Harness LLMs with GenAI agents for effective enterprise use

AUTOMATE KNOWLEDGE PROCESSING

The ability to automate knowledge processing quickly and at scale has arrived.

GenAI, an advanced form of machine learning powered by LLMs, has ushered in a bold new era in computing. The potential of GenAI has been demonstrated by the technology’s ability to generate new and customized content quickly and at scale. Examples include the generation of code, text, images, lyrics, musical compositions, and even videos. The benefit of GenAI is that it’s all done without any human input. Not only can this technology save valuable time and effort, but it frees up human resources to focus on more complicated problems like achieving goals and driving outcomes. 

A type of LLM, generative pre-trained transformers (GPTs) are designed for natural language processing. OpenAI’s GPT-3.5 and the more powerful GPT-4, as well as Google’s PaLM, are examples of these models, which have been pre-trained on large datasets. These LLMs generate responses that are often difficult to distinguish from those written by humans. Since models are pre-trained, they allow enterprises to quickly build new capabilities. 

Optimizing the benefit of LLMs, however, requires the fine-tuning of models to make them more relevant to an enterprise’s domain, along with setting guardrails to make LLMs safe for use by employees. So how can this best be accomplished?

By leveraging a framework that optimizes the use of LLMs with autonomous AI agents, businesses can unleash the possibilities of GenAI while helping to ensure responsible usage across an enterprise. The implications are significant for a wide range of industries—this approach supports activities from technical design and marketing to product discovery and engineering, customer service, and data analysis.

Making models relevant

LLMs represent a significant leap in AI evolution, demonstrating an ability to generate knowledge in the form of content that is not only coherent but also creative and engaging, whether the results are technical documents, engineering designs, or even new product ideas.

But because they are generic pre-trained models, domain and sub-domain adaptations of LLMs are a prerequisite in making them relevant for enterprise usage. A calibrated approach for fine-tuning and better prompt engineering is needed to maintain an LLM’s relevance. 

Enabling a calibrated training strategy

For domain and sub-domain adaptations, LLMs must be either fine-tuned or pre-trained from scratch, depending on use cases and user scenarios. In the case of fine-tuning, enterprises can embrace four different training approaches:

• Combined fine-tuning: An enterprise prepares relevant data, driven by use cases, in a specific format as required by the selected LLM. Because this approach lumps domain, sub-domain, and transactional data into one corpus, model learning is largely restricted to use case scenarios, as opposed to staggered learning, where the model becomes domain-aware before becoming enterprise-aware.

• Phased: Addressing the domain learning limitations of combined fine-tuning, this approach imposes a fine-tuning sequence. The focus is on building domain awareness in a model before going deeper into sub-domains and context.  

• Domain and task-adapted: This approach is critical for use cases and user scenarios where a model’s performance and differentiated outcomes need to be more responsive than a traditional pre-trained or fine-tuned model. It entails expanding the pre-training envelope to include a domain-specific and task-specific data corpus, in addition to a generic open data corpus, for pre-training. 

Fine-tuning to optimize enterprise outcomes

A good training regimen does not guarantee enterprise outcomes. Post-fine-tuning optimization is essential to align an LLM to enterprise needs and drive desired outcomes.  

To address the structural limitations created by involuntary bias injection, which can plague an LLM’s response, newer optimization strategies, from reinforcement learning with human feedback to reward-based models and low-rank adaptations, are critical. These strategies can align an LLM to enterprise demands of "mostly consistent" truthfulness, reliability of responses, and repeatability under the same context and user scenarios. 

Eliminate variability by design

An LLM is only as good as the prompts used to elicit a response. Enterprises must embark on large-scale business change programs to train organization members on prompting. Because prompting is part art and part science. However, even the best training can’t eliminate variability. It also depends on how well employees are deploying prompting techniques to realize desired outcomes. 

To remove variability in responses, enterprises can build a library of prompts and chained prompts with defined purposes, which can be selected and deployed by employees with minor enhancements. This approach not only eliminates variability but also helps standardize user scenarios. By leveraging an agent-based modeling framework, this approach can be extended, eliminating the need for human prompting altogether. Such an approach can help automate prompting, allowing enterprises to govern and constrain (if required) end-user interactions, with codified enterprise policies and guardrails.
 

There are multiple training strategies to get the best out of LLMs.

Setting up guardrails for governance

How can we leverage the promise of AI in a safe manner? 

