Multi‑Vector search advances explained
A new Weaviate podcast episode dives into Multi‑Vector Search techniques (ColGrep, MUVERA, PLAID, ColBERT‑Zero) that blend keyword granularity with semantic retrieval for code, reasoning and multimodal data — not just one‑size‑fits‑all vectors. The discussion frames multi‑vector approaches as key to better precision in developer and reasoning use cases. (x.com) (x.com)
Rajesh Jayaram (Google Research) and Roberto Esposito (Weaviate) were the guests on Weaviate Podcast #123, a long-form episode released alongside Weaviate’s June 2025 coverage of MUVERA. (youtube.com) Weaviate added native support for MUVERA encodings in its v1.31 release announced June 5, 2025, and published a technical blog explaining how MUVERA’s fixed-dimensional encodings (FDEs) fit into Weaviate’s multi-vector pipeline. (newsletter.weaviate.io) MUVERA converts variable-length multi-vector representations into single fixed-dimensional encodings so retrieval can reuse optimized MIPS libraries, and the paper reports FDEs retrieve roughly 2–5× fewer candidates while matching prior heuristics’ recall. (research.google) ColBERT‑Zero, released with code and checkpoints under Apache‑2.0, is a fully multi-vector pre‑trained ColBERT model that reports 55.43 nDCG@10 on the BEIR benchmark and outperforms GTE‑ModernColBERT for models under ~150M parameters. (huggingface.co) PLAID’s late‑interaction engine reports reductions in late‑interaction search latency of up to 7× on GPU and 45× on CPU versus vanilla ColBERTv2, yielding GPU latencies in the tens of milliseconds at scales up to 140 million passages in the paper’s experiments. (arxiv.org) LightOn’s FastPlaid positions itself as a production multi‑vector engine: FastPlaid v1.10.0 added incrementally‑updatable indexes and claims a 6.5× faster index update speed versus Stanford PLAID and a +554% increase in QPS for multi‑vector workloads. (lighton.ai) ColGREP/NextPlaid targets local, code‑centric multi‑vector search with a design that stores ~300 token vectors per document at 128 dimensions for code units and runs as a single Rust binary for local, incremental indexing. (lightonai.github.io)
