AWS re:Invent 2025 - Knowledge Graphs for AI and Intelligent Systems (DAT209)

AWS Events

11 views • 7 months ago Save 12 min 5 min read

Video Summary

Enterprise AI projects face a significant hurdle with only 5% reaching production, largely due to data being siloed, inaccessible, or not "AI-ready." This video explores how knowledge graphs can address these challenges by transforming enterprise data into a connected, context-rich asset. Context engineering, an evolution of prompt engineering, aims to provide Large Language Models (LLMs) with the right information at the right time and structure. A particularly interesting fact is that a study found knowledge graphs to be 3x more accurate in LLM responses compared to NoSQL and SQL databases.

Knowledge graphs, utilizing a property graph model with nodes and relationships, offer a structured way to store and access connected data, overcoming the query diversity problem and enabling simpler, more powerful querying. This approach is crucial for agentic AI, which requires reasoning and multi-step actions. By providing this rich context, knowledge graphs make data AI-ready, leading to more trustworthy and explainable AI applications.

Short Highlights

Only 5% of AI projects reach production due to data challenges like silos and lack of AI readiness.
Context engineering provides LLMs with the right data at the right time and structure to perform tasks.
Knowledge graphs, using nodes and relationships, organize connected data, simplifying complex queries and improving context.
Graph RAG replaces traditional vector databases with graph databases to provide richer, connected context to LLMs.
A study found knowledge graphs to be 3x more accurate in LLM responses compared to NoSQL and SQL databases.

Key Details

AI Challenges with Enterprise Data [00:21]

Enterprise data presents significant challenges for AI adoption, with only 5% of projects reaching production due to data being in silos, security constraints, and regulations.
Key AI challenges include hallucinations in LLMs (especially with enterprise data), the "black box" nature of AI preventing clear data understanding, and significant regulatory and compliance hurdles.
A major hurdle is the lack of structured and connected context, leading to issues like data being in disparate sources (data warehouses, lakehouses, document stores), differing schemas and formats, and a problem of query diversity where multiple queries are needed for a single answer.

"The reason might be it's uh the data is in silos. There might be security constraints. There are a lot of regulations and and and stuff like that."

Context Engineering and Knowledge Graphs [05:31]

Context engineering is an evolution of prompt engineering, focused on providing LLMs with sufficient information to reach the next stage or step, ensuring the right data is available at the right time and in the right structure.
Traditional LLM prompts often fail with enterprise data, which requires multi-step reasoning, planning, and tool calls. Context engineering aims to equip agents with the necessary context to understand their actions.
Sources of context include user interactions (prompts, feedback), LLM memory (short-term and long-term), structured outputs from APIs and data tools, and retrieval from vector or graph databases (Graph RAG).

"It's basically um it's basically a evolution of prompt engineering right so uh it the definition by definition it says that you just give enough information to the LLM so that it can uh get to the next stage or next step right"

Understanding Knowledge Graphs [09:00]

Knowledge graphs are designed to organize and access connected data, treating everything as interconnected, whether it's supply chain, financial, or networking data.
They utilize a property graph model, storing data as nodes (entities) and relationships (verbs) with properties attached to both. For example, a "supplier" node "supplies" a "raw material" node, with properties like lead time or capacity.
This inherent structure simplifies querying, eliminates the need for complex joins across tables, and provides a connected view of business data, solving query diversity and context provision issues.

"Supplier supplies raw materials. So supplies that's a verb. That's a relationship. So that's how we store data."

Knowledge Graphs for AI Readiness and Graph RAG [11:20]

Knowledge graphs make data "AI-ready" by providing all necessary connections and context inherently within the database, acting as a knowledge layer that complements the LLM's reasoning and creativity.
Graph RAG is an advancement of traditional RAG (Retrieval Augmented Generation), replacing vector databases with graph databases to leverage the connected view of information and provide richer context to LLMs.
This allows for simpler, single queries for complex tasks, such as finding the shortest path in a supply chain, rather than extensive digging through multiple systems.

"So that's the reason why we say it's a it's a AI ready data because um so we have you have all the connections uh and u context within within the database"

Advantages and Architecture of Knowledge Graphs [13:18]

Knowledge graphs lead to more accuracy, with a study indicating they are 3x more accurate in LLM responses compared to NoSQL and SQL databases, due to their semantic richness and ability to perform pattern matching and pathfinding.
The architecture typically involves a data platform (e.g., AWS services) feeding data into a knowledge graph, which then acts as a context layer for LLMs. This process may involve projecting only relevant data into the knowledge graph for specific algorithms or contextual needs.
Beyond RAG, knowledge graphs facilitate context engineering by providing the right level of context, store user interactions as memory, and enable advanced querying through languages like Cypher for visualization and exploration.