LoginSign Up
← Back to all sparks
D

ducksemantics

DEVOPS
Velocity3.8

DuckDB-Native Semantic and Retrieval Relations

github.com/RGenomicsETL/ducksemantics ↗

ducksemantics puts ontology grounding and late-interaction retrieval inside DuckDB, from R

duckdbrretrievalontologiesembeddingsbioinformatics
◆Current state
ducksemantics has exactly one release — a 0.1.0 debut that arrived with an unusually complete validation record. It provides a DuckDB-native semantic graph plus lexical grounding, dense retrieval, exact ColBERT late-interaction scoring, structured judgment and benchmark APIs for R. The release notes report the full HPO 2026-06-23 ontology loaded and queried: 19,836 active nodes, 50,029 aliases, and 182,884 128-dimensional ColBERT token vectors alongside EmbeddingGemma dense vectors.
◆Where it's heading
With a single release there is no arc to read yet, but the shape of the first one is informative. Stored DuckDB MaxSim was checked against a reference implementation to an absolute delta of 0.0000021, embedding caches are described as atomic, content-aware and resumable, and graph writes as transactional and idempotent — this is a package built for reproducible pipelines rather than interactive exploration. The dependency note is the constraint to watch: optional components ship through the project's R-universe rather than CRAN.
◆Prediction
The release notes name HPO as the loaded ontology and describe generic OBO parsing, so the immediate direction is likely more OBO ontologies validated at the same scale rather than new retrieval methods. Beyond that, one release is not enough to predict cadence.

◆Recent moves

  1. 25d ago

    ducksemantics 0.1.0 brings ColBERT retrieval into DuckDB for R

    ⚡ SPARK

    A first release that arrives fully formed: semantic graph, lexical grounding, dense retrieval and exact late-interaction scoring, all DuckDB-native, validated against the complete HPO ontology and a reference MaxSim implementation.

    View source ↗