pgvector Integration

9+ Distance Operators via GraphQL

FraiseQL exposes all pgvector distance operators through type-safe GraphQL filters. Dense and sparse vectors, custom distance functions, and vector aggregations.

Operator	GraphQL Filter	Use Case	Range
Dense Vectors
<=> cosine_distance	cosineDistance	Text similarity, recommendations	0.0 (identical) to 2.0 (opposite)
<-> l2_distance	l2Distance	Euclidean distance	0.0+ (smaller = more similar)
<+> l1_distance	l1Distance	Manhattan distance	0.0+ (smaller = more similar)
<#> inner_product	innerProduct	Negative inner product	-∞ to 0.0 (larger = more similar)
<~> hamming_distance	hammingDistance	Binary vectors	0+ bits differ
<%> jaccard_distance	jaccardDistance	Set similarity	0.0 (identical) to 1.0 (no overlap)
Sparse Vectors
sparse_cosine_distance	sparseCosineDistance	Sparse text embeddings	0.0 to 2.0
sparse_l2_distance	sparseL2Distance	Sparse Euclidean distance	0.0+
sparse_inner_product	sparseInnerProduct	Sparse inner product	-∞ to 0.0

GraphQL VectorFilter Schema

Type-safe vector filtering with automatic PostgreSQL query generation. No SQL injection, full type safety.

GraphQL Schema

input VectorFilter {
  # Dense vector operators
  cosineDistance: VectorDistanceFilter
  l2Distance: VectorDistanceFilter
  l1Distance: VectorDistanceFilter
  innerProduct: VectorDistanceFilter
  hammingDistance: VectorDistanceFilter
  jaccardDistance: VectorDistanceFilter

  # Sparse vector operators
  sparseCosineDistance: VectorDistanceFilter
  sparseL2Distance: VectorDistanceFilter
  sparseInnerProduct: VectorDistanceFilter

  # Additional functions
  customDistance: CustomDistanceFilter
  vectorNorm: VectorNormFilter
  quantizedDistance: QuantizedDistanceFilter
}

input VectorDistanceFilter {
  vector: [Float!]!
  threshold: Float
  limit: Int
}

Sub-millisecond Performance

HNSW

Hierarchical Navigable Small World

Industry-standard ANN indexing for billion-scale vector search with 99.9% recall.

IVFFlat

Inverted File Index

PostgreSQL's built-in IVF implementation for high-dimensional vector search.

< 0.5ms

Rust Pipeline

Vector filtering passes through FraiseQL's Rust pipeline for sub-millisecond processing.

Semantic Search Query

GraphQL Query

query SemanticSearch($embedding: [Float!]!) {
  products(
    filter: {
      descriptionEmbedding: {
        cosineDistance: {
          vector: $embedding
          threshold: 0.8
          limit: 10
        }
      }
    }
  ) {
    id
    name
    description
    similarity: _distance
  }
}

Generated PostgreSQL

SELECT
  id, name, description,
  1 - (embedding <=> $1) as _distance
FROM tb_product
WHERE embedding <=> $1 <= 0.8
ORDER BY embedding <=> $1
LIMIT 10;

Why pgvector in PostgreSQL?

✅ FraiseQL + pgvector

• Single PostgreSQL database
• ACID transactions with vectors
• GraphQL API included
• No sync overhead
• Cost-effective scaling
• Native JSONB + vectors

❌ External Vector DBs

• Separate vector database
• Complex sync architecture
• Additional GraphQL layer needed
• Higher operational cost
• Eventual consistency issues
• Multiple data stores

AI/ML Integrations

🤗

Hugging Face

Sentence transformers, CLIP embeddings, local inference

🟣

OpenAI

text-embedding-ada-002, GPT-4 integration patterns

🤖

Cohere

Embed API, multilingual embeddings, custom models

🔍 Vector Search (pgvector)