Qdrant compressed a PDF page to 64 bytes in memory and kept 99.2% of search quality
PDFs full of tables and charts are usually searched either through OCR, which can lose the links between elements on a page, or through ColPali, which stores about a thousand vectors per page. Qdrant found that Cohere Embed 5 Pro needs only one vector per page image. In memory, each page takes 64 bytes instead of 8 KB, and the top 20 pages are re-checked with full vectors read from disk. Search quality reaches 99.2% of the full-vector baseline. For large PDF archives, this means skipping OCR without inflating RAM usage.
AI-processed from Qdrant Blog; edited by Hamidun News
Imagine an archive of 10 million PDF pages filled with reports, manuals, and tables, and you need to find answers to questions across all of it. If you store each page as a vector (a set of numbers the database uses to compare meaning), the memory needed for search quickly becomes the largest cost. The Qdrant team has published a recipe that fits a page into 64 bytes of RAM instead of 8 KB, while keeping 99.
2% of the search quality of the full vector. Here, 100% is the average nDCG@10 score of 65.85.
This metric shows how high the relevant pages rank in the results. For comparison, 8 KB is a vector of 2048 float32 numbers, and 64 bytes is the same vector compressed 128 times.
Why one vector per page at all
Answers in PDFs are often hidden in charts, tables, and diagrams. The classic approach is OCR, which converts an image into text. It can lose the connections between words, for example which number belongs to which table column.
For search without OCR, Qdrant usually recommends late-interaction models such as ColPali and ColQwen. They store a separate vector for each small patch of the image, and each query word is compared with its best-matching patch. This is accurate, but it means about a thousand vectors per page. The MaxSim similarity function works poorly with a standard HNSW index, so at billions of points teams have to rely on workarounds such as MUVERA.
The ideal is one regular vector per page. The engineers did not expect this approach to compete, but out of curiosity they tested it on the Cohere Embed 5 models. The family has two models: the higher-quality Pro and the lightweight Fast. Both accept text and images, support more than 100 languages, and share a single vector space. Cohere itself evaluated them on the ViDoRe V3 benchmark using recognized text, while Qdrant fed Embed 5 Pro page images directly.
The test ran on five English-language ViDoRe V3 datasets: computer science textbooks, financial reports, EU labor market documents, US Air Force technical manuals, and FDA papers. In total there were 12,969 pages and 1,489 questions. Embed 5 Pro, using one vector per page, outscored several late-interaction models.
How a page shrinks to 64 bytes
The recipe combines three techniques:
- Matryoshka representation learning (MRL). The first N numbers of the vector carry the core meaning, so you can keep the first 512 of 2048.
- Binary quantization. Each number is reduced to one bit: whether it is greater than zero or not. 512 bits amount to 64 bytes per page.
- Asymmetric quantization. Pages are stored as single bits, while the query stays at 8 bits. If both the page and the query are reduced to signs, every dimension votes equally, even when its value is just above zero. With a precise query, the strong dimensions carry more weight.
The shorter the stored vector, the more the precise query helped. After a coarse search, the top 20 candidates are re-scored with full vectors stored on disk, while only the compact copies stay in RAM. The result is 99.2% of the full-vector quality and 128 times less memory.
In a estimate for 10 million pages, uncompressed ColPali vectors take about 4.8 TB, full Embed 5 vectors take about 76 GB, and the search vectors under this recipe take about 610 MB in RAM. Another roughly 95 GB of full and truncated vectors sit on disk.
Caveats and a bonus with Rerank 5
These numbers should be read carefully. Embed 5's training data has not been published, so it cannot be ruled out that the model has already seen these corpora. The tests were run in English, and five datasets are not your documents. The memory calculation counts only vector bytes, without the index or metadata. Reading from disk slows search, and the vectors read over time occupy free RAM in the operating system's cache. Finally, these are measurements by a company that sells a vector database.
At the end, Qdrant also tested Cohere Rerank 5, which was released while the post was being prepared. It works with text rather than images, so it is useful for teams already using OCR. The ViDoRe V3 texts were indexed with Embed 5 Pro, and Rerank 5 Pro reordered the top 20 pages. Average nDCG@10 rose from 66.8 to 73.2, and with the 64-byte scheme it rose from 66.2 to 71.7.
The practical takeaway: if your archive consists of reports, manuals, and presentations with complex layouts, run Embed 5 Pro on a small sample of your own PDFs and compare it with your current OCR pipeline. If the quality is close, the RAM savings will be significant at scale. Qdrant Cloud also accepts a page image as base64 and calls Cohere itself, so no separate embedding step is needed in your code.
- Nov 21, 2026Opus rewrote llama.cpp code: Qwen runs 40% faster on a CPU alone
- Nov 21, 2026MiniMax H3 Image in ComfyUI: Three Fixes Remove the "Crunchy" Skin Without Changing the Model
- Nov 21, 2026llama.cpp 1.0: Stable Interface, 35 Functions Removed, and a Cache for Large MoE Models
- Nov 21, 2026Gemma 4 on a mini PC writes 63 tokens per second, Qwen3.8 only 12.8: a test of 14 LLMs
Key Questions Answered
Why one vector per page at all?
Answers in PDFs are often hidden in charts, tables, and diagrams. The classic approach is OCR, which converts an image into text. It can lose the connections between words, for example which number belongs to which table column. For search without OCR, Qdrant usually recommends late-interaction models such as ColPali and ColQwen. They store a separate vector for each small patch of the image, and each query word is compared with its best-matching patch.
How a page shrinks to 64 bytes?
The recipe combines three techniques: Matryoshka representation learning (MRL). The first N numbers of the vector carry the core meaning, so you can keep the first 512 of 2048. Binary quantization. Each number is reduced to one bit: whether it is greater than zero or not. 512 bits amount to 64 bytes per page. Asymmetric quantization. Pages are stored as single bits, while the query stays at 8 bits.
What are the key details on Caveats and a bonus with Rerank 5?
These numbers should be read carefully. Embed 5's training data has not been published, so it cannot be ruled out that the model has already seen these corpora. The tests were run in English, and five datasets are not your documents. The memory calculation counts only vector bytes, without the index or metadata. Reading from disk slows search, and the vectors read over time occupy free RAM in the operating system's cache. Finally, these are measurements by a company that sells a vector database.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.