rag-plusplus-core 0.1.0

High-performance retrieval engine with SIMD-accelerated vector search and trajectory memory
Documentation

rag-plusplus-core

High-performance retrieval engine in Rust with SIMD-accelerated vector search and trajectory memory.

Features

  • SIMD-accelerated vector distances - AVX2 optimized L2, cosine, inner product
  • HNSW index - Approximate nearest neighbor search with configurable parameters
  • Flat index - Exact brute-force search for small datasets
  • Trajectory memory - 5D coordinate system (depth, sibling, homogeneity, temporal, complexity)
  • Write-ahead log - Durable operations with checkpoint/recovery
  • Query result caching - LRU cache with configurable TTL
  • Product quantization - Memory-efficient vector compression (SQ8, PQ)

Quick Start

use rag_plusplus_core::{HNSWIndex, HNSWConfig, MemoryRecord, RecordId};

// Create an HNSW index for 768-dimensional vectors
let config = HNSWConfig::default();
let mut index = HNSWIndex::new(768, config);

// Add vectors
let record = MemoryRecord::new(
    RecordId::new("doc-1"),
    vec![0.1; 768],
    Default::default(),
);
index.add(&record)?;

// Search for nearest neighbors
let query = vec![0.15; 768];
let results = index.search(&query, 10)?;

for (id, distance) in results {
    println!("Found: {} at distance {}", id, distance);
}

Modules

Module Description
index Vector indices (HNSW, Flat) with fusion strategies
distance SIMD-optimized distance functions
retrieval Query engine with reranking
store Record storage trait and implementations
trajectory 5D trajectory coordinates, salience scoring, ring structures
quantization Vector compression (SQ8, PQ)
wal Write-ahead log for durability
cache Query result caching
filter Metadata filter expressions
observability Metrics and tracing

Feature Flags

  • parallel (default) - Enable rayon-based parallel processing
  • salience-debug - Enable debug logging for salience distribution

Benchmarks

Run benchmarks with:

cargo bench

Available benchmarks:

  • distance - Distance computation performance
  • index - Index build/search performance
  • retrieval - Query engine latency
  • serialization - rkyv serialization throughput
  • wal - Write-ahead log performance

License

MIT