knowledge-graph-builder（知识图谱构建器）

Knowledge Graph Builder

Build structured knowledge graphs for enhanced AI system performance through relational knowledge.

Core Principle

Knowledge graphs make implicit relationships explicit, enabling AI systems to reason about connections, verify facts, and avoid hallucinations.

When to Use Knowledge Graphs

Use Knowledge Graphs When:

• ✅ Complex entity relationships are central to your domain
• ✅ Need to verify AI-generated facts against structured knowledge
• ✅ Semantic search and relationship traversal required
• ✅ Data has rich interconnections (people, organizations, products)
• ✅ Need to answer "how are X and Y related?" queries
• ✅ Building recommendation systems based on relationships
• ✅ Fraud detection or pattern recognition across connected data

Don't Use Knowledge Graphs When:

• ❌ Simple tabular data (use relational DB)
• ❌ Purely document-based search (use RAG with vector DB)
• ❌ No significant relationships between entities
• ❌ Team lacks graph modeling expertise
• ❌ Read-heavy workload with no traversal (use traditional DB)

6-Phase Knowledge Graph Implementation

Phase 1: Ontology Design

Goal: Define entities, relationships, and properties for your domain

Entity Types (Nodes):

• Person, Organization, Location, Product, Concept, Event, Document

Relationship Types (Edges):

• Hierarchical: IS_A, PART_OF, REPORTS_TO
• Associative: WORKS_FOR, LOCATED_IN, AUTHORED_BY, RELATED_TO
• Temporal: CREATED_ON, OCCURRED_BEFORE, OCCURRED_AFTER

Properties (Attributes):

• Node properties: id, name, type, created_at, metadata
• Edge properties: type, confidence, source, timestamp

Example Ontology:

# RDF/Turtle format
@prefix : <http://example.org/ontology#> .

:Person a owl:Class ;
    rdfs:label "Person" .

:Organization a owl:Class ;
    rdfs:label "Organization" .

:worksFor a owl:ObjectProperty ;
    rdfs:domain :Person ;
    rdfs:range :Organization ;
    rdfs:label "works for" .

Validation:

• [ ] Entities cover all domain concepts
• [ ] Relationships capture key connections
• [ ] Ontology reviewed with domain experts
• [ ] Classification hierarchy defined (is-a relationships)

Phase 2: Graph Database Selection

Decision Matrix:

Neo4j (Recommended for most):

• Pros: Mature, Cypher query language, graph algorithms, excellent visualization
• Cons: Licensing costs for enterprise, scaling complexity
• Use when: Complex queries, graph algorithms, team can learn Cypher

Amazon Neptune:

• Pros: Managed service, supports Gremlin and SPARQL, AWS integration
• Cons: Vendor lock-in, more expensive than self-hosted
• Use when: AWS infrastructure, need managed service, compliance requirements

ArangoDB:

• Pros: Multi-model (graph + document + key-value), JavaScript queries
• Cons: Smaller community, fewer graph-specific features
• Use when: Need document DB + graph in one system

TigerGraph:

• Pros: Best performance for deep traversals, parallel processing
• Cons: Complex setup, higher learning curve
• Use when: Massive graphs (billions of edges), real-time analytics

Technology Stack:

graph_database: "Neo4j Community" # or Enterprise for production
vector_integration: "Pinecone" # For hybrid search
embeddings: "text-embedding-3-large" # OpenAI
etl: "Apache Airflow" # For data pipelines

Neo4j Schema Setup:

// Create constraints for uniqueness
CREATE CONSTRAINT person_id IF NOT EXISTS
FOR (p:Person) REQUIRE p.id IS UNIQUE;

CREATE CONSTRAINT org_name IF NOT EXISTS
FOR (o:Organization) REQUIRE o.name IS UNIQUE;

// Create indexes for performance
CREATE INDEX entity_search IF NOT EXISTS
FOR (e:Entity) ON (e.name, e.type);

CREATE INDEX relationship_type IF NOT EXISTS
FOR ()-[r:RELATED_TO]-() ON (r.type, r.confidence);

Phase 3: Entity Extraction & Relationship Building

Goal: Extract entities and relationships from data sources

Data Sources:

• Structured: Databases, APIs, CSV files
• Unstructured: Documents, web content, text files
• Semi-structured: JSON, XML, knowledge bases

Entity Extraction Pipeline:

class EntityExtractionPipeline:
    def __init__(self):
        self.ner_model = load_ner_model()  # spaCy, Hugging Face
        self.entity_linker = EntityLinker()
        self.deduplicator = EntityDeduplicator()

    def process_text(self, text: str) -> List[Entity]:
        # 1. Extract named entities
        entities = self.ner_model.extract(text)

