AI-Powered Search at Scale
summary
Frame.io uses Llama 4 via Bedrock → OpenSearch DSL to turn natural language into structured queries and CLIP-style and Adobe Research “vegan” models to produce visual embeddings. The ingest moves from Oban to a Broadway + SQS pipeline replacing Oban for independent scaling and buffering; most embeddings finish under a minute but outliers can take hours—one hit eight hours—so backpressure and DLQs matter. A 2,000-line system prompt plus injected schema drives query generation; telemetry, regression suites and latency-first metrics keep error rates and malformed queries very low at ~10k Bedrock calls/hour.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:05Talk goals
- 2:29Frame.io video search
- 3:56Semantic embeddings
- 7:47Broadway for long jobs
- 9:56Media transcode pipeline
- 12:44Embeddings in OpenSearch
- 13:05Natural language search
- 17:59LLM prompt tuning
- 21:32Production AI lessons
- 26:17Monitoring AI pipelines
- 29:50Q&A
related talks
26mAdvancing Multi-Modal Search Capabilities in Search Pipeline
01h 30mBuilding a Search Engine
27mAI Demystified Unveiling the Intelligence Behind the Machine
01h 28mBuilding a Semantic Search Engine
43mBird's Eye View of Open-Source AI Infrastructure
21mHow Web Data Infrastructure Powers the Next Generation of AI
31mAgents Need More than Vector Search
01h 08mGoing multi-modal