In his recent presentation at the NICE Analyst Summit 2024, NICE CEO Barak Eilam made a bold prediction: “Public data harvesting is reaching its limit” for training large language models (LLMs). He argued that we’re entering an era when proprietary enterprise data will emerge as a new asset class.

Initially, I was somewhat skeptical about this. The track record for fine-tuning models with domain-specific knowledge has been mixed, with attempts often failing to outperform general-purpose models. Take, for instance, Bloomberg’s effort several years ago to create a financial language model—it was less effective than the generalized GPT-3.5, despite Bloomberg’s deep industry knowledge. For many companies, Retrieval-Augmented Generation (RAG) has seemed the go-to method to integrate LLMs with corporate knowledge, avoiding the common pitfall of hallucinations.

In the GenAI arms race among hyperscalers, IBM isn’t usually the first name that comes to mind. However, IBM’s recent developments with its Granite series and InstructLab approach offer a new perspective. The Granite models can be “right-sized” for specific purposes, a significant move toward adaptable AI. More importantly, InstructLab provides an approach that isn’t just about fine-tuning a model but “aligning” it with enterprise-specific data. This process is relatively straightforward and doesn’t require specialized hardware or extensive technical expertise. Let’s dive into how InstructLab works, its synergy with RAG, and why this dual approach could reshape enterprise AI.

The Complementary Roles of InstructLab and RAG

InstructLab and RAG each offer unique ways to enhance LLMs, but they serve different purposes. RAG has become popular for its ability to incorporate real-time, external data into model responses without fundamentally changing the model itself. By retrieving relevant information during an interaction, RAG can reduce hallucinations, as the model is pulling up-to-date context to answer questions accurately. This method works well in dynamic industries where knowledge rapidly evolves and information must be constantly refreshed.

In contrast, InstructLab focuses on embedding an enterprise’s proprietary knowledge within the model itself. By “aligning” the model with domain-specific data, InstructLab allows companies to build an internal framework that doesn’t require real-time retrieval for every response. This approach makes InstructLab ideal for tasks where the model must consistently understand and represent a company’s specialized context.

How InstructLab Works

InstructLab has taken a significant step toward accessibility and efficiency, making it possible for companies to align models with enterprise-specific knowledge without needing GPUs or intensive technical expertise. Here’s how it works:

1. Data Collection and Preparation: Enterprises start by collecting relevant knowledge artifacts, such as documents, FAQs, and support logs. These artifacts provide the context and content that will guide the model’s responses.
2. Automated Instruction Generation: Here’s where InstructLab’s automation shines. When knowledge artifacts are provided, InstructLab analyzes them to create preliminary question-and-answer pairs. This automated step simulates potential interactions without the need for teams to manually develop instructions for each scenario, minimizing labor-intensive setup. Human reviewers then refine the automatically generated instructions, ensuring they align with the company’s communication style and specific needs.
3. Model Training: Using these synthetic question-and-answer pairs, InstructLab trains the model on enterprise-specific data, updating its internal parameters. This alignment allows the model to produce responses that are contextually accurate for the organization, effectively embedding company-specific knowledge into the model’s “memory.”
4. Evaluation and Adjustment: Once trained, the model’s responses are evaluated and iteratively refined until they meet the organization’s standards. This step ensures a consistent quality of response that aligns with company goals.

This process enables businesses to directly update a model with proprietary knowledge in a streamlined, accessible way. By embedding this data into the model itself, companies can rely on the model’s internal understanding of their specific context rather than depending solely on external data retrieval.

Why This Could Be a New Era for Enterprise AI

The democratization of AI customization is the real breakthrough here. Historically, fine-tuning or adapting models required technical skills, powerful hardware, and significant resources—an investment that only large enterprises could afford. InstructLab removes those barriers, allowing even mid-sized companies to align AI with their proprietary knowledge without specialized hardware or extensive technical expertise. This shift brings the power of AI customization to a wider range of businesses, enabling them to tailor models in-house and with standard resources.

This alignment method doesn’t eliminate the need for RAG. Rather, RAG and InstructLab can work in tandem, creating a robust, flexible system. For real-time, dynamic knowledge needs, RAG provides essential, up-to-date information. Meanwhile, InstructLab ensures the model carries a reliable foundation of proprietary knowledge, reducing dependency on external sources for everyday interactions.

Maximizing the Value of AI with Enterprise Data

The rise of proprietary data as a “new asset class” for AI, as NICE’s Barak Eilam predicted, may indeed be upon us. With InstructLab, IBM has introduced an accessible path for companies to truly embed enterprise-specific knowledge into AI models. Combined with RAG for real-time data needs, InstructLab empowers businesses to create AI solutions that are highly accurate, contextually aware, and consistently aligned with organizational goals. This dual approach—enhancing the model’s internal knowledge through InstructLab and augmenting with external data via RAG—opens a new era for enterprise AI, where models don’t just retrieve information; they deliver contextually relevant responses grounded in enterprise-specific knowledge.

Categories: Conversational Intelligence, Intelligent Assistants, Articles