Knowledge Graphs

Knowledge graphs provide structured representations of entities and their relationships, enabling AI agents to reason over interconnected data rather than processing information in isolation. By combining knowledge graphs with LLMs through patterns like GraphRAG, agents gain the ability to perform multi-hop reasoning, track entity relationships, and maintain factual grounding that flat vector search cannot achieve.

A knowledge graph represents information as a network of entities (nodes) connected by relationships (edges), with properties on both. Unlike tabular databases, knowledge graphs excel at representing complex, interconnected domains where the relationships between things are as important as the things themselves.

• Nodes represent entities: people, companies, concepts, documents
• Edges represent relationships: “works_at”, “cites”, “related_to”
• Properties store attributes on both nodes and edges

Formally, a knowledge graph is a set of triples $G = \{(h, r, t) \mid h, t \in \mathcal{E}, r \in \mathcal{R}\}$, where $\mathcal{E}$ is the set of entities, $\mathcal{R}$ is the set of relation types, $h$ is the head entity, $r$ is the relation, and $t$ is the tail entity.

Property graphs (used by Neo4j, Amazon Neptune, Memgraph) dominate agent applications due to their flexibility and developer-friendly query languages.

Knowledge graph embedding methods learn low-dimensional vector representations for entities and relations, enabling link prediction and reasoning. Common scoring functions include:

• TransE: Models relations as translations, scoring $f(h, r, t) = -||\mathbf{h} + \mathbf{r} - \mathbf{t}||$ where entities and relations are embedded as vectors
• DistMult: Uses bilinear scoring $f(h, r, t) = \mathbf{h}^\top \text{diag}(\mathbf{r}) \, \mathbf{t} = \sum_i h_i \cdot r_i \cdot t_i$
• ComplEx: Extends to complex-valued embeddings $f(h, r, t) = \text{Re}(\sum_i h_i \cdot r_i \cdot \bar{t}_i)$ for asymmetric relations

Knowledge graphs serve as centralized, dynamic hubs that allow specialized agents to access and share contextual data. Key capabilities include:

• Structured retrieval — Agents traverse relationships to find connected information that vector similarity search misses
• Multi-hop reasoning — Following chains of relationships to answer complex questions (e.g., “Which suppliers serve customers in regions with regulatory changes?”)
• Entity resolution — Disambiguating references to the same entity across different data sources
• Temporal tracking — Recording when facts were true, enabling reasoning about change over time
• Shared context — Multiple agents access the same knowledge graph as a shared source of truth

Microsoft's GraphRAG combines knowledge graph construction with retrieval-augmented generation:

1. Entity extraction — LLMs extract entities and relationships from source documents
2. Graph construction — Extracted entities are organized into a knowledge graph with community detection
3. Hierarchical summarization — Communities are summarized at multiple levels of abstraction
4. Graph-enhanced retrieval — Queries traverse the graph structure for context, supplementing vector search

GraphRAG excels on complex analytical queries that require understanding entity relationships across many documents.

from neo4j import GraphDatabase
 
class KnowledgeGraphAgent:
    def __init__(self, uri, auth, llm_client):
        self.driver = GraphDatabase.driver(uri, auth=auth)
        self.llm = llm_client
 
    def query_graph(self, question: str) -> str:
        # LLM generates a Cypher query from natural language
        cypher = self.llm.invoke(
            f"Convert this question to a Neo4j Cypher query:\n{question}\n"
            f"Schema: (Person)-[:WORKS_AT]->(Company)-[:LOCATED_IN]->(City)"
        )
 
        # Execute the graph query
        with self.driver.session() as session:
            results = session.run(cypher)
            data = [dict(record) for record in results]
 
        # Generate natural language response from graph data
        response = self.llm.invoke(
            f"Question: {question}\n"
            f"Graph results: {data}\n"
            f"Provide a clear answer based on these results."
        )
        return response
 
    def add_knowledge(self, text: str):
        # Extract entities and relationships from text
        entities = self.llm.invoke(
            f"Extract entities and relationships from this text as "
            f"(subject, relationship, object) triples:\n{text}"
        )
        # Store in graph
        with self.driver.session() as session:
            for subject, rel, obj in parse_triples(entities):
                session.run(
                    "MERGE (a:Entity {name: $subj}) "
                    "MERGE (b:Entity {name: $obj}) "
                    "MERGE (a)-[r:RELATES {type: $rel}]->(b)",
                    subj=subject, obj=obj, rel=rel
                )
 
agent = KnowledgeGraphAgent(
    "bolt://localhost:7687",
    ("neo4j", "password"),
    llm_client
)
agent.add_knowledge("Anthropic, founded by Dario Amodei, is headquartered in San Francisco.")
answer = agent.query_graph("Where is the company founded by Dario Amodei located?")

Modern entity extraction for knowledge graphs uses LLMs to identify entities and relationships from unstructured text. Key approaches:

• Zero-shot extraction — Prompt LLMs to extract entities without examples
• NER + relation extraction — Combine named entity recognition with relationship classification
• Iterative refinement — Multiple passes to resolve coreferences and disambiguate entities
• Schema-guided extraction — Constrain extraction to a predefined ontology

• Microsoft GraphRAG
• Neo4j Graph Database
• Knowledge Graphs Reshaping AI Workflows 2026

• Retrieval Augmented Generation — RAG patterns including GraphRAG
• Embeddings — Vector representations complementing graph retrieval
• Agent Memory Frameworks — Memory systems using knowledge graphs
• Agent Orchestration — Multi-agent systems sharing knowledge graphs