Australian National University

Qingqin Fang

**An Introduction to Retrieval Augmented Generation (RAG) and Knowledge
Graph**

<figure>
<img src="https://cdn-images-1.medium.com/max/904/0*UtyEkTLv-rFAJV5a" />
<figcaption>Image from Pexel</figcaption>
</figure>

### Author

* Qingqin Fang (ORCID: [0009--0003--5348--4264](https://orcid.org/0009-0003-5348-4264))

### **Introduction**

Large Language Models (LLMs) have transformed the landscape of natural
language processing, demonstrating exceptional proficiency in generating
text that closely resembles human language. However, a significant
challenge faced by these advanced models is the phenomenon known as
"hallucination." Hallucination in LLMs refers to the generation of
content that appears factual but lacks a basis in reality. This issue
poses a critical obstacle to the safe deployment of LLMs in real-world
applications, where accuracy and reliability are paramount.

In this post, we'll embark on a journey to understand hallucination, its
causes and two effective mitigation techniques: **Retrieval Augmented
Generation** and **Knowledge Graph** integration, including the
classification and evolution. By combining these strategies, we aim to
counter hallucination challenges and contribute to the responsible
development of LLMs.

### **What is Hallucination?**

Hallucination in LLMs occurs when the model generates information that
is factually incorrect or entirely fabricated. This phenomenon stems
from the training phase of the model, where the LLM has been exposed to
vast amounts of online text data. While this extensive training data
enables LLMs to exhibit impressive language fluency, it also opens the
door to potential pitfalls.

The generation of hallucinations in LLMs can be attributed to
several factors:

1\. **Biases in Training Data**: LLMs learn from the data they are
trained on, which may contain biases or inaccuracies. These biases can
influence the model's output, leading to the generation of hallucinated
content.

2\. **Ambiguity in Prompts**: LLMs rely on prompts to generate text.
Ambiguous or unclear prompts can result in the model extrapolating
information incorrectly, leading to hallucinations in the generated
text.

3\. **Lack of Comprehension:** Despite their language generation
capabilities, LLMs may lack true comprehension of the content they
generate. This can result in the model modifying information to
superficially align with the input, even if it is factually incorrect.

### **Hallucination Mitigation**

**Prompt engineering** is the crafting of instructions for Artificial
Intelligence (AI) text generation models, adjusting the parameters of a
sophisticated tool to achieve optimal results, ensuring more refined and
tailored content, resulting in more accurate and reliable outputs.

Some studies have focused on developing innovative models to address
hallucinations, emphasising a continuous and progressive process that
combines algorithmic advancements and improvements in data quality.
Instead of fine-tuning models, these methods implement a comprehensive
model architecture to combat hallucinations.

<figure>
<img
src="https://cdn-images-1.medium.com/max/1024/0*9fVvSrInAjXrXVW3" />
<figcaption>Hallucination Mitigation Techniques in LLMs</figcaption>
</figure>

In this post, we delve into two effective mitigation
strategies --- Retrieval Augmented Generation (RAG) and Knowledge Graph
(KG) integration --- to address hallucination challenges in LLMs.

### **Retrieval Augmented Generation (RAG) in LLM**

RAG is a technique used to enhance the responses of LLMs by
incorporating external, authoritative knowledge bases during the text
generation process. By tapping into external sources of information, RAG
provides additional context and verifiable data to guide the generation
of text, ultimately improving the accuracy and reliability of the
generated content.

**RAG Categories Based on its Implementation:**

1\. Before Generation: Techniques involving information retrieval before
text generation, like LLM-Augmenter.

2\. During Generation: Methods utilising external knowledge retrieval
mechanisms during text generation, such as Knowledge Retrieval and
Decompose-and-Query framework.

3\. After Generation: Approaches focusing on post-generation
verification and refinement, like RARR and High Entropy Word Spotting
and Replacement.

4\. End-to-End: Holistic integration of retrieval mechanisms throughout
the entire text generation process, exemplified by RAG.

By categorising RAG into these subcategories, researchers and developers
can better understand the diverse applications and methodologies of
retrieval augmented generation in mitigating hallucination in large
language models.

<figure>
<img
src="https://cdn-images-1.medium.com/max/1024/0*Vqf5oczQPp_leeS-" />
<figcaption>Comparison between the three paradigms of RAG. Source:
Shirui Pan et al., 2023. Link: <a
href="https://arxiv.org/abs/2306.08302">https://arxiv.org/abs/2306.08302</a></figcaption>
</figure>

### The Evolution of RAG Paradigms

1\. Naive RAG: Initial approach following a "Retrieve-Read" methodology,
cost-effective but with limitations.

2\. Advanced RAG: Introduced sophisticated elements like query rewriting
and chunk reranking to enhance LLM performance.

3\. Modular RAG: Further refinement with new modules like Search and
Memory Modules to improve retrieval accuracy.

