Pinterest: Serving Two-Tower Models Using GPUs (Feb 2026)
This post is an index entry from a curated collection (awesome-list style) covering search, recommendation, and personalization research and engineering. It catalogs a Pinterest resource from February 2026 on serving two-tower models using GPUs.
Source
- Original link: https://lnkd.in/g4eyypSh
- Resource type: blog post / whitepaper
- Organization: Pinterest
- Date: February 2026
- BM25 / sparse retrieval: fast and robust, but no semantic matching.
- Cross-encoders: higher accuracy, but document representations cannot be precomputed, making them unsuitable for first-stage retrieval.
- Two-tower dense retrieval: millisecond-scale ANN recall with learned semantics, though sensitive to domain shift and long-tail queries.
- Generative retrieval / LLM agents: newer paradigms that treat retrieval as an iterative, plannable process, shifting attention to reasoning budgets and tool-call policies.
- IR: Information Retrieval
- RAG: Retrieval-Augmented Generation
- nDCG: Normalized Discounted Cumulative Gain, a ranking quality metric
- Agentic Search: modeling search as sequential decision-making with tool calls
- Gen-IR: Generative Information Retrieval
Context
Two-tower models are a foundational architecture in large-scale retrieval and recommendation. A query/user encoder and an item/document encoder independently produce embeddings; candidate retrieval is then performed via approximate nearest-neighbor (ANN) search over precomputed item embeddings. Compared with alternatives:
Engineering considerations highlighted by the collection
| Concern | Recommendation | |---------|----------------| | Data | PII handling, partitioned indexes, versioned/rollback-able embeddings | | Latency | p99 budgets, cascaded retrieval with early stopping, caching hot queries, async reranking | | Quality | Interleaving experiments, human audits, verifying offline gains translate to online CTR/satisfaction | | Safety | Source whitelists, poisoning/bias detection, output filtering | | Cost | Per-query GPU usage, model routing, distillation, hybrid sparse+dense retrieval |
Notes
The forum post does not reproduce detailed experimental figures from the Pinterest source; readers should consult the original publication for quantitative results. Related entries in the same collection include Netflix's foundation model for personalized recommendation, LLM-as-judge search relevance evaluation (Phi-3), and clickstream analyses of ChatGPT Search.