Bedrock's managed video search embeds your footage in 512 dimensions every four seconds

Sync is where the work happens. Managed MKB extracts frames, transcribes the audio, generates a Marengo Embed 3.0 embedding for each segment, and writes the vectors to the index ^[6]. Query text and video segments share one 512-dimensional space ^[3]. That is what makes a natural language string comparable with a frame.

The setting that defines a segment sits in the Audio/video segmentation configuration section under Advanced configurations, and it defaults to 4 seconds for both audio and video ^[8]. That default is your retrieval floor. Results come back with a chunk start time and a chunk end time ^[10], and the narrowest window you can get is the one you set before the first ingest.

The walkthrough runs on a 10-minute clip of the 2022 FIFA World Cup final ^[9]. Ten minutes is 600 seconds, so at the default that is 150 segments per modality ^[1]. Audio and video are segmented separately at the same duration, putting the clip at roughly 300 embeddings of 512 dimensions each ^[2]. The need the post describes is finding specific moments in hours of footage ^[15]. One hour of video is 900 segments per modality at the same setting ^[3].

In the console, Amazon Titan Text is the selected embeddings model, and Marengo is behind the pen icon in the Additional Configuration pane ^[7]. Accept the defaults on Create Managed KB over a bucket of MP4 files and you have built yourself a text embedding index. Marengo also has to be enabled for the account in a supported Region before it is selectable at all ^[13].

The post carries no retrieval accuracy numbers and no pricing ^[16]. What it shows is the console Test feature, a query of "show me the penalty kicks from this soccer match" ^[11], and the finding that the top results identify moments where penalty kicks were attempted ^[12]. For that to say anything about your archive, your queries would have to name events as visually distinct as a penalty kick, and your audio track would have to carry commentary as descriptive as a broadcast feed. Fixed-camera security footage with no speech is the harder case, and it is not the case demonstrated.

AWS describes the alternative as stitching together transcription services, frame extraction pipelines, embedding models, vector databases, and synchronization logic ^[4]. Deleting that plumbing moves the coupling into the knowledge base, and the cleanup steps show where it lands: you note the IAM service role and the ARN of the S3 Vector index the knowledge base created, then delete both ^[14]. A later change of embedding model means re-running ingest across the whole bucket, because a Marengo vector is comparable only with other Marengo vectors ^[3].