Qdrant has released a research preview of Constella, a family of query models designed to allow developers to swap encoders without re-embedding document collections. The models, including Zero and Nano, aim to reduce compute costs and latency for vector search applications by offering varying levels of model complexity.
The ability to swap query encoders without re-indexing entire document collections addresses a major bottleneck in scaling vector search for streaming discovery and recommendation engines. By decoupling query complexity from document storage, engineers can deploy lightweight models like Zero for search-as-you-type features while reserving high-parameter models for final retrieval. This flexibility reduces the compute bill for high-volume APIs and enables sophisticated search on low-power edge devices. As vector databases become central to AI-driven video navigation, watch for whether this asymmetric encoder approach becomes a standard for balancing retrieval accuracy against real-time performance requirements.
Qdrant has been building out its vector database platform with complementary tools that position Constella within a broader retrieval stack. The company's FastEmbed library, which generates the document embeddings that Constella's query models operate against, supports dense, sparse, multi-vector, and image embedding modes alongside cross-encoder rerankers, giving developers a unified pipeline from ingestion through retrieval. FastEmbed's default text model uses Flag Embedding from the MTEB leaderboard, and the library runs on ONNX Runtime without requiring a GPU, making it suitable for serverless environments like AWS Lambda where Constella's lightweight Zero model could handle automated scene detection workflows at the edge.
Qdrant has launched the Constella research preview, allowing developers to swap query encoders without re-indexing document collections. By decoupling query complexity from storage, the new Zero and Nano models significantly reduce latency and compute costs. This flexibility enables high-performance vector search for streaming discovery and recommendation engines on low-power edge devices.
Constella allows developers to switch query encoders per request without needing to re-embed entire document collections, which reduces compute costs and latency.
The Constella Zero model reduces query latency to 0.081 ms on an Apple M5 Pro, making it approximately 480 times faster than the full 400M-parameter Stella model.
Yes, the Nano model retains 91% of the Stella model's average retrieval score across 15 BEIR datasets while utilizing only 34.5M parameters.
Developers can switch between Zero, Nano, and Stella encoders by simply changing the model name within the FastEmbed configuration.
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source