v0.beta.2026.08.18(v0.beta.2026.08.18)
Astonishingly Fast Search
One of Poly’s key value proposition is to help you find anything in your file system, no matter how large. Today, we’re excited to announce a leap forward in that mission, with an industry-first sub-100ms Terabyte scale smart search. We’ll have more details to share in the coming days, as well as a technical blog post to dive into the internals of this system as well as its key architectural decisions that make it a best-in-class system. Here’s a preview!
Why is it hard to do Terabyte-scale smart search?
Let’s be clear about what we mean by smart search. Poly’s search system has some fairly challenging requirements: a) it must be able to search over millions of embedded (vector-based) representations of a user’s files. b) it must be able to search over millions of full-text representations of a user’s files, including text files, OCR, transcripts, and other files down to a page/paragraph/frame level granularity. c) it must be able to search in parallel over a set of up to 100 independent drives, with no clear isolation (multi-tenant), and d) it must support a vast amount of facets to narrow the search including within-folder, filtering by tag, mime-type, text query, and even custom user-defined properties. Our old, legacy search implementation would routinely slide into spinning for multiple seconds, sometimes as long as 10 seconds for a complex search over large drives. Our users would routinely complain of missing search results when in actuality they just weren’t waiting long enough for these slow searches to finish.
Our initial attempt at a solution was by no means primitive. However, even 5-year old “legacy” systems for handling this level of complexity are not up to the task. It’s for this reason that today’s data storage companies from consumer to enterprise hardly make complex, smart search widely available. When they do, it’s often non-interactive or gated behind enterprise sales. When they do ship something, it’s handicapped: no frame or page or paragraph level granularity, no hybrid text/embed search, wildly inaccurate, hidden behind some new product name or AI agent feature.
Faster than your agent can think
One of the crucial reasons why a fast search layer is a must-win for Poly is that it accelerates the frontier for both human use as well as agent use. When you want an agent to help you accomplish things, you want it to spin off dozens of smart searches at once and receive results instantly. Essentially, you want the search to return results before the model is done thinking. This creates a noticeable positive feedback cycle — the agent does more searches because it can receive broad results quickly, it prefers the search to wanton browsing of the file system, and (eventually) if we’re successful our technology will make it inside the training gyms of the large modeling companies, improving its use further.
What does this mean for you? Primarily, it means that over the next months, you should see Poly’s search become more common with models (ideally as common as web searches are today with LLMs). And most importantly, we hope that it simply brings a dose of delight to your everyday work.
Where we go from here
Our next goal with search is to bring a range of our new ideas and insights to polyembed-v3, as we continue to experiment at the boundaries of what’s possible with file systems. We have taken your feedback seriously, especially in where the searches work and where they don’t yet clear the bar. Keep bringing us this valuable feedback as we push towards our next scale of models.
🧹 Improvements
- Flashback mode now contains more information about your folder, making it easier to find out when something happened. Events for items within the folder, like file renames, edits, etc. are shown on the timeline now.
- You can now shift+drag a region of the flashback view to instantly zoom into that region
- Performance improvements under the hood reduce the number of unnecessary requests from the client to the server, especially on creating new windows.
🐞 Bug Fixes
- An issue with flashback prevented you from seeing the correct file names and details if you scrub from after to before the item was changed
- We have fixed an issue with retrying file uploads as they were not using exponential backoff, making retries worse rather than better
- We fixed an issue with the MCP server having a misconfigured root, preventing users from using it after an update
- An issue with duplicate hidden windows was fixed that caused issues like ghost playing audio or youtube videos