        # 2. Link to existing entities (entity resolution)
        linked_entities = self.entity_linker.link(entities)

        # 3. Deduplicate and resolve conflicts
        resolved_entities = self.deduplicator.resolve(linked_entities)

        return resolved_entities

Relationship Extraction:

class RelationshipExtractor:
    def extract_relationships(self, entities: List[Entity],
                            text: str) -> List[Relationship]:
        relationships = []

        # Use dependency parsing or LLM for extraction
        doc = self.nlp(text)
        for sent in doc.sents:
            rels = self.extract_from_sentence(sent, entities)
            relationships.extend(rels)

        # Validate against ontology
        valid_relationships = self.validate_relationships(relationships)
        return valid_relationships

LLM-Based Extraction (for complex relationships):

def extract_with_llm(text: str) -> List[Relationship]:
    prompt = f"""
    Extract entities and relationships from this text:
    {text}

    Format: (Entity1, Relationship, Entity2, Confidence)
    Only extract factual relationships.
    """

    response = llm.generate(prompt)
    relationships = parse_llm_response(response)
    return relationships

Validation:

• [ ] Entity extraction accuracy >85%
• [ ] Entity deduplication working
• [ ] Relationships validated against ontology
• [ ] Confidence scores assigned

Phase 4: Hybrid Knowledge-Vector Architecture

Goal: Combine structured graph with semantic vector search

Architecture:

class HybridKnowledgeSystem:
    def __init__(self):
        self.graph_db = Neo4jConnection()
        self.vector_db = PineconeClient()
        self.embedding_model = OpenAIEmbeddings()

    def store_entity(self, entity: Entity):
        # Store structured data in graph
        self.graph_db.create_node(entity)

        # Store embeddings in vector database
        embedding = self.embedding_model.embed(entity.description)
        self.vector_db.upsert(
            id=entity.id,
            values=embedding,
            metadata=entity.metadata
        )

    def hybrid_search(self, query: str, top_k: int = 10) -> SearchResults:
        # 1. Vector similarity search
        query_embedding = self.embedding_model.embed(query)
        vector_results = self.vector_db.query(
            vector=query_embedding,
            top_k=100
        )

        # 2. Graph traversal from vector results
        entity_ids = [r.id for r in vector_results.matches]
        graph_results = self.graph_db.get_subgraph(entity_ids, max_hops=2)

        # 3. Merge and rank results
        merged = self.merge_results(vector_results, graph_results)
        return merged[:top_k]

Benefits of Hybrid Approach:

• Vector search: Semantic similarity, flexible queries
• Graph traversal: Relationship-based reasoning, context expansion
• Combined: Best of both worlds

Phase 5: Query Patterns & API Design

Common Query Patterns:

1. Find Entity:

MATCH (e:Entity {id: $entity_id})
RETURN e

2. Find Relationships:

MATCH (source:Entity {id: $entity_id})-[r]-(target)
RETURN source, r, target
LIMIT 20

3. Path Between Entities:

MATCH path = shortestPath(
  (source:Person {id: $source_id})-[*..5]-(target:Person {id: $target_id})
)
RETURN path

4. Multi-Hop Traversal:

MATCH (p:Person {name: $name})-[:WORKS_FOR]->(o:Organization)-[:LOCATED_IN]->(l:Location)
RETURN p.name, o.name, l.city

5. Recommendation Query:

// Find people similar to this person based on shared organizations
MATCH (p1:Person {id: $person_id})-[:WORKS_FOR]->(o:Organization)<-[:WORKS_FOR]-(p2:Person)
WHERE p1 <> p2
RETURN p2, COUNT(o) AS shared_orgs
ORDER BY shared_orgs DESC
LIMIT 10

Knowledge Graph API:

class KnowledgeGraphAPI:
    def __init__(self, graph_db):
        self.graph = graph_db

    def find_entity(self, entity_name: str) -> Entity:
        """Find entity by name with fuzzy matching"""
        query = """
        MATCH (e:Entity)
        WHERE e.name CONTAINS $name
        RETURN e
        ORDER BY apoc.text.levenshtein(e.name, $name)
        LIMIT 1
        """
        return self.graph.run(query, name=entity_name).single()

    def find_relationships(self, entity_id: str,
                         relationship_type: str = None,
                         max_hops: int = 2) -> List[Relationship]:
        """Find relationships within specified hops"""
        query = f"""
        MATCH (source:Entity {{id: $entity_id}})
        MATCH path = (source)-[r*1..{max_hops}]-(target)
        RETURN path, relationships(path) AS rels
        LIMIT 100
        """
        return self.graph.run(query, entity_id=entity_id).data()

    def get_subgraph(self, entity_ids: List[str],
                    max_hops: int = 2) -> Subgraph:
        """Get connected subgraph for multiple entities"""
        query = f"""
        MATCH (e:Entity)
        WHERE e.id IN $entity_ids
        CALL apoc.path.subgraphAll(e, {{maxLevel: {max_hops}}})
        YIELD nodes, relationships
        RETURN nodes, relationships
        """
        return self.graph.run(query, entity_ids=entity_ids).data()