While the generative capabilities of LLMs offer immense potential, they also present organizations with a new set of challenges that must be addressed. The open-ended nature of these models means they can produce outputs that may not align with an organization’s requirements, policies, or ethical guidelines. Without a set of safety controls or guardrails in place for governance, there is a real risk of generating inappropriate, biased, or misleading content that can harm an organization’s reputation, or violate regulatory and legal obligations.

Putting up guardrails is essential to ensure the responsible and secure usage of generative LLMs across the enterprise by defining the boundaries within which they may operate. Organizations can leverage the creative potential of GenAI technology while maintaining control over outcomes—you don’t have to enable one aspect at the expense of others. The guardrails approach for LLMs arises from three key areas of business need to:

• Enforce policies for their use.

• Enhance their contextual understanding.

• Support their ability to continuously adapt. 

By implementing guardrails, enterprises can harness the power of LLMs while safeguarding against security breaches, compliance violations, and ethical concerns. These guardrails serve as guidelines to shape the behavior of the models, allowing organizations to build generative capabilities on a foundation of trust. Organizations are thereby empowered to channel their new capabilities in a controlled and responsible manner.

Various routes to create GenAI guardrails

While the need for guardrails is readily apparent, the route an enterprise should take to implement them is not so clear. There are broadly four possible routes an enterprise can take to create foolproof guardrails around GenAI: 

• Building enterprise specific LLMs on trusted information sources: Taking this route is easier said than done. Pre-training LLMs from scratch and subsequently fine-tuning them, although effective in preventing obvious risks around bias and toxicity, requires massive investments in computing infrastructure. Even then, outcomes are not likely to be completely free of bias and toxicity. The risks around security, data protection, and variability in model responses remain unaddressed—these risks are inherent in LLMs. 

• Optimizing LLMs for enterprise policies and technology guardrails: This route entails augmenting LLMs with advanced optimization techniques, such as reward-based models, reinforcement learning with human feedback, or low-rank adaptations. These techniques are not just governed by generic rules of bias elimination, toxicity detection, and improving hallucinations like the ones prevalent in the latest GPT or PaLM models. In addition, enterprises can leverage these optimization techniques on tailor-made rules and policies that are unique to an industry and enterprise context. Although promising, the closed nature of these techniques makes them hard to implement on commonly used LLMs.

• Red teaming, using a manual, human-driven verification, and assurance capability: With this route, the model is verified and checked for all possible vulnerabilities and risks. This is a slow and very expensive process that is only useful for identifying vulnerabilities in an LLM. Fixing those vulnerabilities would entail subsequent fine-tuning and optimizations, if possible.

• Augmenting LLMs with agent-based models as verifier and governor: This route recognizes the complexity of an enterprise while focusing on creating an automated LLM verifier and governor. Enterprises today have varying degrees of complexity, and their resultant policies, including risk mitigations and optimizations, will invariably become less effective over time. Instead of focusing on making LLMs safe for an enterprise, agent-based modeling can be used to verify and govern all LLM interactions. Even if the underlying LLMs are not safe, the use of agent-based models can help an enterprise ensure its interactions are safe as well as outcome-oriented. 

Among these potential routes for making GenAI and LLMs safe for enterprise use, the best-suited and more viable option is agent-based modeling. By augmenting LLMs with agent-based models, enterprises can enforce technology and security guardrails for all GenAI interactions.

Letting AI agents do the work

Automation meets creation.

Combining LLMs with, for instance, an agent-based modeling framework opens new possibilities for the development of an autonomous AI agent ecosystem. Intelligent agents, powered by LLMs, can work collaboratively, adaptively, and autonomously in this ecosystem to realize user goals and outcomes for content creation without a heavy reliance on explicit prompt design and engineering by users.

A shift in focus

Traditionally, users have been responsible for designing prompts and providing explicit instructions to LLMs to obtain the desired outputs. This process requires domain expertise and can be time-consuming, limiting a user’s ability to explore the full potential of a model. With the introduction of agent-based models, the level of required expertise is significantly reduced.

By incorporating AI agents into a framework, users can shift their focus from prompt design and engineering to defining goals and outcomes. The agents, equipped with the ability to understand user intents and interactions with LLMs, can take charge of formulating appropriate queries, seeking clarifications, and generating outputs aligned with user objectives. This outcome-driven approach enables users to engage more intuitively with the AI system, making it accessible to a broader range of users, including those without deep technical expertise.

