Core Features

Search

Copy page

Cortex implements hybrid search that combines three search methods for optimal results: vector similarity, keyword matching, and knowledge graph traversal.

Search Methods

Vector Search (Semantic)

Uses embedding similarity to find conceptually related content, even with different wording.

• Powered by OpenAI text-embedding-3-small (1536 dimensions)
• Stored and queried via Neo4j's native vector index
• Finds related concepts, synonyms, and paraphrases

Keyword Search (Full-Text)

Traditional full-text search for exact term matching.

• Uses Neo4j's full-text indexes
• Great for specific terms, names, and codes
• Supports phrase matching

Graph Traversal

Follows relationships in the knowledge graph to find connected information.

• Traverses up to 2 hops from matched entities
• Finds related entities even if not mentioned in the query
• Leverages GraphRAG entity extraction

Hybrid Search Architecture

API Usage

Basic Search

Code
 
curl -X POST "http://localhost:8000/api/search" \
  -H "X-API-Key: your-api-key" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "machine learning algorithms",
    "limit": 10
  }'

Response

Code
 
{
  "results": [
    {
      "id": "chunk_abc123",
      "content": "Machine learning algorithms can be categorized into supervised, unsupervised, and reinforcement learning...",
      "score": 0.92,
      "document_id": "doc_xyz789",
      "document_title": "ML Fundamentals.pdf",
      "metadata": {
        "page": 15,
        "chunk_index": 3
      }
    }
  ],
  "total": 45,
  "query_time_ms": 127
}

Advanced Search

Code
 
curl -X POST "http://localhost:8000/api/search" \
  -H "X-API-Key: your-api-key" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "neural network training",
    "limit": 20,
    "collection_id": "research-papers",
    "search_type": "hybrid",
    "filters": {
      "document_type": "pdf"
    }
  }'

Search Types

Type	Description	When to Use
hybrid	All methods combined (default)	General queries
vector	Semantic search only	Conceptual/abstract queries
keyword	Full-text search only	Exact term matching
graph	Graph traversal only	Relationship exploration

Reciprocal Rank Fusion (RRF)

Cortex uses RRF to combine results from multiple search methods:

Code
 
RRF_score(d) = Σ 1 / (k + rank(d))

Where:

• k is a constant (default: 60)
• rank(d) is the position in each result list

The weights for each method are configurable:

Code
 
VECTOR_WEIGHT=0.5
KEYWORD_WEIGHT=0.3
GRAPH_WEIGHT=0.2

Cross-Encoder Re-ranking

After RRF fusion, a cross-encoder model re-ranks the top candidates for higher precision:

• Model: cross-encoder/ms-marco-MiniLM-L-6-v2
• Considers query-document pairs together
• Significantly improves relevance of top results

Enable/disable via:

Code
 
ENABLE_RERANKING=true
RERANKING_MODEL=cross-encoder/ms-marco-MiniLM-L-6-v2

Collection-Scoped Search

By default, search queries span all collections. To narrow results to a specific collection, pass collection_id:

Code
 
# Search all collections (default — omit collection_id)
curl -X POST "http://localhost:8000/api/search" \
  -H "X-API-Key: your-api-key" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "quarterly results"
  }'

# Search a specific collection
curl -X POST "http://localhost:8000/api/search" \
  -H "X-API-Key: your-api-key" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "quarterly results",
    "collection_id": "financial-reports"
  }'

Graph Context in Search

When graph search is enabled, results may include entity relationships:

Code
 
{
  "results": [...],
  "graph_context": {
    "entities": [
      {"name": "Neural Networks", "type": "Concept"},
      {"name": "Deep Learning", "type": "Concept"}
    ],
    "relationships": [
      {
        "source": "Neural Networks",
        "target": "Deep Learning",
        "type": "PART_OF"
      }
    ]
  }
}

Performance Tips

1. Use collection scoping when needed - Search spans all collections by default; narrow to a specific collection for focused results
2. Adjust weights - Tune VECTOR_WEIGHT, KEYWORD_WEIGHT, GRAPH_WEIGHT for your use case
3. Limit results - Start with smaller limits and paginate
4. Enable re-ranking - Improves precision for the final results