Instagram's OneShot retrieval system boosts recall 20% at 10x speed

Researchers at Meta (Instagram) have published a paper on arXiv describing OneShot, a holistic retrieval framework designed to fix a long-standing problem in large-scale recommendation systems. Traditionally, retrieval has two conflicting goals: ranking accuracy (predicting what users will engage with) and indexing efficiency (structuring item representations for fast search). These are optimized separately, forcing systems to rely on simple dot-product scoring—a bottleneck that limits expressiveness.

OneShot addresses this by using an in-model index that is learned end-to-end together with ranking objectives, making the index itself ranking-aware. This joint learning removes the dot-product limitation, allowing more powerful neural scoring during retrieval. The system is already fully deployed in Instagram's short-video recommendation pipeline, where it drives significant gains in daily sessions, engagement, and time-spent. Benchmarks show a 20% recall improvement at operational ranking volume and a 10x efficiency gain at equivalent recall, proving that aligning index and ranking can deliver both better quality and lower computational cost.