Book a Demo Start Instantly

Knowledge graphs are fundamental tools in the realm of data science and artificial intelligence, used to organize and integrate information through an interconnected network of entities and their interrelations. By structuring data in this way, knowledge graphs facilitate more intuitive data retrieval and sophisticated machine learning models. In this blog, we’ll explore how large language models (LLMs) can be utilized to extract entities and their relationships to build knowledge graphs.

Knowledge Graph

Understanding Knowledge Graphs

A knowledge graph represents a collection of interlinked descriptions of entities — objects, events, or concepts — where each entity is connected by edges that describe the relationships between them. This model helps in representing data in a way that reflects real-world scenarios, allowing for easier navigation and querying of connections.

The Extraction Process

Here’s the prompt we’ll be discussing, which outlines the process of extracting knowledge graph entities and relationships:

After reviewing the initial extraction results, assess whether all relevant entities and their relationships have been identified from the text.

Firstly, consider if any key details, entities, or relationships might still be missing, focusing particularly on technical terms, named entities, and interactions that may not have been captured in the first round.
- Please respond with 'YES' if you believe all entities and relationships have been captured, based on the provided text only, ensuring that no external knowledge or assumptions are used.
- Respond with 'NO' if you identify gaps and think additional entities or relationships need to be identified.

In the case of a 'NO' response, please revisit the text, paying close attention to potentially overlooked or subtle entities and their connections.
List any additional entities and relationships you find, ensuring each is clearly defined, contextualized, and strictly based on the text.

Some key points to follow while identifying additional entities and relationships are:
- Entities must be meaningful, and it should be specific object or concept, avoid ambiguous entities.
  1. Specificity: Entity names must be specific and indicative of their nature.
  2. Diversity: Entities should be identified at multiple levels of detail, from general overviews to specific functionalities.
  3. Uniqueness: Similar entities should be consolidated to avoid redundancy, with each entity distinctly represented.

- Metadata must directly relate to and describe their respective entities.
  1. Accuracy: All metadata should be factual, verifiable from the text, and not based on external assumptions.
  2. Structure: Metadata should be organized in a comprehensive JSON tree, with the first field labeled "topic", facilitating structured data integration and retrieval.

- Relationships:
  1. Entities must be in entities list, don't use non-existing entities.
  2. Carefully examine the text to identify all relationships between clearly-related entities, ensuring each relationship is correctly captured with accurate details about the interactions.
  3. Clearly define the relationships, ensuring accurate directionality that reflects the logical or functional dependencies among entities. 
     This means identifying which entity is the source, which is the target, and what the nature of their relationship is (e.g., $source_entity depends on $target_entity for $relationship).
  4. Extract as many relationships as possible.

- Please endeavor to extract all meaningful entities and relationships from the text, avoid subsequent additional gleanings.
- Maintain language consistency in the terminology used to describe these entities and relationships, except it is necessary to preserve the original meaning.

Your task is to ensure that all extracted entities and their relationships are factual and verifiable directly from the text itself, without relying on external knowledge or assumptions.

Breaking Down the Prompt

1.Initial Assessment:

This section asks the extractor (either human or LLM) to review previously identified entities and relationships to determine completeness. The focus should be on ensuring no key elements are missing from the extraction.

2.Identifying Gaps:

If the initial review finds missing entities or relationships, the prompt directs further scrutiny of the text to find these omissions. This requires a meticulous approach to scan the content for any overlooked details.

3.Criteria for Entities and Relationships:

The prompt outlines specific criteria for entities and relationships:

  • Entities: Must be specific, diverse, and unique.
  • Metadata: Should accurately describe the entities without assumptions.
  • Relationships: Must be clearly defined, with accurate directionality and context.

4.Extraction Methodology:

It emphasizes the necessity to extract all relevant information in one go, maintaining consistency in how entities and relationships are described, based entirely on the text given.


Free free to checkout these demos:


Using LLMs to extract entities and relationships for knowledge graphs can vastly improve the efficiency and accuracy of data organization. LLMs, with their deep understanding of language nuances, are particularly suited to this task, making them invaluable tools for building robust, reliable knowledge graphs. This process, as detailed in the prompt, ensures a thorough and precise extraction, crucial for the integrity of the resulting knowledge graph.

Last updated May 27, 2024

Spin up a Serverless database with 25GiB free resources.

Start Now