November 2026 Summaries

No summary generated yet.

Qdrant 1.20 introduces turbo8 and turbo16 vector datatypes alongside per-element indexing for nested metadata fields, aiming to reduce search infrastructure costs while maintaining performance and accuracy. Based on TurboQuant compression, turbo8 and turbo16 store vectors using 8 or 16 bits per dimension, reducing storage to one-quarter or one-half of float32 respectively while retaining recall comparable to full-precision vectors and keeping server-side latency within roughly 10%. Unlike scalar or binary quantization approaches that may require rescoring against stored float32 vectors, the new types discard the originals and use asymmetric scoring to avoid disk-read delays. The release also improves filtering on arrays of objects by indexing individual nested elements, allowing Qdrant to evaluate conditions such as a single author matching both a first and last name entirely through indexes rather than reading payloads after an initial lookup. Benchmarks on one million arXiv papers indicated that per-element indexes could deliver up to four times the throughput of regular nested indexes with equivalent correct results, while flat filters may be faster but can produce false positives for object arrays.