PulseWire

Journalists needed to find archival footage to contextualize breaking stories, but the existing keyword search over manually tagged clips returned relevant results for only 35% of queries. Tagging backlogs meant that 800,000 clips had minimal or no metadata. Newsrooms were spending an average of 45 minutes per story searching for supporting footage, causing missed deadlines and stale reporting.

Mixpeek's pipeline extracts visual, audio, and textual features from every video clip, creating a unified multimodal index. Journalists search by describing a scene in natural language, uploading a reference frame, or combining both. The retriever ranks results using scene-level embeddings that capture composition, subjects, locations, and mood, far beyond what keyword tags can express.

The 1.8 million clip backlog was processed in a two-week batch run. New content is indexed in near-real-time through a collection trigger that fires on bucket upload. The search API was exposed via a custom plugin in PulseWire's editorial CMS, so journalists never leave their existing workflow. Latency targets of sub-500ms were achieved with Qdrant-backed vector search.