Large white library with many rows of bookshelves and a modern design with glass windows. [Large language models and search engines working together / Librarian using search engine to find information]

In this new amazing age of generative AI, large language models (LLMs) like ChatGPT and Gemini are often viewed as standalone solutions for managing vast amounts of information. It’s critical to address a common misconception here: no matter how sophisticated an AI becomes, it cannot replace the foundational systems that feed it with information. 

Consider this analogy… 

Like hiring a librarian for a library or a research assistant to consume and distill vast amounts of content into what is most valuable, an LLM can be a potent tool in summarizing content into concise and actionable answers. However, hiring a librarian or research assistant does not eliminate the need for a well-stocked library. In the same way, the most powerful uses of LLMs are those backed by a robust and well-maintained search index. In the case of our analogy, this search index is your library, and your LLM librarian or research assistant cannot function properly without it. 

The Limitations of Standalone AI

LLMs have garnered global attention for their ability to generate human-like text responses, but they operate within certain limitations. The most significant limitation they have now, and likely will have for many years, is that they are not continuously retrainable in real-time. They rely heavily on the data they were last trained on. For instance, an LLM trained on data up to 2020 would not know of events or developments occurring after that year. That’s like a historian who has not read any new research published in the last three years. This temporal limitation necessitates continual updates from an external, current data source. Managing that external, current data source requires many integrated elements, as shown in the exhibit below. 

A diagram showing a general orchestration engine with seven steps. The steps include defining a use case, tracking business case, tracking access control, tracking security and accuracy, and tracking cost control. [General orchestration engine / LLM workflow / Information management process]

These integrated components make up the Generative AI Orchestration Engine. Much like the role a data engine plays as the backbone for integrated best-of-breed applications, the Gen AI Orchestration Engine provides that backbone for a comprehensive Generative AI solution. Each component plays a critical role in maximizing the value available in the use of the core LLM. 

Data Acquisition: Connectors Are the Unsung Heroes

Effective knowledge management and e-commerce platforms require diverse connectors to integrate data from various sources. Just as a library divides its collection into sections — reference for quick facts, periodicals for current events, and archives for historical data — a well-curated search index employing diverse connectors can access and organize as many data sources as are available in an organization. 

This ensures that all the knowledge the organization creates can be leveraged all at once to make the most informed decisions. These connectors are what draw diverse data in with their associated permissions. They allow the search engine to provide instant, security-aware information to the LLM, enabling it to function like a well-informed assistant who knows exactly where everything is stored and who is allowed to access it.

Security & Accuracy: The Power of a Hybrid Search Index

To truly harness the capabilities of LLMs in applications such as knowledge management and ecommerce, a robust search engine is critical. A hybrid search index, for example, combines the precision of traditional lexical search algorithms with the contextual understanding of vector embeddings. This dramatically improves search relevance by understanding query intent rather than just matching keywords. 

Furthermore, this combination creates a dynamic index that not only understands textual content in depth but also adapts to new information as it becomes available. Granting an LLM access to such a search index to search on a user’s behalf ensures the most up-to-date information is presented with the convenience and simplicity of LLM responses.

A row of colorful, horizontal stripes against a white background. [Colorful stripes / Abstract background / Visual representation of data for large language models]

A well-maintained search index acts like a constantly evolving library that provides the latest and most relevant information to the LLM. This ensures that the responses it generates are accurate, current, and only contain information to which the end-user has permission. Just as a library needs to continually acquire new books to remain relevant, a search engine must continuously update its index to provide the most accurate and secure information.

LLM Models: Don’t Lock Yourself In

As of late 2022, the landscape of large language models (LLMs) featured just a handful of notable players. Fast forward to mid-2024, and the scene has expanded dramatically with hundreds of models. The growth trajectory of LLMs is not just steep—it’s practically vertical. Each new model introduced to the market carries distinct enhancements or novel capabilities, occasionally altering the landscape dramatically, though perhaps only temporarily, as newer models quickly overtake their predecessors.

Despite this brisk pace of development, there is no singular LLM that reigns supreme across all applications. The variability among models in terms of speed, accuracy, currency, and cost efficiency underscores the necessity of tailored selection based on specific use cases. For instance, one model might sacrifice speed and be optimal for internal organizational uses where cost considerations are paramount. Another model might prioritize accuracy, making it ideal for customer service. And yet another might optimize for the speed needed in ecommerce applications.

It’s crucial to understand that the best LLM for data indexing and abstraction may differ significantly from the one best suited for handling direct user queries. This is where the Gen AI Orchestration Engine comes into play, designed to future-proof your choice in LLMs. By constructing your systems to be adaptable rather than fixed to a specific LLM, you allow for flexibility as the capabilities of these models evolve. This ensures that your solutions can continue to evolve alongside them to be the most efficient and valuable.

Access Control: Security Through Structured Authorizations

Another important aspect of this integrated approach is security. A search engine with robust connectors can draw data from a variety of sources and manage document-level permissions effectively. This ensures that when a user interacts with an LLM, the information retrieved is not only relevant but also compliant with the permissions defined for the end user. Essentially, the LLM can only access information that the user is permitted to see, much like a military records custodian might restrict access to certain materials based on clearance levels.

A person sitting at a desk using a laptop computer to type. The person is wearing a black shirt and jeans, and they are holding a pen in their right hand. The laptop screen shows text and data. [Person using laptop to access information / Large language model user research / Information retrieval]

Cost Control: Manage The Hidden Costs of LLMs

When discussing the costs of large language models (LLMs), it’s vital to note that while individual users might access them for free or through a modest subscription, the commercial implications are significantly steeper. The expense associated with commercial-grade LLMs can be substantial—often 10 to 20 times that of traditional lexical or semantic search tools. This could equate to a tenfold increase in costs related to ecommerce, internal searches, or customer support. Deciding to incur such expenses should be a strategic choice, not merely a pursuit of the latest technology.

There are, however, strategies for managing these costs effectively. Organizations might consider adopting an open-source LLM, balancing the upfront investment in customization against ongoing API costs. Another approach is to pre-index responses to frequently asked questions, allowing for the reuse of answers without additional expense. Additionally, query routing can optimize resource allocation by matching the complexity of a query to the most appropriate search method—simple inquiries might only need a basic lexical search while the most complex or valuable might require the capabilities of an LLM.

These cost management techniques are essential for organizations aiming to integrate LLMs into their operations without compromising their financial efficiency or strategic objectives.

A Synergistic Future

The Lucidworks Platform represents the future of information management — bringing together advanced search engines with large language models through the Gen AI Orchestration Engine. By ensuring all the components work in harmony, organizations can leverage AI not just as a tool for interaction but as a cornerstone of a comprehensive, secure, and intelligent information management strategy. 

Remember, even the most knowledgeable librarian needs a well-stocked, up-to-date library to provide the best service. So, don’t burn your library. Instead, evaluate your organization’s current technological infrastructure and consider how enhancing your search engine capabilities could not only preserve but significantly enhance the effectiveness of your AI tools. Innovate your library to stay ahead in a competitive landscape.

About Guy Sperry

Read more from this author


Contact us today to learn how Lucidworks can help your team create powerful search and discovery applications for your customers and employees.