Qdrant Tests Cohere Embed 5 for PDF Search With One Vector per Page
Qdrant says Cohere's new Embed 5 Pro model, which represents a whole PDF page image with a single dense vector, scored above several late-interaction models on five public English ViDoRe V3 corpora. The vector database company also describes a recipe that cuts per-page memory from 8 KB to 256 bytes using Matryoshka dimensions and its asymmetric quantization.
- Qdrant reports Cohere Embed 5 Pro, using one dense vector per PDF page image, scored above several late-interaction models on five public English ViDoRe V3 corpora.
- The test covered 12,969 page images and 1,489 English questions, measured with standard nDCG@10 over the full corpus.
- Combining MRL with 1-bit binary quantization cuts a full vector from 8 KB to 256 bytes, 32x smaller.
- Qdrant's asymmetric quantization keeps queries at 8 bits while page vectors stay at 1 bit, followed by reranking with full vectors.
- Qdrant cautions that Embed 5's training data is not published, so test contamination cannot be ruled out.
Dense vectors versus late interaction
Cohere released Embed 5, a family of embedding models with two members: Embed 5 Pro (embed-v5.0-pro), focused on quality, and Embed 5 Fast (embed-v5.0-fast), a lighter option. According to Qdrant, both accept text and images, support 100+ languages and share an embedding space.
For searching PDF page images, Qdrant usually advises late-interaction visual models such as ColPali and ColQwen, which avoid text parsing and extraction. They keep a vector for each small image patch, which gives fine granularity but means about a thousand vectors per page. Qdrant notes that MaxSim, the similarity used by these models, cannot directly use the HNSW index, which leaves MUVERA-like approaches at billions of points.
Known tricks reduce the load. Qdrant's PDF retrieval tutorial combines pooling with rescoring, averaging patch vectors by rows or columns to cut the number of first-stage vectors 20-30 times. MUVERA turns a list of token vectors into one fixed-length vector that approximates the multi-vector score.
Qdrant says it did not expect dense embeddings to compete, but tested them out of curiosity.
What the ViDoRe V3 test showed
Qdrant tested Embed 5 on the five public English corpora of ViDoRe V3: computer science textbooks, financial reports, EU labor reports, US Air Force technical orders and FDA pharmaceutical documents. Together they contain 12,969 page images and 1,489 English questions.
Qdrant points out a difference in method. As it understands it, Cohere's own ViDoRe V3 results used parsed text outputs, an OCR-like setting, and its launch results use RCP-nDCG. Qdrant instead fed page images directly and, out of habit, used standard nDCG over the full corpus to compare with published ViDoRe results. On mean nDCG@10 across the five corpora, Embed 5 Pro beat several late-interaction models.
Qdrant adds a caveat: the training data is not published, so it may be accidentally testing on training data. If that is not the case, a single dense vector per page would mean no OCR and ordinary HNSW search, avoiding the MaxSim limitation while keeping the page image as input.
Shrinking vectors with Matryoshka and asymmetric quantization
A full Embed 5 vector has 2048 dimensions by default, which is 8 KB per page as float32. The family supports Matryoshka representation learning (MRL), where the first N numbers of a vector keep most of its meaning. Cohere recommends 1024 Matryoshka dimensions in int8 as the production default, which is 1 KB per vector with what Cohere describes as near-full-precision quality. For the smallest footprint, Cohere advises binary vectors as a fast first pass before higher-precision re-ranking.
Qdrant combines MRL with 1-bit binary quantization, which stores only whether each dimension is above zero. For a full Embed 5 vector, memory falls from 8 KB to 256 bytes, 32x smaller. To offset the information loss, Qdrant's asymmetric quantization keeps the query at 8 bits per number while page vectors stay at 1 bit. If both sides keep only signs, every dimension counts equally, so a value barely above zero gets the same vote as a large one. An 8-bit query lets stronger dimensions carry more weight.
The workflow uses the compressed vectors to find a short list of candidates, then reranks them with the full vectors, as Cohere advises. Compressed copies stay in RAM, while full vectors stay on disk until needed for re-scoring. Qdrant also suggests experimenting with Embed 5 Fast for queries.
The recipe in practice
Qdrant published code tested with Qdrant 1.19.2 and qdrant-client 1.20.0, using Qdrant Cloud inference. The collection holds two named vectors. The "full" vector has 2048 dimensions, is stored cold and has no HNSW graph (m = 0), since it is only used to re-score candidates. The "small" vector has 512 dimensions with 1-bit binary quantization, 8-bit query encoding and pinned memory.
When uploading a page, the same image is embedded twice with cohere/embed-v5.0-pro: once in full and once with the option mrl: 512. At search time, a prefetch step queries the small vector with a limit of 20 candidates, using the example question "What drove the drop in posted workers in Germany in 2022?", followed by rescoring against the full vectors.
What it means for PDF search at scale
According to Qdrant, the result is a path to visually rich document retrieval with one vector per page, no OCR step and standard HNSW indexing. Teams that find late-interaction models hard to scale may find this approach worth testing. The results come from Qdrant's own evaluation, and the company flags the unknown training data as a possible confound.
