miniCOIL

Definition

miniCOIL is a lightweight Learned Sparse Retrieval model from Qdrant, introduced by Evgeniya Sukhodolskaya in May 2025. Where SPLADE replaces lexical weights entirely with learned ones, miniCOIL augments the BM25 formula with a small semantic component — described by its author as standing “on the shoulders of BM25”.

The problem it targets: BM25 scores bat in “fruit bat” and “baseball bat” identically, because it sees only term statistics. Dense retrievers resolve the ambiguity but give up keyword precision.

Mechanism

Derived from COIL (Contextualized Inverted Lists), which stored per-term vectors in an inverted index rather than collapsing each term’s meaning to one number. COIL’s practical problems were specialized indexes, domain-specific training, and subword tokenization; miniCOIL works at word level instead.

Word → contextual embedding (jina-embeddings-v2-small-en, 512d)
     → linear layer + Tanh
     → 4 dimensions per word in the sparse vector

• Vocabulary: the 30,000 most common English words, stemmed and cleaned
• Per word: 4 consecutive cells, one per semantic dimension
• Scoring: the BM25 formula extended with a Meaning component comparing query and document term semantics

The Out-of-Vocabulary Fallback

miniCOIL’s distinguishing property — and the vault’s worked example of Out-of-Vocabulary as a design constraint: a word with no miniCOIL training falls back to plain BM25 scoring within the same sparse vector.

This is the direct answer to a structural limit of SPLADE, raised in Q&A at Evgeniya Sukhodolskaya - Fine-Tuning Sparse Neural Retrievers for E-Commerce — SPLADE’s output dimensions are its base model’s vocabulary, so terms outside that vocabulary simply cannot be represented, and SPLADE has no graceful degradation path. With SPLADE you must choose a base model that already knows your tokens.

Benchmarks

On BEIR datasets it had not been trained on, nDCG@10 vs BM25:

Dataset	miniCOIL	BM25
MS MARCO	0.244	0.237
NQ	0.319	0.304
Quora	0.802	0.784
FiQA-2018	0.257	0.252
HotpotQA	0.633	0.634

Wins on four of five. The gains are deliberately modest — the point is generalization without domain-specific training bias, in contrast to a fine-tuned SPLADE, which buys large in-domain gains at the cost of transfer (see Fine-Tuning Sparse Embeddings for E-Commerce Search).

vs SPLADE

	SPLADE	miniCOIL
Relationship to BM25	Replaces the weighting	Extends the formula
Term expansion	Yes (synonyms)	No
Out-of-vocabulary	Unrepresentable	Falls back to BM25
Vector size	~100–200 non-zero of 30,522	4 dims per word
Domain adaptation	Fine-tune per catalog	Designed to generalize

Related Concepts

• BM25 — the formula miniCOIL extends
• SPLADE — the alternative learned-sparse approach
• Learned Sparse Retrieval · Sparse Embeddings · Sparse Vector Retrieval
• Hybrid Search — miniCOIL is positioned as a BM25 replacement in the sparse leg
• Out-of-Vocabulary — the constraint its BM25 fallback is designed around
• Tokenization — word-level vs COIL’s subword tokenization

People

• Evgeniya Sukhodolskaya — author