Collaboration and adaptation across an ecosystem

Within an autonomous AI ecosystem, agents can collaborate and share knowledge to enhance their collective capabilities. Each agent can specialize in a specific domain or task, leveraging the expertise of LLMs in that area. By sharing insights, learned behaviors, and successful strategies, agents can collectively improve their performance and achieve better outcomes.

These agent-based models also have the capability to adapt and learn from user interactions and feedback. As users interact with the system and provide guidance, agents can refine their understanding of user preferences, contextual nuances, and desired outcomes. This adaptive behavior enables the agents to deliver more personalized and accurate responses attuned to individual user needs over time.

Role of agent-based models

An agent-based modeling framework introduces a new approach to enhance the usage of LLMs by combining them with an agent-based layer. Computational in nature, agent-based models simulate the actions and interactions of autonomous agents to analyze complex systems. 

With LLMs, an agent-based layer acts as a mediator between the user and the language model, facilitating controlled and guided interactions. Agent-based models bring several advantages:

• Policy enforcement: Enables the enforcement of policies and guardrails by defining rules, constraints, and guidelines for interactions with the language model. These policies can encompass security measures, compliance requirements, ethical guidelines, and other enterprise-specific regulations.

• Contextual understanding: Enhances the understanding of user queries and requests. By leveraging external data sources, user preferences, historical interactions, and other relevant information, agent-based models can provide a more accurate and context-aware input to the underlying language model, resulting in more relevant and reliable responses.

• Adaptive behavior: Learns and adapts over time. Agent-based models can incorporate machine learning techniques to analyze user feedback, improve their understanding of user intents, and dynamically adjust the policies and guardrails based on evolving enterprise needs and changing user requirements.

Long-term memory management

Despite their impressive capabilities, LLMs typically possess a short-term memory, restricted to the context window of a few thousand words. They excel at processing and generating content based on recent context but may struggle to retain information over extended periods. This limitation poses challenges in maintaining coherent and consistent interactions with users, especially when dealing with multi-turn conversations or complex tasks requiring a broader context.

An agent-based modeling framework addresses the short-term memory limitation of LLMs beyond the context window of a few thousand words by incorporating long-term memory management within the agent layer. Agents, acting as intermediaries between users and language models, can effectively maintain and recall contextual knowledge and user preferences across multiple interactions.

• Context preservation: Agents store and preserve relevant context from previous interactions, allowing for continuity in conversations and helping to ensure a more coherent understanding of user intents and requirements. This leads to more personalized and contextually relevant outputs.

• Learning from historical data: Agents enhance their understanding and decision-making capabilities by leveraging historical data. Through analyzing past interactions and outcomes, agents can identify patterns, learn from successful strategies, and adapt their behavior to better serve user goals. 

• Knowledge integration: The agent layer acts as a knowledge repository, integrating information from various sources such as external data, domain-specific knowledge bases, and user-provided context. This aggregated knowledge becomes a valuable resource that agents can tap into to generate more comprehensive and accurate responses, even when the language model’s short-term memory alone might not suffice.

Ring-fencing the language model

In TCS’ approach, an agent-based layer acts as a ring-fence, separating the language model from direct interactions. Ring-fencing serves multiple purposes:

• Controlled interactions: By placing the language model behind an agent-based layer, organizations can ensure all their interactions with the model are channeled through defined policies and guardrails. This allows for fine-grained control over outputs, reducing the risk of generating content that does not comply with enterprise-specific guidelines.

• Security and privacy: Ring-fencing provides an additional layer of security, protecting an LLM from potential breaches and unauthorized access. It also helps preserve user privacy by preventing direct exposure of sensitive data to a language model.

• Flexible integration: The ring-fence approach allows for flexibility in integrating different LLMs within an agent-based framework. Organizations can choose the most suitable language models for their specific needs while maintaining a consistent interface and policy enforcement mechanism through the agent-based layer.

Policies and controls

An agent-based modeling framework provides the means for defining policies and enforcing controls to govern the behavior and outputs of underlying language models. These guardrails can address the specific requirements and constraints of an organization through:

• Policy customization: Organizations can define policies that align with their security, compliance, ethical, and operational requirements. These policies can address aspects such as content filtering, language constraints, sensitivity to specific topics, adherence to regulatory guidelines, or any other criteria deemed important by the organization.

