thoughts and observations.
The agent utilizes the Search and Calculator tools, along with its observations, to determine the final answer, as illustrated in the image below.
Credit: Screenshot from Blog.
Conclusion: Augmented LLMs as the Core of AI Agents
From this analysis, it’s evident that external data is pivotal in making LLMs more effective. In the case of AI Agents, this external data encompasses action specifications, thoughts and observations, and action responses. While action responses may be seen as a form of retrieved data, other inputs, such as thoughts and observations, do not strictly fit into this category. Nonetheless, all these components act as augmentation mechanisms, enriching the LLM’s capabilities to reason, decide, and act. This underscores the transformative role of external data in turning a stateless LLM into a dynamic, decision-making system. Regardless of its form, external data fundamentally augments the LLM, enabling it to perform tasks far beyond its original design.
This understanding aligns with a recent article by Anthropic, which emphasizes that augmented LLMs are the foundational unit of AI Agents. Read more here.
Thus, we see that AI Agents are essentially an evolution of RAG, with added capabilities for decision-making and action-taking. At their core, they are built on the same principle of augmenting stateless LLMs with external data to unlock their full potential.
To answer the original question, “Are AI Agents Nothing But Multiple RAGs?”—the answer is both yes and no. Yes, in the sense that AI Agents rely on the concept of augmenting LLMs with external data, much like RAG. However, they go beyond simply retrieving information. AI Agents are fundamentally augmented LLMs that integrate various forms of external data. So, AI Agents Are Nothing But Multiple Augmented Generations.
At Innowhyte, we examine trends like AI Agents with careful consideration, grounding our approach in thoughtful analysis rather than hype. Instead of being swept up by the hype, we focus on reasoning from first principles to truly understand their implications. This deliberate approach ensures that we create deterministic and reliable AI systems, built on robust foundations to deliver consistent and meaningful outcomes.