Retrieval-augmented generation, or RAG, has played an important role helping organizations leverage their data for advanced applications like AI assistants, document search, and automated summarization. It’s flexible, scalable, and cost-effective.

Now, steady progress in the field of applied generative AI (GenAI) for enterprises is expanding on these benefits and giving businesses options to improve both the accuracy and richness of their data retrieval.

With GraphRAG for example, they can draw on data and text that has been more rigorously connected and defined to enhance the reliability of responses. A recent analysis by AWS partner Lettria found that GraphRAG can improve the precision of replies by up to 35%.

Another emerging retrieval method is agentic search, which uses AI agents to carry out deep, iterative searches to help organizations delve into their data for richer, accurate, and more context-aware answers.

In this article, we’ll explain these two methods – GraphRAG and agentic search – while illustrating their benefits, drawbacks, and examples of how they can be used.

What is GraphRAG?

GraphRAG uses a database called a knowledge graph to improve the accuracy, reliability, and traceability of responses. It does this by having the relationships among different pieces of information like names, places, dates, often called entities, tied together with explicit pathways.

Conventional RAG uses semantic similarities to connect data and text. Consider the example of the word ‘bank.’ It can mean several things: a financial institution, a raised side of the river, or perhaps a turning plane.

If a RAG system is asked ‘How does a bank prevent flooding?’ it needs to understand the question as humans do: that this involves a river bank and rising waters. It does this by noting the word ‘bank’ is being used together with the word ‘flooding.’ This creates a vector that pulls the question’s intent toward a swollen river and away from financial institutions.

This new vector is then used to search for text and data in a database that has similar vectors associated with flooding rivers. This enables an embedding system to generate responses that mirror the context and human-like language of the initial question.

Benefits of GraphRAG

RAG is a very effective and reliable way to find and retrieve information, but it can create ambiguity. For example, if we changed our original question from ‘How does a bank prevent flooding?’ to ‘How does a bank help with flooding?’ the meaning becomes more nuanced.

GraphRAG can cut through this ambiguity. In addition to organizing data with vector embeddings, it uses a knowledge graph to organize data through clearly defined pathways.

In a knowledge graph, the different meanings of bank would be represented as nodes within the database, with some nodes representing a river bank and others representing a financial institution and other meanings. Each of these nodes is linked directly to other nodes representing words like flood or insurance.

These pathways between the nodes for bank and flood, or bank and insurance, are explicitly defined, leaving less room for misinterpretation or error.

The explicit nature of the pathways between data points in GraphRAG can enhance organizations’ control over their data, reducing the risk of information leaking or being seen by unauthorized users. GraphRAG systems still need the same level of permission control implemented as semantic search, but it can potentially be managed in a stricter, more organized way.

With concerns rising about privacy and accuracy, GraphRAG has the potential to give business leaders more peace of mind to move forward with transformative AI solutions in finance, healthcare, and every industry where safety and governance must come first.

Challenges of GraphRAG

While GraphRAG can improve accuracy and traceability, it can also cost more to build and maintain. This is because the explicit nodes and pathways among the data and text take time and money to build. That can also make it harder to scale. This characteristic of GraphRAG also makes it better suited for use cases where data sources are relatively stable rather than being updated regularly.

GraphRAG in action — better accuracy

Organizations that need very high levels of accuracy, traceability, and explainability stand to benefit in particular from GraphRAG compared to standard RAG solutions.

Consider an automaker that needs to manage a highly complex, international supply chain. It might incorporate a knowledge graph to define every supplier and the parts they provide, providing a clear picture of their supplier network and the ability to trace defective parts back to their origin.

Compared to conventional RAG, GraphRAG could better enable the company to trace complex supply chain relationships, giving it advantage in the event of product recalls or compliance audits, as well as for overall risk management.

GraphRAG could also help pharmaceutical companies, providing a detailed map of the proteins, enzymes and other components used to create a new drug. If a side effect emerges later in the drug development process, the company could trace the chain of events that led to the side effect.

Because it is grounded in defined scientific relationships, a GraphRAG solution in this setting could enhance the explainability and causality of outputs – a key benefit for companies in regulated industries.

What is agentic search?

Agentic search gives organizations the ability to delve more deeply into the data, resulting in responses that provide more understanding and context.

While knowledge graphs are like prebuilt maps, agentic search is more like a smart research assistant that can figure out answers even when the information it needs is scattered or ambiguous.

Instead of giving an immediate, simple response to a query, it deploys an AI agent to engage in a search that builds on one response to generate increasingly more informative responses until all attempts at finding an answer are exhausted.

Benefits of agentic search

This back-and-forth, iterative process — similar to one that a human researcher would use — leads to richer retrieval results. Instead of single-pass search that may result in a simple yet unsatisfying response, it keeps delving for patterns and understanding until a fuller picture can be provided.

Agentic search can also mean taking this exhaustive search for an answer and delivering it in a detailed format like a table, CSV or documents that enhance a user’s ability to analyze that answer. Technically this is also possible with GraphRAG, however the generation from agentic search will more often than not be much more comprehensive than the former.

It has the ability to interpret and connect information more dynamically than GraphRAG, and can be particularly powerful in environments where data hasn’t been strictly organized such as emails and Slack messages.

Unlike GraphRAG, it doesn’t need a pre-built database of explicit relationships to create responses, making it more flexible and reducing the initial set-up time and costs.

It also works well when a question is vague or open-ended because the user isn’t sure of exactly what they’re looking for. The interactive process of one response building a better response can sharpen the user’s idea of what they really need.

Challenges of agentic search

The more intensive work that agentic search systems put in to produce answers can lead to higher computation costs. More computation can also increase latency, which could turn a one to two second response into one that takes a minute or more.

And because agentic search is searching wider for more meaning, there’s an increased risk that an answer may stray too far from the original question.

Agentic search in action – unlocking deeper connections

If a candy company is building an ecommerce platform they may want to organize items with very explicit relationships. This might mean a customer can search for a chocolate bar with tags such as ‘nuts,’ ‘dark,’ or ‘sugar.’

It scores high on accuracy but it’s less helpful if a customer uses an unorthodox search term, or inquires about a product based on a recent trend that hasn’t been incorporated into the company’s database.

In a different context, let’s say a user had the following question: “What game systems run Flight Simulator?” A knowledge graph would struggle to answer that correctly if it doesn’t have specific connections encoded to tell it that Flight Simulator is a game that runs on Microsoft XBox. It lacks the flexibility to just “figure it out.”

An agentic system, by contrast, would break the question into parts, thinking through what Flight Simulator is and what “game systems” are. It would then search trusted sources to discover which systems support the game, even doubling back and double checking to ensure it gave the most accurate answer.

Conclusion

The world of enterprise GenAI is constantly changing. And while vector-based retrieval will continue to excel at providing accurate responses in human-like interactions, opportunities are growing to improve the accuracy, depth, and richness of responses with GraphRAG, agentic search, and other emerging methods.

Whether it’s in healthcare or other industries, keeping abreast of the latest retrieval systems and new solutions will help organizations save money, improve lives, and position themselves for the AI innovations still to come.

Building enterprise AI agents

A comprehensive blueprint for developing scalable AI agents in regulated industries.

Download the eBook

Brandon Hoang

Solutions Architect at Cohere.