Skip to main content
Every document you insert into Vrin contributes to a growing knowledge graph. This graph is what enables multi-hop reasoning, temporal queries, and fact-level provenance.

​
What’s in the Graph

The knowledge graph consists of:
  • Entities (nodes): People, organizations, products, concepts, events, locations
  • Relationships (edges): Typed, directional connections between entities (e.g., “CEO of”, “acquired”, “reported revenue”)
  • Properties: Confidence scores, temporal markers, source document references

​
How It’s Built

During ingestion, Vrin uses LLMs to extract structured facts from your documents: 
Document: "Jane Smith became CEO of ACME Corp in March 2025"

Extracted:
  Entity: Jane Smith (person)
  Entity: ACME Corp (organization)
  Relationship: Jane Smith --[became_ceo_of]--> ACME Corp
  Temporal: valid_from = 2025-03
  Source: "CEO Announcement Q1 2025"
  Confidence: 0.94
Facts are deduplicated using content hashing. Re-ingesting the same document won’t create duplicates.

​
How It’s Queried

When you ask a question, Vrin uses Personalized PageRank to traverse the graph starting from entities mentioned in your query. This enables multi-hop reasoning: 
Query: "Who leads the company that reported the highest revenue growth?"

Traversal:
  1. Find all entities with "revenue growth" relationships
  2. Rank by growth percentage
  3. Follow "leads" / "CEO of" edges from the top entity
  4. Return: "Jane Smith, CEO of ACME Corp (23% YoY growth)"
This traversal is deterministic (not dependent on LLM attention) and doesn’t degrade with hop count, unlike transformer-based reasoning which struggles beyond 2-3 hops.

​
Temporal Awareness

Every fact in the graph can have temporal bounds:
  • valid_from: When this fact became true
  • valid_to: When this fact stopped being true (null = still current)
This enables queries like “Who was ACME’s CEO in 2023?” even if leadership has changed since.

​
Graph + Vector = Hybrid Retrieval

The knowledge graph and vector index complement each other:
ComponentStrengthsUsed For
Knowledge GraphStructural reasoning, multi-hop, temporalFinding the right entities and connections
Vector IndexSemantic similarity, fuzzy matchingFinding supporting text and context
Both results are fused before being sent to the LLM, giving it structured facts and natural-language evidence.