One interesting aspect of the evolution from Naive RAG to Advanced RAG
and Modular RAG, signifying significant progress in the field. Naive
RAG, initially adopted after ChatGPT, followed a traditional
"Retrieve-Read" approach but had limitations despite being
cost-effective. Advanced RAG introduced advanced elements like query
rewriting and chunk reranking, enhancing LLM performance. Modular RAG
further improved this by adding modules like Search and Memory Modules
for better retrieval accuracy.

This progression addressed the limitations of Naive RAG by integrating
more advanced technologies. Advanced RAG and Modular RAG aimed to
optimise retrieval-augmented generation systems through complex
architectures and new modules. The integration of RAG with fine-tuning,
a notable technique, combines retrieval-based methods and fine-tuning to
enhance RAG model performance effectively. By exploring optimal
integration strategies like sequential or joint training, this hybrid
approach shows promise in enhancing large language models through a
blend of retrieval and fine-tuning methods.

While RAG is a cutting-edge technology with growth potential, it still
faces challenges. Improving user-friendliness, query understanding, data
compatibility, and evaluation methods are key areas for development.
Ongoing research is crucial to unlock RAG's full potential and shape its
future in language models and generative tasks, leveraging its strengths
in integrating retrieval and fine-tuning methods.

### **Knowledge Graph in LLM**

A Knowledge Graph (KG) is an organised collection of data that
represents entities, their attributes, and the relationships between
them in a structured format. KGs enable machines to understand the
semantic meaning and connections within the data, facilitating
sophisticated reasoning, data analysis, and information retrieval.
Integrating KGs into LLMs involves incorporating external knowledge
sources into the pre-training and inference stages of the models. This
integration aims to enhance the understanding of acquired knowledge by
LLMs, enabling them to generate more informed and contextually accurate
responses.

**Knowledge Graph Categories Based on its Implementation:**

1\. KG-enhanced LLMs: This framework focuses on integrating KGs into
LLMs to enhance knowledge understanding during pre-training and
inference. By leveraging external knowledge, this approach aims to
improve LLM performance and interpretability.

2\. LLM-augmented KGs: In this framework, LLMs are utilised for various
KG tasks such as embedding, completion, construction, text generation,
and question answering. The emphasis is on using LLMs to enhance
KG-related tasks by incorporating textual information and improving
downstream task performance.

3\. Synergized LLMs + KGs: This framework emphasises a collaborative
partnership between LLMs and KGs, enabling bidirectional reasoning
through data and knowledge exchange. By leveraging the strengths of both
models, this approach aims to enhance the capabilities of LLMs and KGs
in a mutually beneficial manner.

<figure>
<img
src="https://cdn-images-1.medium.com/max/1024/0*G7rOC_hNfsgwuhaF" />
<figcaption>The general roadmap of unifying KGs and LLMs. Source: Shirui
Pan et al., 2023. Link: <a
href="https://arxiv.org/abs/2306.08302">https://arxiv.org/abs/2306.08302</a></figcaption>
</figure>

The frameworks, including KG-enhanced LLMs, LLM-augmented KGs, and
Synergized LLMs + KGs, emphasise the importance of integrating external
knowledge sources into LLMs to enhance knowledge understanding,
interpretability, and task performance. This structured approach
facilitates bidirectional reasoning, data exchange, and collaborative
utilisation of data and knowledge, leading to more informed and
contextually accurate responses generated by LLMs.

Overall, the integration of KGs into LLMs not only enhances model
performance but also broadens the scope of applications in natural
language processing tasks. By leveraging external knowledge sources,
LLMs can access a wealth of information, improve response accuracy, and
provide more contextually relevant outputs, ultimately advancing the
field of computational linguistics and contributing to the development
of more sophisticated language models.

### **Conclusion**

Addressing hallucination mitigation in LLMs involves navigating a
complex challenge through a range of creative methodologies. Looking
ahead, a crucial direction lies in countering hallucinations through the
integration of multiple mitigation strategies, specifically by combining
RAG and Knowledge Graph (KG). Furthermore, diminishing reliance on
labelled data and exploring unsupervised or weakly supervised learning
methods could contribute to enhanced scalability and adaptability.

By adeptly leveraging these techniques, LLMs hold the potential to
produce text that is not only more reliable but also contextually
relevant and factually accurate. This advancement is pivotal for the
field of natural language processing, ensuring responsible AI
development and fostering a more profound understanding of
language nuances.

As we chart the course forward, ongoing research remains paramount.
Continued exploration into the intricacies of combining RAG and KG,
improving user-friendliness, refining query understanding, accommodating
various data types, and establishing robust evaluation methods will be
instrumental. The collaborative efforts of researchers and developers
are key to unlocking the full potential of these approaches and shaping
the future landscape of language models and generative tasks.

![](https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a96c6a33bf02){width="1"
height="1"}


<strong>
 An Introduction to Retrieval Augmented Generation (RAG) and Knowledge Graph
</strong>
Author  Qingqin Fang (ORCID: 0009–0003–5348–4264)
<strong>
 Introduction
</strong>
Large Language Models (LLMs) have transformed the landscape of natural language processing, demonstrating exceptional proficiency in generating text that closely resembles human language.

Messaggi di Rogue Scholar

Hallucination in Large Language Models and Two Effective Alleviation Pathways