How Social Department built culture-aware video discovery
Social Department built QuickSearch, a set of single‑action retrieval pipelines (Shareable Scenes, Quotable Quotes, Memeable Moments, Trend Align) that let marketers find culturally resonant video clips without typing detailed queries. The pipelines run on Amazon Bedrock using TwelveLabs Marengo 3.0 and Pegasus 1.2 for scene‑level video reasoning and Anthropic’s Claude models, with retrieval backed by Amazon OpenSearch Serverless. Responsible‑AI safeguards include Bedrock Guardrails for prompt‑attack screening and post‑generation checks against episode transcripts before human approval. This architecture lets entertainment marketers scale editorial‑driven discovery while keeping proprietary video data inside the AWS boundary.
⚡ Key Takeaways
- QuickSearch provides four predefined actions (Shareable Scenes, Quotable Quotes, Memeable Moments, Trend Align) that return relevant clips without manual description.
- Scene‑level video reasoning is powered by TwelveLabs Marengo 3.0 and Pegasus 1.2 models hosted on Amazon Bedrock.
- Retrieval is performed via Amazon OpenSearch Serverless, enabling a serverless, scalable index of video assets.
- Bedrock Guardrails screen user search text for prompt attacks, and generated outputs are validated against source transcripts before publication.
- The system relies on Anthropic’s Claude models for additional language understanding, all running within the AWS data boundary.
Engineers can now deploy editorial‑driven video discovery at scale, reducing manual review time from weeks to seconds while maintaining data privacy and compliance through fully managed AWS services.
✅ Practical Steps
- Deploy the ingest and indexing pipeline on Amazon OpenSearch Serverless as described, feeding video metadata and transcripts into the index.
- Configure Amazon Bedrock to invoke TwelveLabs Marengo 3.0, Pegasus 1.2, and Anthropic Claude models for the QuickSearch actions.
- Integrate Bedrock Guardrails into the query front‑end to filter user input and set up post‑generation transcript verification before content is approved.
Want the full story? Read the original article.
Read on AWS ML Blog ↗