• Compliance enforcement: Organizations can define policies to ensure that generated outputs comply with legal, regulatory, and industry-specific guidelines. Compliance measures include guarding against the dissemination of sensitive information, preventing the generation of discriminatory or biased content, and maintaining intellectual property rights.

• Dynamic updates: Organizations can update policies and guardrails dynamically as needs evolve. Policies can be modified and updated in real time, helping to ensure continuous compliance and adaptability to changing business environments.

Key considerations for the deployment of agent-based models

While agent-based models offer a well-suited approach to making GenAI and LLMs safe for enterprise use, the systems are quite complex to design and deploy. Before embarking on the approach, enterprises need to keep these key considerations in mind:

• Design an environment closer to enterprise reality for better performance: Agents are strictly bound by the environment design in which they operate. The environment, in this case, should be very close to how an enterprise operates and the outcomes it delivers. 

• Model enterprise interactions and dynamics faithfully: Agent-based modeling needs to reflect enterprise user interactions with LLMs and the purpose those interactions are expected to serve. These dynamics are critical and need to be captured as part of guardrails for agent behavior. 

• Provide dedicated training for agents and agent-based systems: Agent-based models are effective only if they are trained for defined tasks and policies. This training is not to be confused with the pre-training and fine-tuning regimen of LLMs. 

• Optimize agents for enterprise-specific needs: Enterprises must calibrate their approaches based on industry and enterprise-specific requirements. Optimization techniques such as reward-based models or reinforcement-based learning techniques such as proximal policy optimization need to be carefully evaluated based on clear appreciation of their potential and impact on enterprise scenarios. It’s important to emphasize that not all optimization techniques will yield similar outcomes without careful evaluation and selection. 

• Maintain consistency in policy definition for predictable agent performance: Policy design needs to be consistent and manageable. Inconsistencies or vagueness in policies would force agents off the rails and result in inexplicable behavior. Policies can be atomic, conditional, or nested, so long as the definitions are clear and unambiguous. 

• Create a process to manage exceptions: Agents can and will fail. What matters more is how the failures are handled. Incorporating exception management routines for agent-based models is key in helping to ensure that agents fail gracefully—and with minimal enterprise impact. 

• Integrate and extend to new frameworks for explainability and security: An overarching integration framework is essential to bring explainability and enterprise-specific solutions into the GenAI capability set. Instead of building and rebuilding capabilities that are not central to GenAI, a technology integration backbone allows enterprises to grow the "art of the possible" with GenAI. This integration framework can be enabled without necessarily building new capabilities, for instance, around data management, explainability, and content filtering in LLMs or in the agent-based model layer. 
   

Opportunities to leverage LLMs for industry-specific applications are extensive and continue to unfold, ranging from content analysis to creation and generation to summarization, automation, and translation.

Demonstrating THE POWER OF LLMs

What can we do with large language models?

The opportunities that await adopters of GenAI technology extend to virtually all industries. The true value of leveraging generative LLMs within the enterprise lies in augmenting the abilities of people to work smarter and faster. Especially in functions where there is a significant amount of knowledge dependency, LLMs can improve overall work efficiency and human productivity. 

Knowledge dependency in a business function generally manifests as either a large number of people doing repetitive tasks or an overreliance on a few subject matter experts for critical thinking and decision-making. Enterprises have the opportunity to redefine how work gets done in this new GenAI era, automating time-consuming knowledge processing tasks and elevating the roles of their people to focus on tasks that require human cognition.

Here are just a few examples of how LLMs may be used to augment the roles of humans today to drive increased value:

• Logistics: LLMs can be given a set of criteria and then create or generate well-written, plagiarism-free content as a result. This generative capability may include codes, reports, recommendations, questions and answers, commentary, and design documents. 

• Banking: LLMs can automate many generative tasks, such as reports or dashboards. An investment research firm may use an LLM to create an automated portfolio optimization capability to improve returns. 

• Manufacturing: The amount of content abstraction and contextualization provided by LLMs to summarize tasks can be leveraged for various utilities across domains and industries. An automotive manufacturer may use an LLM to summarize engineering notes for its assembly line workers or market research reports for product improvements.  

• Capital markets: LLMs can analyze environmental parameters quickly and recommend the next-best course of action to take. Capital market firms might use an LLM to create an investment advisory bot for wealth management.

• Health: LLMs can translate text, generating human-like output in response to user input. A medical devices company might use an LLM to produce high-quality translations of spoken or written text on adverse drug events.