Phase 6: AI Integration & Hallucination Prevention

Goal: Use knowledge graph to ground LLM responses and detect hallucinations

Knowledge Graph RAG:

class KnowledgeGraphRAG:
    def __init__(self, kg_api, llm_client):
        self.kg = kg_api
        self.llm = llm_client

    def retrieve_context(self, query: str) -> str:
        # Extract entities from query
        entities = self.extract_entities_from_query(query)

        # Retrieve relevant subgraph
        subgraph = self.kg.get_subgraph(
            [e.id for e in entities],
            max_hops=2
        )

        # Format subgraph for LLM
        context = self.format_subgraph_for_llm(subgraph)
        return context

    def generate_with_grounding(self, query: str) -> GroundedResponse:
        context = self.retrieve_context(query)

        prompt = f"""
        Context from knowledge graph:
        {context}

        User query: {query}

        Answer based only on the provided context. Include source entities.
        """

        response = self.llm.generate(prompt)

        return GroundedResponse(
            response=response,
            sources=self.extract_sources(context),
            confidence=self.calculate_confidence(response, context)
        )

Hallucination Detection:

class HallucinationDetector:
    def __init__(self, knowledge_graph):
        self.kg = knowledge_graph

    def verify_claim(self, claim: str) -> VerificationResult:
        # Parse claim into (subject, predicate, object)
        parsed_claim = self.parse_claim(claim)

        # Query knowledge graph for evidence
        evidence = self.kg.find_evidence(
            parsed_claim.subject,
            parsed_claim.predicate,
            parsed_claim.object
        )

        if evidence:
            return VerificationResult(
                is_supported=True,
                evidence=evidence,
                confidence=evidence.confidence
            )

        # Check for contradictory evidence
        contradiction = self.kg.find_contradiction(parsed_claim)

        return VerificationResult(
            is_supported=False,
            is_contradicted=bool(contradiction),
            contradiction=contradiction
        )

Key Principles

1. Start with Ontology

Define your schema before ingesting data. Changing ontology later is expensive.

2. Entity Resolution is Critical

Deduplicate entities aggressively. "Apple Inc", "Apple", "Apple Computer" → same entity.

3. Confidence Scores on Everything

Every relationship should have a confidence score (0.0-1.0) and source.

4. Incremental Building

Don't try to model entire domain at once. Start with core entities and expand.

5. Hybrid Architecture Wins

Combine graph traversal (structured) with vector search (semantic) for best results.

Common Use Cases

1. Question Answering:

• Extract entities from question
• Traverse graph to find answer
• Return path as explanation

2. Recommendation:

• Find similar entities via shared relationships
• Rank by relationship strength
• Return top-K recommendations

3. Fraud Detection:

• Model transactions as graph
• Find suspicious patterns (cycles, anomalies)
• Flag for review

4. Knowledge Discovery:

• Identify implicit relationships
• Suggest missing connections
• Validate with domain experts

5. Semantic Search:

• Hybrid vector + graph search
• Expand context via relationships
• Return rich connected results

Technology Recommendations

For MVPs (<10K entities):

• Neo4j Community Edition (free)
• SQLite for metadata
• OpenAI embeddings
• FastAPI for API layer

For Production (10K-1M entities):

• Neo4j Enterprise or ArangoDB
• Pinecone for vector search
• Airflow for ETL
• GraphQL API

For Enterprise (1M+ entities):

• Neo4j Enterprise or TigerGraph
• Distributed vector DB (Pinecone, Weaviate)
• Kafka for streaming
• Kubernetes deployment

Validation Checklist

• [ ] Ontology designed and validated with domain experts
• [ ] Graph database selected and set up
• [ ] Entity extraction pipeline tested (>85% accuracy)
• [ ] Relationship extraction validated
• [ ] Hybrid search (graph + vector) implemented
• [ ] Query API created and documented
• [ ] AI integration tested (RAG or hallucination detection)
• [ ] Performance benchmarks met (query <100ms for common patterns)
• [ ] Data quality monitoring in place
• [ ] Backup and recovery tested

Related Resources

Related Skills:

• rag-implementer - For hybrid KG+RAG systems
• multi-agent-architect - For knowledge-graph-powered agents
• api-designer - For KG API design

Related Patterns:

• META/DECISION-FRAMEWORK.md - Graph DB selection
• STANDARDS/architecture-patterns/knowledge-graph-pattern.md - KG architectures (when created)

Related Playbooks:

• PLAYBOOKS/deploy-neo4j.md - Neo4j deployment (when created)
• PLAYBOOKS/build-kg-rag-system.md - KG-RAG integration (when created)