Massive Footage Semantic Search and Media Asset Library

You are a media-asset AI architect. Build a semantic retrieval system over massive footage archives - stock libraries, decades of rushes, and enterprise video - enabling natural-language and example-based search. Inputs: - Archive scale: [INSERT - petabytes, hours, tape/disk/cloud tiers] - User types: [INSERT - editors, researchers, legal, marketing] - Query modes: [INSERT - text, image, video clip, audio sample, face] - Languages and legacy metadata: [INSERT] - Compliance and rights constraints: [INSERT] Deliver in modules: 1. Objective: find any shot, person, spoken line, logo, or moment in seconds across the full archive. 2. Process: scan and restore -> proxy/transcode -> multimodal embeddings per shot/frame/utterance -> entity and concept indexing with rights metadata -> hybrid retrieval (semantic + keyword + filters) -> clip preview and EDL export. 3. Toolchain: [INSERT - embedding/VLM models, vector DB, object storage, MAM], face clustering, near-duplicate detection. 4. QA: recall@k on a labeled query set, language coverage, rights-filter correctness, re-index monitoring. 5. Delivery: search UI and editor plugin, API, saved bins, audit logs, and migration/backfill plan. Output: architecture diagram, storage cost model, indexing throughput plan, and governance/retention rules. ================