Vector Protocol

we-trust-vector is a dedicated protocol extension designed for the AI era, specifically for handling the storage and retrieval of high-dimensional vectors (Vector Search).

1. Protocol Features

Unlike traditional KV protocols, the Vector protocol supports the following at the message level:

• Dimension Definition: Specify vector dimensions (e.g., 768, 1536) during write operations.
• Distance Metrics: Supports specifying L2 (Euclidean distance), Cosine (Cosine similarity), or IP (Inner Product) in query messages.
• Top-K Queries: A dedicated instruction set for quickly returning the K most similar results.

2. Data Encoding

Vector data is transmitted using a compact binary floating-point array format:

• f32 Arrays: Standard precision.
• f16 / Int8 Quantization: Supports data quantization and compression at the protocol layer, saving 50% - 75% of network bandwidth.

3. Search Operator Mapping

Vector instructions map directly to YYKV's underlying vector indices (such as HNSW or IVF-Flat):

• MessageType::Query carries the vector payload.
• Storage nodes utilize the vector acceleration module in yykv-compute for parallel retrieval.

4. Use Cases

• RAG (Retrieval-Augmented Generation): Provides real-time, efficient knowledge base retrieval for Large Language Models.
• Image/Video Search: Stores feature vectors and enables fast reverse image/video search.
• Recommendation Systems: Real-time personalized recommendations based on user behavior vectors.

Programming Interface (Pseudo-code)

// Pseudo-code: Initiating a vector search
let query_vector = vec![0.1, 0.2, 0.3, ...];
let request = VectorQuery::new("face_embeddings")
    .vector(query_vector)
    .metric(MetricType::Cosine)
    .limit(10);

let results = conn.vector_search(request).await?;