Knowledge Graph Entity Linking

Category: Knowledge

Problem

Vector search finds semantically similar memories but cannot reason about relationships. Questions like "Who works at the company that provides our cloud hosting?" require traversing entity relationships — something flat memory stores cannot do.

Architecture

Dakera automatically extracts entities from stored memories using GLiNER (a zero-shot NER model). Entities are linked with 4 edge types and stored in a traversable knowledge graph. BFS traversal answers multi-hop questions by following relationship chains.

Edge Types

• RELATES_TO — general association between entities
• WORKS_AT — person-organization relationship
• LOCATED_IN — geographic containment
• PART_OF — compositional relationship

Implementation

from dakera import Dakera

client = Dakera(base_url="http://localhost:3300", api_key="dk-...")

# Store memories — entities are extracted automatically
client.memory.store(
    content="Alice works at Acme Corp as a senior engineer",
    namespace="team",
    metadata={"type": "fact"}
)

client.memory.store(
    content="Acme Corp uses AWS for their cloud infrastructure",
    namespace="team",
    metadata={"type": "fact"}
)

client.memory.store(
    content="Bob is Alice's manager at Acme Corp",
    namespace="team",
    metadata={"type": "fact"}
)

# Query entities extracted from memories
entities = client.knowledge.entities(
    namespace="team"
)
# Returns: [
#   {"name": "Alice", "type": "PERSON", "mentions": 2},
#   {"name": "Acme Corp", "type": "ORGANIZATION", "mentions": 3},
#   {"name": "Bob", "type": "PERSON", "mentions": 1},
#   {"name": "AWS", "type": "ORGANIZATION", "mentions": 1}
# ]

# Traverse the knowledge graph
graph = client.knowledge.graph(
    entity="Alice",
    namespace="team",
    depth=2  # BFS depth
)
# Returns relationships:
# Alice --WORKS_AT--> Acme Corp
# Acme Corp --RELATES_TO--> AWS
# Bob --WORKS_AT--> Acme Corp
# Bob --RELATES_TO--> Alice (manager)

# Multi-hop query: "What cloud provider does Alice's company use?"
# Traversal: Alice -> WORKS_AT -> Acme Corp -> RELATES_TO -> AWS
# Answer: AWS

REST API Example

# Get entities for a namespace
curl -X POST http://localhost:3300/v1/knowledge/entities \
  -H "Authorization: Bearer dk-..." \
  -H "Content-Type: application/json" \
  -d '{"namespace": "team"}'

# Traverse the graph from an entity
curl -X POST http://localhost:3300/v1/knowledge/graph \
  -H "Authorization: Bearer dk-..." \
  -H "Content-Type: application/json" \
  -d '{
    "entity": "Alice",
    "namespace": "team",
    "depth": 2,
    "edge_types": ["WORKS_AT", "RELATES_TO"]
  }'

When to Use This Pattern

• CRM systems tracking people, companies, and relationships
• Research assistants connecting papers, authors, and institutions
• Any application requiring multi-hop reasoning over stored knowledge
• Building organizational knowledge maps from conversations

Key Considerations

• Entity extraction happens automatically on store — no extra API calls needed
• GLiNER runs on-device via ONNX, so entity extraction is private and fast
• Use depth limits on traversal to avoid returning the entire graph
• Combine graph results with vector recall for comprehensive answers
• Knowledge graphs complement (not replace) semantic search — use both