Xiaohongshu Search Advertising Recall: Practices, Metrics, and Large‑Model Integration

Xiaohongshu combines content sharing, community interaction, and e‑commerce, and its search function has become a crucial channel for users to obtain information and make purchasing decisions. The rapid growth of advertising material and diversified user needs pose multiple business and technical challenges for the search advertising recall system.

The article examines the business‑driven recall practice during a growth phase, describing how data loops and complex recall models are built to ensure high‑quality ad distribution while preserving user experience.

Business characteristics include strong semantic constraints, explosive material growth, and coexistence of multiple advertising goals (clicks, conversions, lead capture). These constraints require the recall stage to enforce strict relevance while handling a massive and fast‑expanding candidate pool.

Water‑level metrics are defined to monitor recall effectiveness: "should‑recall‑as‑much-as‑possible", "noise‑correction", and "instant‑recall". By sampling queries across frequency tiers and scoring inventory offline, the team measures coverage, noise ratio, PVR gaps, and cold‑start success rates, enabling clear identification of improvement space.

Semantic vs. efficiency trade‑off is addressed through a two‑track strategy: a semantic model that prioritizes relevance on long‑tail traffic, and an efficiency model that maximizes platform revenue on head traffic. The combined approach balances user experience, advertiser fairness, and platform profitability.

Evolution stages are outlined:

Stage 1 – Advertiser‑driven keyword bidding with query rewrite and inverted‑index recall.

Stage 2 – Decoupling from keyword bidding using BERT‑based vector retrieval to capture unbidded but relevant ads.

Stage 3 – Multi‑objective recall (relevance + CTR + ECPM) with separate semantic and efficiency channels.

Stage 4 – Strengthening retrieval models and indexes to mitigate selection bias.

Stage 5 – Large‑model era (LLM‑driven query rewriting, ad‑selling‑point extraction, and representation learning).

Model upgrades move from simple dual‑tower inner‑product models to deeper MLP + Target‑Attention architectures, combined with HNSW hierarchical indexes for efficient nearest‑neighbor search.

Large‑model integration includes real‑time LLM‑based suggestion generation, low‑cost fine‑tuning via LoRA/DPO, and embedding alignment to the advertising semantic space. These advances enable knowledge‑enhanced retrieval, multimodal understanding, and scaling‑law‑driven performance gains.

Engineering optimizations focus on GPU‑accelerated training, mixed‑precision, operator fusion, and a TensorFlow‑based full‑graph retrieval engine that supports minute‑level index updates and high QPS.

Overall, the system achieves coverage of 80 % of clicks and 60 % of conversions in the efficiency channel, contributing a +5 % CPM lift after five iterative launches. The article concludes that search technology will increasingly serve as a retrieval‑augmented generation (RAG) backbone for LLMs, reshaping human‑machine interaction in the AI era.