Qdrant 1.20 adds TurboQuant datatypes and per-element nested indexing

Qdrant 1.20 introduces two quantization-native datatypes, turbo8 and turbo16, which store each vector dimension in 8 or 16 bits. According to Qdrant, a collection using turbo8 takes a quarter of the space of float32, and turbo16 takes half. The datatypes use the TurboQuant rotation-based compression algorithm and have no rescoring step, so they do not read float32 originals from disk. Qdrant says turbo8 and turbo16 stay within 10% of float32 server-side latency.

Queries are rotated but kept at float32 precision, which keeps the query overhead small. Per-element indexing for nested fields lets Qdrant check a nested filter with index lookups alone, skipping the step of reading payloads. On 1 million arXiv papers, Qdrant reports that the per-element index served up to four times as many queries per second as regular indexes, with the same correct results. Flat filters are close in speed but miss between a fifth and a third of the right papers.

Turbo8 and turbo16 are selected when a collection is created, and nested is set on payload indexes behind nested filters. Turbo4, introduced in version 1.19, saves more storage but gives up a few points of recall on some datasets.

Teams short on RAM or disk can fit more vectors on existing nodes with turbo8 or turbo16 and keep recall close to float32. Adding nested to payload indexes can speed up nested filters without changing search requests. Flat filters should not be used on arrays of objects, because they return false positives.