Bedrock Knowledge Bases gains multimodal video search via TwelveLabs Marengo 3.0

Per AWS's announcement and ML blog, Marengo 3.0 brings true multimodal embeddings to Bedrock Managed Knowledge Bases, letting developers build natural-language search over video, audio and image content rather than relying on transcripts. The key distinction is that Marengo encodes visual scenes directly — so a query like 'find the clip where a person opens a red door' can match on visual content, not just spoken words captured by ASR.

Mechanically, media stored in S3 is embedded via Marengo and indexed in the managed knowledge base, unifying multimodal retrieval under the same Bedrock RAG plumbing developers already use for text. This makes it the first video-understanding model listed in the service, extending RAG beyond documents into rich media libraries — useful for media/entertainment archives, surveillance, e-learning and content moderation.

It's part of AWS's broader week of Bedrock and inference updates (prefix-aware routing, HyperPod caching, AgentCore MCP Apps and Evaluations, DevOps Agent). Separately, AWS and Qualcomm disclosed a deepening chip relationship — Qualcomm designing custom AI inference chips for AWS across multiple generations while using Bedrock to design them, with optical interconnects up to 1.6 terabits — signaling AWS's push to diversify inference silicon beyond NVIDIA.

Competitively, native video embeddings put Bedrock ahead of some rival managed RAG offerings on multimodal breadth, and the TwelveLabs partnership gives AWS a specialist model rather than a general-purpose compromise. Caveats: multimodal embedding quality varies by domain, video indexing is compute-heavy (cost), and retrieval precision on nuanced visual queries remains to be proven at scale. Watch pricing, latency on large video corpora, and how it compares to Google's multimodal retrieval on Vertex.