PLAID

PLAID is the retrieval engine behind ColBERTv2 for multi-vector Late Interaction search. It generates candidates with the single-vector (SV) heuristic: each query token is searched individually against the document token vectors, matching documents are collected (and deduplicated), and the candidates are then scored with full MaxSim/Chamfer similarity.

---

As the Baseline for MUVERA

PLAID is the state-of-the-art baseline the MUVERA paper measures against (MUVERA - Multi-Vector Retrieval via Fixed Dimensional Encodings):

• The SV heuristic can miss true nearest neighbours and is costly and sensitive to parameter tuning.
• To match the deduplicated SV heuristic’s recall on MS MARCO, MUVERA’s fixed dimensional encodings retrieve 1.75–3.75× fewer candidates at 4096 dims and 2.6–5× fewer at 10240 dims.
• End to end, MUVERA reports 10% higher Recall@k and 90% lower latency than PLAID on average across six BEIR datasets. On MS MARCO it stays within 0.4% of PLAID rather than beating it, which the authors suggest is possibly because PLAID was carefully tuned for that dataset.

Practitioners in the Hacker News discussion framed the difference as PLAID needing specific index structures and clustering assumptions, where MUVERA reuses any single-vector MIPS index.

Related Concepts

• ColBERT — the model PLAID serves
• Late Interaction — the scoring paradigm
• MUVERA — the fixed-dimensional alternative to the SV heuristic
• Maximum Inner Product Search

Related Articles

• MUVERA - Multi-Vector Retrieval via Fixed Dimensional Encodings
• MUVERA - Transforming Multi-Vector Information Retrieval Through Fixed Dimensional Encodings
• Muvera - Making Multi-Vector Retrieval as Fast as Single-Vector Search (HN Discussion)

Companies

• LightOn AI — ships NextPlaid / FastPlaid multi-vector engines