miniCOIL EN-ES: how to search across a language barrier without translation
Qdrant and Pento have built miniCOIL EN-ES, a sparse neural search model that works across English and Spanish simultaneously. The English word "home" and the Spanish word "casa" get written into the same vector cells, avoiding machine translation in the hot path. The system contains 2,398 concepts (translation clusters), each allotted 8 dimensions. The vocabulary was built via a graph constructed from 112k bilingual pairs in the MUSE dictionaries, with communities detected using the Louvain algorithm, then filtered by frequency of occurrence in Wikipedia. The model is trained on quintuplets — five examples per query — explicitly mixing languages. The backbone is multilingual-e5-small (384 dimensions, 118M parameters). • **MUSE dictionaries + Louvain**: 112k pairs → graph → communities → ~80k concepts → 12k after the Wikipedia filter → 2,398 for the final version (based on mMARCO dev.small) • **multilingual-e5-small**: 384-dimensional output, 118M parameters; 100% recall on parallel EN/ES pairs with cosine similarity of 0.91–0.93 • **Quintuplet loss**: five points (an anchor plus four contrasts) for explicit language mixing during training, replacing the triplets used in miniCOIL v1
AI-processed from Qdrant Articles; edited by Hamidun News
Qdrant has released miniCOIL EN-ES, a sparse retrieval model for bilingual search without machine translation. The joint project with Pento tackles a long-standing pain point in sparse retrieval: how do you search across translations when BM25 only sees exact word matches? The Spanish word "bajo" can mean low, bass, short, or under — but to a classical algorithm, it's just another token in another posting list.
The traditional fix is to translate queries or documents, but that adds a model to the hot path along with a dependency. Dense multilingual encoders work, but they don't guarantee exact-keyword search. miniCOIL EN-ES takes a third route: a sparse vector that reasons in concepts rather than words.
A vocabulary of concepts
The foundation is 2,398 concepts — clusters of translations such as {cat, gato} or {bat, murciélago, bate}. Each concept is allotted 8 dimensions within the sparse vector.
The vocabulary was built via a graph: the team started with 112k bilingual pairs from the MUSE dictionaries (English–Spanish), where each pair forms an edge between words. They ran the Louvain algorithm to detect communities within this graph. Polysemy — for example, the Spanish word "bajo" with four English translations — was pulling a hundred words into a single cluster. The solution: cap it at the top 2 translations per word, with a maximum of 20 words per cluster. That produced roughly 80k concepts.
Next, they ran English and Spanish Wikipedia through the model and counted the occurrence frequency of each concept: that left 12k. The final cut, based on real queries from mMARCO dev.small (the translated version of MS MARCO), came to 2,398 concepts, which capture 90% of all matches found by the full 12k vocabulary.
Architecture: multilingual-e5-small and quintuplet loss
The model's backbone is multilingual-e5-small (384-dimensional outputs, 118M parameters). Tested on parallel EN/ES pairs, it achieved 100% recall with cosine similarity of 0.91–0.93. Each concept gets one trainable head: a projection from 384 down to 8 dimensions.
The innovation lies in the training. miniCOIL v1 used triplets (an anchor, a positive with the same word in the same sense, and a negative with a different sense). In a bilingual setting, most triplets end up being single-language, and the model "cheats" by splitting languages into separate regions of the vector. miniCOIL EN-ES instead trains on quintuplets — five examples covering four quadrants: a match within one language, a match within the other, a mismatch within one language, and a mismatch within the other. The quintuplet loss explicitly forces a single concept to remain unified regardless of language. Words outside the vocabulary fall back to proven BM25 behavior.
The authors plan to scale up both the vocabulary and the training process, and the approach transfers readily to other language pairs.
Key Questions Answered
What are the key details on A vocabulary of concepts?
The foundation is 2,398 concepts — clusters of translations such as {cat, gato} or {bat, murciélago, bate}. Each concept is allotted 8 dimensions within the sparse vector. The vocabulary was built via a graph: the team started with 112k bilingual pairs from the MUSE dictionaries (English–Spanish), where each pair forms an edge between words. They ran the Louvain algorithm to detect communities within this graph. Polysemy — for example, the Spanish word "bajo" with four English translations — was pulling a hundred words into a single cluster. The solution: cap it at the top
What are the key details on Architecture: multilingual-e5-small and quintuplet loss?
The model's backbone is multilingual-e5-small (384-dimensional outputs, 118M parameters). Tested on parallel EN/ES pairs, it achieved 100% recall with cosine similarity of 0.91–0.93. Each concept gets one trainable head: a projection from 384 down to 8 dimensions. The innovation lies in the training. miniCOIL v1 used triplets (an anchor, a positive with the same word in the same sense, and a negative with a different sense). In a bilingual setting, most triplets end up being single-language, and the model "cheats" by splitting languages into separate regions of the vector.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.