Snap’s EGR uses one shared LLM to power retrieval and lifts conversions 2.91%
_reachsumit · x · 2026-07-28
Snap presents EGR, an embedding-native generative retrieval framework for recommendation and advertising.
- EGR uses a single shared LLM to encode item metadata and user histories into the same embedding space.
- It removes the need for semantic-ID quantization and indexes items directly as dense vectors.
- The system is trained with joint contrastive learning so user queries and target items align better.
- Snap reports gains on Amazon Reviews and Snap DPA, with the production deployment delivering a +2.91% conversion-rate lift.
The paper’s practical angle is that it simplifies retrieval architecture while improving ad performance, including cold-start handling and multimodal inputs.
More from Companies & People
- HeyNox teases Caspian, an AI assistant that connects email, calendar and apps — avlok · 2026-07-28
- Big companies are hiring again as AI wipeout fears meet reality — jbarbier · 2026-07-28
- Ex-Shopify insider says Tobi Lütke was right to fix a culture problem — morganb · 2026-07-28
- Google, Meta and OpenAI sign open AI coalition letter as Anthropic stays out — FinanceYF5 · 2026-07-28
- AI engineering guide adds 4,894 job posts and real interview patterns across 51 companies — Al_Grigor · 2026-07-28
- A Stanford talk highlights a startup that cut its AI bill from $1.2M to about $100k a month — HeyAmit_ · 2026-07-28