Changelog
New capabilities and performance improvements across the Mixpeek platform. Every entry links to relevant documentation and examples.
The Footage Intelligence template page now runs real searches over a public sample library of stock clips from Pexels and NASA. Pick one of the suggested queries, such as an aerial view of a coastline, or type your own, and each result plays the matching clip with its time range, credit and license. Deploying the template runs the same scene search over the footage you connect.
Build ($25 a month) and Scale ($250 a month) are flat monthly fees, and usage is billed on top: on Managed, the usage beyond the plan's included pool ($10 on Build, $100 on Scale); on MVS, all metered vector usage. The pricing page, the MVS page and the billing docs had called the fee a minimum that usage counts toward. They now match the invoice, as do the storage rate ($1.56 per million vectors per month plus $0.023 per GB-month), Scale's 20% discount on overage in the estimator, and the speech transcript add-on at $0.024 per minute.
On a video, the image and ColQwen extractors sample frames and embed each one. The frame index was dropped before the document id was built, so the frames of a clip collapsed into a single document and only one of them could come back from a search. Each frame now gets its own document with its timestamp, so a search can return the frame that matched. PDF pages were already separate documents. Videos indexed before this change keep one document per clip until they are processed again.
When a bucket's schema had required plain-value fields such as ad_id or spend, Studio's upload dialog sent them alongside the files, so an upload with every field filled in was refused with "Required field is missing". Studio now sends those fields at the top level of the object, in both file and URL mode.
When a signed-in user's session key expired, Studio showed "API key is not active (status: expired)" on every page. It now renews the key and retries the request.
A Studio link with ?ns=<namespace>, such as one to a retriever, now switches to that namespace before the page loads, on every page that works inside a namespace. It used to take effect only on the namespaces list, so a link into another namespace showed a not-found page.
The marketplace tab offered only the categories found on the current page of listings, and choosing one hid the others. The filter now lists every category in the catalog, and you can switch between them directly.
The bring-your-own object storage page listed the variables the API, coordinator and Ray workers read, and left out the ones the MVS shard reads. A self-hosted deployment that followed it started a shard with no store. The page now has a Shard settings section: mvs.objectStore for the Helm chart, or MVS_STORE_BACKEND (s3 or gcs) with the MVS_S3_* or MVS_GCS_BUCKET variables, pointed at the same bucket as the rest of the deployment. MVS can also keep its data on a POSIX filesystem such as local NVMe, NFS or Lustre. That backend is in private beta, and you turn it on for a deployment by contacting us.
POST /v1/retrievers/{retriever_id}/execute/batch runs one retriever over a list of queries in a single round trip. The list now holds at most 16 queries. A request with 17 or more is refused with 422, and error.details.limits.max_queries says 16, so a client can split the set without parsing the message. The limit had been 50, and no customer request in the previous 30 days sent more than 16.
When every job in a batch failed, the batch error read "All N jobs failed. First error: ...", which looked like one cause even when there were several. It now adds a count of the distinct error types, for example "(4 distinct causes: ...)". A batch whose rows all failed also left each object reading COMPLETED with no error. Those objects now read FAILED with the error type and message, the same as objects that fail in a partly successful batch.
A batch waiting for capacity showed "N of N concurrent batches are running" with N taken from the limit, even when fewer were running. The queue reason now counts the batches actually holding a slot. When the engine turns a batch away for capacity, the reason keeps the engine's own explanation and adds when the batch will be retried, or that it starts when another batch finishes.
A batch with more than one extraction tier holds a slot from its first tier to its last. When the last tier had nothing to extract, such as a clip with no speech, or a later tier failed early, the finished batch kept its slot and the batches queued behind it waited until something else freed one. In one run that wait was more than 25 minutes. Both paths now give the slot back and start the next queued batch.
A filter that matches a field at several paths at once, which the managed attribute filter does, made each count and scroll collect every point on the shard before narrowing. On the shared store that meant 14.1 million ids per call and scroll times around 5.4 seconds. Those groups now resolve from the index entries alone. Separately, listing a bucket's objects newest first now reads the order straight from the date index: on a 222,765-object bucket a 500-object page at offset 3,000 went from 0.31 to 0.19 seconds, and the first page from 0.08 to 0.045 seconds.
The first signup step now opens with a search that runs as the page loads and shows four live results from a public demo, with a switch between a film-scene collection and an art collection, suggested queries and a free-text box. After paying, a new account lands on the dashboard rather than on billing settings.
A default alert whose condition stayed true sent a new email and in-app notification on every check. It now notifies once, then waits for the alert's cooldown or re-notify interval, or 24 hours when neither is set, before notifying again. Every check is still recorded in the alert's history. The "no new objects" alert now carries the bucket it measured, so the notification says which bucket stopped receiving objects.
Each new account gets a default namespace, collection and retriever. If a step failed during an outage, the account stayed half set up until someone fixed it by hand. A scheduled job now retries setup every 15 minutes for accounts more than an hour old, up to five times, and records the last error on the account.
A large image batch on CPU workers had been embedding about 185 images an hour. The model ran in bfloat16 on processors with no hardware support for it, so each image took about 20 seconds, and only one of the batch's worker CPUs was ever put to work. CPU batches now embed in fp32 and start one model copy per worker CPU from the beginning. In our test at the same worker shape, 3,414 images took 20 minutes, about 10,250 an hour, where the previous code managed about 180 an hour. GPU batches are unchanged.
When the audio fingerprint extractor could not read a file, the batch failed with a message saying it returned 0 rows, followed by a list of possible causes. Each object now carries a named error: an unsupported format, a video with no audio track, or a decode failure with FFmpeg's last lines of output. A video with no audio track is not retried. When a Vertex multimodal embedding fails, the object's error now carries the provider's message, where it had read "None/empty". A finished batch also takes its extracted and failed counts from its per-object record, so a row that failed extraction is counted as failed.
When a required stage hit its time limit before any results existed, the retriever answered 200 with an empty result marked degraded, which reads the same as a search that matched nothing. In our test of 30 executions over a million documents, 5 came back that way. It now answers a retryable 503 whose error names the stage and its time limit. A stage that times out after earlier stages returned results still answers 200 with those results, flagged degraded, and optional stages set to on_error skip behave as before. The attribute_filter stage also fails when the vector store does not answer its scroll; it had returned the documents fetched before the error, usually none.
A retriever stage read from a GET, a clone or a pinned revision in a resource.executed event carries a parameters field, which the API fills in as a copy of config.parameters. Posting that stage back flagged parameters as an undeclared key, and retriever execution flagged settings, including settings.limit, which the API translates itself. Both now count as declared, so they get no warning today and no 422 after November 2. A misspelled key next to them is still reported.
Everything above is live. Point Mixpeek at the storage you already have and search your video, images, audio, and documents through one API. Build starts at $25/mo for up to 1M vectors.