TechPrufer
Get started
Weaviate
Weaviate logoWeaviate · Feature trail

HFresh: Memory-Efficient Vector Search

2 posts · Mar 2026 – Sep 2026

What this trail covers

HFresh is Weaviate's disk-based vector index designed for memory-efficient applications and large datasets. It utilizes a two-stage search strategy: an in-memory HNSW centroid index routes queries to relevant disk-based postings, which contain compressed vectors. Background operations like splitting, merging, and reassigning postings ensure index freshness without full rebuilds. In v1.39, HFresh reaches technical preview with RQ-8 for centroid index and RQ-1 for postings, supporting cosine and l. This release introduces 4-bit Rotational Quantization (RQ4) with centering and outlier storage, significantly improving memory efficiency (45% heap reduction for a 1536-dim dataset) and import times (37% faster for RQ4). Encoding and distance kernels are optimized with SIMD support, and prefetching is improved for better memory access. Recall remains stable across datasets from 1M to 250M vectors.

All Weaviate trails
  1. 4-bit Rotational Quantization | WeaviateIntroduced 4-bit Rotational Quantization (RQ4) with centering and outlier storage, achieving similar recall to 8-bit RQ but with a 45% heap reduction on a 1536-dimension vector dataset. Optimized encoding and distance kernels using SIMD (FWHT, UDOT, VPDPBUSD) for significant speedups (up to 6.1x for FWHT, 3.8x for RQ8 encoding). Improved memory access with prefetching for AMD64 and ARM64, contributing 7-11% query throughput and 12% faster imports. Demonstrated stable recall across datasets from 1M to 250M vectors.
  2. HFresh: Memory-Efficient Vector Search | WeaviateThis post details the implementation of HFresh, Weaviate's disk-based vector index. It explains the two-stage search strategy using an in-memory HNSW centroid index (with RQ8 quantization for 4x memory reduction) to route queries to disk-based postings (using RQ1 quantization for up to 32x data reduction). It also elaborates on the background maintenance operations (splitting, merging, reassigning postings) that ensure index freshness and balance without full rebuilds, and the rationale for using HNSW as the centroid index.
  3. How to extract meaning from charts and tables in PDFs | WeaviateThis post introduces a new capability for ingesting and querying PDFs using late-interaction multi-vector models. This enables direct search of visual elements like charts and tables without OCR. The process involves rendering PDF pages to images, storing them as BLOBs, and vectorizing them with the multi2multivec-weaviate module. It highlights drag-and-drop ingestion in Weaviate Cloud and demonstrates complex agentic reasoning with the Weaviate Query Agent for synthesized answers with page-image citations. The ingestion pipeline is also provided in Python for production deployment. This extends the HFresh capability by enabling search on visual data within documents, complementing existing text-based vector search.
  4. Weaviate 1.38 Release | WeaviateThis post announces the general availability of the HFresh disk-based vector index in Weaviate v1.38. It details how HFresh is enabled via configuration, its support for RQ-1 quantization and specific distance metrics (cosine, l2-squared), and its incremental rebalancing mechanism for continuous updates. The post also highlights that HFresh is now a production-ready feature, moving beyond its technical preview status.
  5. Your LLM Is Only as Good as What It Retrieves | WeaviateThis post details the critical role of retrieval quality in RAG systems, emphasizing that poor retrieval is the primary driver of LLM hallucinations. It outlines five common retrieval failure modes (drift, truncation, stale index poisoning, low-relevance top-k, inter-agent miscommunication) and argues that scaling the LLM does not solve retrieval problems. The post proposes four dimensions for improving retrieval quality: embedding model selection, chunking architecture, retrieval strategy (hybrid search, cross-encoder re-ranking, relevance thresholding), and index maintenance for freshness. It also introduces a practical measurement framework for evaluating retrieval quality, including context precision and recall.
  7. Multimodal Embeddings and RAG: A Practical Guide | WeaviateThis post details the integration of multimodal embeddings (specifically Gemini Embedding 2) with Weaviate's HFresh vector index. It explains how to configure Weaviate to ingest and embed data across text, image, audio, and video modalities using the `multi2vec-google` module. The post highlights the use of Matryoshka Representation Learning (MRL) for dimension reduction to manage storage efficiently and discusses chunking strategies for non-text data. It provides practical examples of building multimodal retrieval systems, including searching audio without transcripts, semantic search for creative assets, and multimodal RAG for document analysis, demonstrating how these capabilities are enabled by the underlying HFresh architecture.
  8. Weaviate 1.36 Release | WeaviateIntroduces HFresh as a technical preview, a new disk-based vector index inspired by the SPFresh algorithm. HFresh divides vectors into disk-based postings and uses an in-memory HNSW index over centroids for routing. Key capabilities include incremental rebalancing for freshness, HNSW centroid index, and Rotational Quantization (RQ-8 for centroids, RQ-1 for postings). Tuning parameters are searchProbe, replicas, and maxPostingSizeKB. Supports cosine and l2-squared distance metrics. Suitable for large datasets, cost-sensitive deployments, and write-heavy workloads.