Voyage AI × MongoDB Atlas
Search what video shows and says.
Each scene is already stored as one multimodal vector. Ask in plain language and Atlas lands on the exact second, across pixels and speech together.
Pre-indexed public catalog
No uploads or links
No video processing at runtime
Public video catalog
Open NASA footage hosted by Wikimedia CommonsSearch inside the selected video
The selected public video will play here.
Search results include the matched sentence, timestamp, relevance, and live Voyage usage.
Prepared once.
Searched live.
The expensive content pipeline runs before deployment. Public requests cannot mutate the catalog. At runtime, only the text query is embedded and reranked before Atlas returns stored scenes.
OFFLINE 01Scene detectionShot boundaries create focused retrieval units.
OFFLINE 02Speech + keyframeTimestamped transcript and one representative frame.
OFFLINE 03Voyage embedPicture and speech become one multimodal vector.
LIVE 04Atlas search
$vectorSearch filters to the chosen video.LIVE 05Voyage rerankThe strongest sentence supplies the exact second.
Voyage AI understands the scene
voyage-multimodal-3.5 places keyframes and transcripts in one semantic space. rerank-2.5 then scores candidate sentences against the question.
MongoDB keeps vectors with the data
Each Atlas document holds the vector, timestamps, transcript, and source metadata. Filtering and nearest-neighbor retrieval happen in one query.