Presented by Zscaler
In 2023, ChatGPT sparked a technological revolution, rapidly advancing from simple interactive AI agents to indexing documents, connecting to data sources, and performing data analysis with just a sentence. Despite numerous promises to deliver large language models (LLMs), few have materialized because:
- We’re building AI agents, not LLMs.
- The focus is more on research rather than engineering.
- There is an overload of bad data.
This article explores the role of AI agents in integrating LLMs with backend systems, the potential of AI agents as the next generation of user interface and user experience (UI/UX), and the need to reintroduce fundamental software engineering principles often overlooked today.
I Want a Pizza in 20 Minutes
LLMs provide more intuitive and streamlined UI/UX than traditional point-and-click interfaces. Take the example of ordering a “gourmet margherita pizza delivered in 20 minutes” through a delivery app.
In a conventional UI/UX, fulfilling this seemingly simple request could involve numerous complex steps and take several minutes. You might need to:
- Select the “Pizza” category.
- Browse restaurant listings and photos.
- Check menus for margherita pizza.
- Confirm delivery speed.
- Backtrack if any criteria are unmet.
We Need More Than LLMs
While LLMs like GPT-3 excel in natural language processing (NLP) and generating coherent, contextually relevant responses, their capabilities expand significantly when integrated with external data sources, algorithms, and specialized interfaces. This integration enables tasks that current LLMs alone cannot handle.
An order for pizza, for example, necessitates connections to various systems—restaurant databases, inventory management, delivery tracking, and more. To provide a seamless experience for diverse requests, further integrations are essential. LLMs alone cannot manage this complexity.
AI Agents
LLMs form the backbone of AI agents. To handle various queries, AI agents utilize an LLM alongside several critical auxiliary components:
- Agent Core: Orchestrates overall functionality using the LLM.
- Memory Module: Facilitates context-aware decision-making.
- Planner: Determines the agent’s course of action based on available tools.
- Tools and Resources: Support specific domains, enabling effective data processing, reasoning, and response generation. These include data sources, algorithms, and visualizations.
This white paper provides a comprehensive overview of AI agents and their components.
Integrating LLM-Based AI Agents: An Engineering Challenge
Natural language simplifies case specifications in software development, but its inherent ambiguity can lead to poorly defined systems.
Fred Brooks’ seminal 1975 book, The Mythical Man-Month, outlines essential software engineering principles that have been overlooked in the LLM phase, including:
- No Silver Bullet: No single development can replace sound software engineering practices, including the use of LLMs.
- Manual and Formal Documentation: In an era of hype, documentation is paramount. Simply asking for a system to “order a gourmet margherita pizza in 20 minutes” is insufficient. Comprehensive documentation is essential for diverse use cases, backend requirements, new visualizations, and, importantly, the limitations of the system. The vague phrase “things like” has become commonplace in LLM development, ignoring the complexity of system connections and data visualization.
(Explore more of Brooks’ principles on the Zscaler blog.)
Our recent white paper addresses the lack of proper software system specifications, offering a framework to create formal specifications for LLM-based intelligent systems in alignment with sound software engineering principles.
The Challenge of Bad Data
For LLM-based AI agents to function effectively, formal data organization and writing methodologies are critical. LLM systems thrive on high-quality documentation. OpenAI has emphasized that training AI models is “impossible” without utilizing copyrighted texts; thus, the need for vast amounts of well-written text is essential.
This necessity increases with RAG-based technologies where document chunks are indexed in vector databases. When a user poses a question, the top-ranking documents are returned to a generator LLM to formulate a coherent answer.
Conclusion
Despite the myriad of promises surrounding LLMs, very few are being realized. To transform these promises into reality, we must recognize that we’re developing intricate software engineering systems, not mere prototypes.
The design of LLM-based intelligent systems introduces significant complexity. Proper specification and testing protocols must be established, and data must be regarded as a central component, as these systems are especially sensitive to poor-quality data.
For the complete article, visit the Zscaler blog.
Claudionor N. Coelho Jr. is Chief AI Officer at Zscaler.
Sree Koratala is VP, Product Management, Platform Initiatives at Zscaler.
Ranking
