Mubadala Investment Company has used PULSAR since March 2026 to search pitch decks, board packs and other deal documents by their page images. The system has handled 78,000 documents—about 2.4 million pages—across more than 3,000 deals, according to a paper published on arXiv and in the 2026 EMNLP Industry Track.
The paper reports a median vector-search time of 156 milliseconds in production. On the ViDoRe V3 benchmark, PULSAR’s pooled index cut median search latency by 15.1 times compared with an unpooled version, with less than 0.01 absolute loss in two ranking measures, NDCG@10 and Recall@10. Those measures assess how useful the top 10 search results are. Under concurrent load, the pooled index handled about 88 times more queries per second than the unpooled version. These are results reported by the paper’s authors; the supplied information does not include independent validation.
How does PULSAR search pages?
Many document search systems first use optical character recognition, or OCR, to turn text in page images into searchable words. They may also generate descriptions of charts and other figures. The paper says this approach can be expensive to refresh as documents change and can lose chart details.
PULSAR instead indexes page images using a frozen, ColPali-style vision model. It first searches compact summaries of pages, then checks promising results against a more detailed representation. This two-step method is called late interaction: the system compares the query with page details after an initial shortlist, rather than relying on a single summary for the final match.
What did the paper find in production?
The paper estimates that PULSAR’s event-driven ingestion costs about 20 times less per page than the OCR-and-figure-description system it replaced. At the production search cutoff, the authors also report that it more than doubled answer-fact recall over that earlier approach. In a retrieval-augmented generation system, recall measures whether the pages containing needed facts make it into the results available to the answer-generating model.
The results describe one system at one investment company, not a general test of document search products. The paper does not establish that PULSAR is available to other organizations.
Why it matters
For Mubadala staff searching deal materials, PULSAR could make it quicker and cheaper to find useful pages in large collections that change frequently, while retaining details in charts and other visuals. The reported deployment is at Mubadala; whether other organizations can use the system is not established.
