Source: qdrant.tech/blog/ultrabert-embed
DMIS Lab's UltraBERT-Embed-150M is a 151M-parameter dense embedding model built on the UltraBERT English encoder, and it takes inputs up to 32K tokens to produce one 768-dimension vector. Qdrant compared whole-document vectors with 512-token chunks on five LongEmbed tasks and on MLDR, a set of 200,000 English Wikipedia articles. Chunks scored higher on most tasks, and the whole-document vector won only on SummScreenFD, where the query describes an entire episode.
Storing both representations in one Qdrant point as named vectors gave small gains on SummScreenFD and Needle, but lowered scores on 2WikiMultihopQA, QMSum and MLDR. For MLDR's 1.43 million chunk vectors, 1-bit TurboQuant brought the index to about 140 MB and kept 98% of quality with rescoring. Rescoring read original vectors from disk and added 4 to 11 ms per query.
Why it matters if you build agents
Developers building retrieval for agents can index long documents as whole-document vectors, but Qdrant's results suggest chunks remain the better default for finding specific facts. Qdrant recommends starting with 512-token chunks for detail search and testing 256-token chunks for single facts. Whole-document embedding of SummScreenFD took about three times as long as chunk embedding on an NVIDIA L4, so that cost should be measured as well.
Get stories like this every weekday
One short email with the day's agent tooling news.