Qdrant previews Constella models for changing query encoders without re-embedding

Qdrant has introduced Constella as a research preview for changing the model that encodes search queries without re-embedding a document collection. The model family is built around Stella, a 400M-parameter English embedding model that encodes documents. Zero, Nano, or Stella can encode queries against the same Qdrant collection. Qdrant reports results across 15 BEIR datasets.

The design puts more computation into document vectors, which can be reused across searches, and offers different levels of computation for queries. Zero uses token lookup, pooling, and normalization, without modeling word order. Nano adds a 34.5M-parameter transformer to account for relationships and positions between tokens, then maps its output to Stella’s 1024-dimensional space. Qdrant says it trained both smaller models to reproduce Stella’s query embeddings, enabling model changes while stored document vectors remain fixed.

Across the 15-dataset evaluation, average nDCG@10 was 0.4572 for Zero, 0.5081 for Nano, and 0.5614 for full Stella; the metric reflects the ranking of relevant results among the top 10. Qdrant says Nano retained about 91% of Stella’s average score, while Zero performed better than Nano on FEVER, HotpotQA, and Climate-FEVER. The company cautions that Stella reports training or evaluation exposure to four datasets, and that Zero and Nano learn from Stella, so those results are not tests on entirely unseen data. It recommends testing the tradeoff on a builder’s own workload and suggests pairing Zero with BM25 for hybrid search.

In Qdrant’s CPU measurements on an Apple M5 Pro, warm encoding of a 20-word query took 0.081 milliseconds for Zero, 3.131 milliseconds for Nano, and 38.952 milliseconds for Stella. The company measured encoding with FastEmbed and ONNX Runtime; search and network time are additional, and the figures do not include those costs. Qdrant describes possible uses including local search on low-power devices, high-volume retrieval APIs, and using Zero for each keystroke before switching to Nano when a user pauses or submits. For AI builders, the preview offers a way to evaluate query-side compute and retrieval quality separately from rebuilding stored document vectors.