이미지: 트웰브랩스
Broadcast archives with nothing left but handwritten notebooks
Somewhere in a broadcaster's storage room sits a mountain of DigiBeta, DVCam, VHS, film reels, LTO tapes, and 3/4-inch tapes. What's actually on each tape is recorded only in handwritten notes on a shelf, and the only person who can make sense of them is whoever wrote them. Even after digitization, the problem doesn't go away. The bottleneck was never file format — it was metadata, meaning whether a system understood not just where a file lives, but what's actually happening inside it.
Mimir is a serverless media production and collaboration platform that serves as the system of record where content is stored, organized, and moved through production workflows. Twelve Labs acts as the intelligence layer that understands what's inside that content, indexing every scene, utterance, object, and action to make it searchable. Combining the two turns an archive from mere storage into something you can find, understand, and reuse.
Semantic search that works without tags
Mimir's existing search pulls results from clip titles, transcripts, and facial recognition data. Searching "Macron" and layering conditions like "wearing a tie" or "says the word monopoly" can narrow results down to one specific moment. It's powerful, but only works if someone has already tagged things in advance.
Semantic search, powered by Twelve Labs' multimodal media embedding model Marengo, works differently. Type in "stairs," "people running," "voting," "people eating," or "helicopter," and results appear even without a single human-created tag. Marengo reads intent rather than keywords, routing each query to whichever signal — visual, transcript, or audio — best fits. Compound queries also work as a single search, such as "a goal shot from left to right, player in a white shirt, against a team in yellow, in the rain." The two approaches can also be combined: for large archives, users can first narrow the field with Mimir filters, then run semantic search to cut down processing time.
Automatic scene-level compliance logging
Regulators such as the UK's Ofcom impose strict penalties for broadcasting harmful content, but manually watching a growing archive from start to finish to screen for it isn't realistic. Twelve Labs' video-language model Pegasus breaks content into scene-level segments and automatically flags violence, gore, criminal acts, and drug- or alcohol-related content against predefined compliance categories. Each flag comes with a timestamp, a severity rating, and the reason it was flagged, letting reviewers jump straight to the relevant frame.
The same scene-segmentation capability isn't limited to compliance. It's also used to locate "previously on" recaps at the start of an episode or end credits at the close, and multiple segmentation passes can be layered on top of one another.
Metadata generation at archive scale
When processing a single clip, Pegasus extracts both an asset-level summary and scene-by-scene detail — people, objects, and taxonomy-aligned metadata — in a single pass. It can process up to two hours of video at a time. The resulting metadata is stored in Mimir immediately and can be found through ordinary keyword search without needing to switch over to semantic search.
On average, indexing and metadata generation for an hour of content takes 3 to 4 minutes, assuming the proxy is already online in Mimir. Actual completion time varies depending on the source format — formats that can be simply rewrapped process faster than those requiring full transcoding. The resulting metadata belongs to the customer and is stored in Mimir's own database; Twelve Labs does not retain the data it generates.
In live news workflows, there's no need to wait for recording to finish. Because Mimir generates proxy segments as capture happens, editors can begin reviewing footage and building clips within seconds of recording starting.
From natural-language query to editing timeline
Searching "soccer goal" returns ranked results across the entire archive. To narrow it down to a specific player, uploading a single image of that player is enough — Twelve Labs registers the player as a recognizable entity and matches them across the archive without relying on names or tags. Requesting goal-celebration clips featuring that player returns results with the relevant segments highlighted directly on each clip's timeline.
Selected clips can be summarized in natural language to create clip descriptions, or sent straight to Mimir Cutter. Clips arriving in Cutter land as an assembled timeline rather than a bundle of separate files. Since newsrooms often need to edit match highlights before the broadcast even ends, being able to request celebration clips by name or image and get back an immediately trimmable timeline makes a real practical difference.
Integration architecture and current limitations
The two platforms connect through an orchestration layer. When new content arrives in Mimir, it triggers an event; an orchestration layer built with tools like EventBridge, Step Functions, and Lambda directs Twelve Labs to perform indexing and sends the results back to Mimir. Indexing runs on proxy files rather than full-resolution originals and is agnostic to storage type. Whether the files live on Amazon S3, GCP, on-premises, or a local NAS, Mimir renders the media in place, Twelve Labs pulls the lightweight proxy to generate embeddings, and the proxy is then discarded.
Indexing can also be configured to run only on-demand rather than through automatic ingestion. For person identification, Twelve Labs maintains an entity library covering people, objects, brands, and logos, and work is currently underway to connect this directly with customers' own facial recognition data (Mimir's AWS-based face recognition). Specific commercial terms — pricing, billing structure, supported countries — are not disclosed in the source. According to the source, prospective adopters need to discuss details individually with the Twelve Labs sales team.
