Ai Retrieval Graphrag And Agentic Search

North Mini Code. Cohere's first model for developers.

Learn more

Apr 28, 2025

6 minutes read

Boosting AI retrieval: From GraphRAG to agentic search

Businesses can improve both the performance and reliability of RAG solutions with innovations like GraphRAG and agentic search.

!Blog Post Featured Image

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...