GraphRAG Hybrid Search

GraphRAG blends two retrieval strategies: vector similarity (what's semantically similar?) and graph proximity (what's contextually connected?). The result is retrieval that understands not just meaning, but relationships.

Why GraphRAG?

Approach	Finds	Misses
Vector only	Semantically similar content	Connected but differently-worded content
Graph only	Structurally related entities	Content not yet linked to entities
GraphRAG	Both — blended by configurable weight	Nothing (best of both worlds)

Example: You ask "What is Company X's pricing strategy?"

• Vector search finds memories mentioning "pricing" or "strategy"
• Graph traversal finds: Company X → negotiation with Person Y → budget constraint Z → competitor comparison W
• GraphRAG returns all of these, ranked by blended score

How It Works

1. Vector search → top 20 candidate memories
2. Extract entity names from candidates
3. For each entity, traverse 2 hops in the knowledge graph
4. Score: final = (vector_score × 0.7) + (graph_proximity × 0.3)
5. Re-rank and return top N

Graph proximity scoring:

• Directly linked to a traversed entity: 1.0
• 1 hop away: 0.5
• 2 hops away: 0.25

Usage

curl -X POST https://api.ainative.studio/api/v1/public/memory/v2/graph/graphrag \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "What pricing strategies does Company X use?",
    "limit": 10,
    "graph_weight": 0.3,
    "max_hops": 2
  }'

Tuning graph_weight

Value	Behavior	Best For
0.0	Pure vector search	General similarity queries
0.1-0.3	Mostly vector, light graph boost	Default — good for most use cases
0.4-0.6	Balanced	Multi-entity questions
0.7-0.9	Mostly graph	Relationship-focused queries
1.0	Pure graph proximity	"How is X connected to Y?"

Response Shape

{
  "results": [
    {
      "id": "memory-uuid",
      "content": "Company X proposed a 15% discount...",
      "score": 0.82,
      "vector_score": 0.75,
      "graph_boost": 0.98,
      "memory_type": "episodic",
      "entity_id": "company_x",
      "tags": ["negotiation", "pricing"]
    }
  ],
  "query": "pricing strategies",
  "graph_weight": 0.3,
  "max_hops": 2,
  "entities_traversed": ["company_x", "person_y", "competitor_z"],
  "total": 10
}

Prerequisites

GraphRAG works best when:

1. Memories are stored via /remember (builds the vector index)
2. Entities are created via /entity (builds the graph)
3. Edges connect entities via /edge (enables traversal)

Without a graph, GraphRAG falls back to pure vector search (graph_boost = 0 for all results).