Vector Search in Liven: No Separate Database Required

The Problem with Vector Search Today

Semantic search and RAG (Retrieval-Augmented Generation) have become essential for modern applications. But adding vector search to a database typically means:

1. Running a separate vector database (Pinecone, Qdrant, Milvus)
2. Maintaining synchronization between your primary database and the vector index
3. Managing two infrastructure stacks

Liven eliminates this complexity by embedding vector search directly into the storage engine.

How It Works

Liven stores vectors as a first-class data type (DataValue::Vector) alongside the rest of your data. This means:

• Vectors live in the same stream as your other records
• Vector search is just another pipeline stage
• No external index, no synchronization, no extra infrastructure

Int8 Quantization

Liven uses int8 quantized vectors rather than full float32. This gives you:

• 4× smaller storage compared to float32
• Faster scan due to reduced memory bandwidth
• Minimal accuracy loss for most use cases

Using Vector Search

Insert documents with embeddings:

db.insert("documents", "doc1", json!({
    "title": "Machine Learning Fundamentals",
    "embedding": [1, 0, 0, 0, 1]  // int8 quantized
}))?;

Search for similar documents:

let query = vec![1i8, 0, 0, 0, 1];
let results = db.run(
    Pipeline::from("documents")
        .vector_filter("embedding", query, 0.75)
        .limit(5)
)?;

The vector_filter stage:

1. Computes cosine similarity between the query vector and each stored vector
2. Filters results below the threshold
3. Returns records sorted by similarity score

Performance

In our benchmarks, vector search over 100K vectors completes in under 50ms on a single core — competitive with dedicated vector databases for workloads under 10M vectors where Liven's simplicity advantage is most valuable.

When to Use Liven for Vector Search

Liven's vector search is ideal for:

• Edge and IoT devices — single binary, no external dependencies
• Small to medium vector collections (under 10M vectors)
• Applications that already use Liven — no extra infrastructure
• Real-time streaming where vectors arrive with other data

For large-scale vector search with hundreds of millions of vectors, a dedicated vector database may still be the right choice. But for the vast majority of applications, Liven's integrated approach eliminates complexity without sacrificing performance.