Releases·Tool Release
Qdrant's Constella lets RAG pipelines swap query encoders per request without re-embedding documents
Qdrant published a research preview of an approach letting RAG pipelines swap query encoders — Zero, Nano, or full Stella — per request without re-embedding documents. The team trained two smaller query encoders to search vectors produced by Stella's 400M document encoder. Nano nearly matched Stella's retrieval quality while encoding queries much faster, and the models output standard 1,024-dimension vectors usable with any cosine-similarity vector search engine.
12 engagement1 sourceTue, Sep 29, 2026, 01:16 AM
AI Overview
- Qdrant embedded documents with Stella's 400M encoder and trained smaller Zero and Nano query encoders to search the same vectors
- Nano approached Stella's retrieval quality while encoding queries much faster; Zero uses no query-time compute
- Constella outputs standard 1,024-dimension vectors usable with any cosine-similarity vector search engine, not just Qdrant
Related
Products·launched
EmbedFlow: zero-downtime embedding model migration without re-embedding the corpus
A tool that lets teams upgrade embedding models without re-embedding their entire vector store upfront. Instead, the old index keeps serving retrieval while the new model reranks top-K candidates, caching new embeddings incrementally. It targets teams with large RAG corpora (tens of millions to a billion vectors) facing days-to-months backfill costs.
A tool that lets teams upgrade embedding models without re-embedding their entire vector store upfront. Instead, the old index keeps serving retrieval while the new model reranks top-K candidates, caching new embeddings incrementally. It targets teams with large RAG corpora (tens of millions to a billion vectors) facing days-to-months backfill costs.
2106 engagement3 sourcesreddit, hackernewsTue, Sep 8, 2026, 02:08 AM
Social·Community
Stop Renting Your AI's Memory — Dylan Couzon, Qdrant
Dylan Couzon, developer relations engineer at Qdrant, gave a talk at AI Engineer World's Fair 2026 arguing practitioners should own their AI's memory rather than keep it in vendor clouds. He frames model as CPU, context as RAM, and memory as disk, and argues retrieval beats stuffing context. A live fully offline demo shows a drone building searchable memory with sub-millisecond queries in a 15 MB footprint using Qdrant Edge.
Dylan Couzon, developer relations engineer at Qdrant, gave a talk at AI Engineer World's Fair 2026 arguing practitioners should own their AI's memory rather than keep it in vendor clouds. He frames model as CPU, context as RAM, and memory as disk, and argues retrieval beats stuffing context. A live fully offline demo shows a drone building searchable memory with sub-millisecond queries in a 15 MB footprint using Qdrant Edge.
71 engagement1 sourceyoutubeFri, Oct 2, 2026, 07:00 AM
Products·building
Fine-tuned retrieval head for RAG that learns which embedding dimensions matter
A neural network retrieval head that transforms query embeddings before retrieval, enabling the system to weight embedding dimensions according to their predictive value for a specific corpus. This improves hit rate, completeness, and faithfulness in RAG systems by reducing noise from irrelevant dimensions.
A neural network retrieval head that transforms query embeddings before retrieval, enabling the system to weight embedding dimensions according to their predictive value for a specific corpus. This improves hit rate, completeness, and faithfulness in RAG systems by reducing noise from irrelevant dimensions.
4 engagement1 sourceredditSat, Jul 11, 2026, 12:08 PM
Social·Community
Developer reports re-ranking pass beats embedding model swaps for RAG accuracy
A developer reports that adding a re-ranking pass improved RAG retrieval accuracy more than switching embedding models. The second-stage cross-encoder re-scores top results using full query context, catching relevant documents that basic vector similarity misses.
A developer reports that adding a re-ranking pass improved RAG retrieval accuracy more than switching embedding models. The second-stage cross-encoder re-scores top results using full query context, catching relevant documents that basic vector similarity misses.
3 engagement1 sourceredditMon, Aug 3, 2026, 07:01 PM
Developments·Paper
MRVQ: One Resident Index for Dimension- and Rate-Elastic Vector Search
An arXiv paper introduces Matryoshka Residual Vector Quantization (MRVQ), a post-hoc residual quantizer for frozen embeddings whose maximum-rate code truncates both by residual stages (rate) and by embedding coordinates (dimension). One resident index thus serves every evaluated (dimension, rate) pair, avoiding multiple per-rate quantizer states in the retrieval tier. On FiQA and NFCorpus across four embedding families and 4/8/16-byte codes, MRVQ is the lowest-RAM design tested, though per-rate QINCo2 retains a quality edge.
An arXiv paper introduces Matryoshka Residual Vector Quantization (MRVQ), a post-hoc residual quantizer for frozen embeddings whose maximum-rate code truncates both by residual stages (rate) and by embedding coordinates (dimension). One resident index thus serves every evaluated (dimension, rate) pair, avoiding multiple per-rate quantizer states in the retrieval tier. On FiQA and NFCorpus across four embedding families and 4/8/16-byte codes, MRVQ is the lowest-RAM design tested, though per-rate QINCo2 retains a quality edge.
0 engagement1 sourcearxivFri, Oct 2, 2026, 05:33 PM