CardRewriter: Leveraging Knowledge Cards for Long-Tail Query Rewriting on Short-Video Platforms

TL;DR

CardRewriter tackles long-tail query rewriting on short-video platforms by grounding rewrites in concise knowledge cards synthesized from platform videos, similar queries, and open-domain sources. A two-stage training regime (supervised fine-tuning plus group relative policy optimization) with a customized reward system balances semantic relevance and retrieval impact, enabling effective card-guided rewrites. Offline experiments show improved query relevance and retrieval coverage, while online A/B tests on Kuaishou reveal meaningful gains in long-view rate and click-through rate alongside reduced initiative reformulations. The approach is deployed at scale, enhancing search quality for hundreds of millions of users and demonstrating the practical value of knowledge-grounded, retrieval-augmented rewriting on proprietary short-video content.

Abstract

Short-video platforms have rapidly become a new generation of information retrieval systems, where users formulate queries to access desired videos. However, user queries, especially long-tail ones, often suffer from spelling errors, incomplete phrasing, and ambiguous intent, resulting in mismatches between user expectations and retrieved results. While large language models (LLMs) have shown success in long-tail query rewriting within e-commerce, they struggle on short-video platforms, where proprietary content such as short videos, live streams, micro dramas, and user social networks falls outside their training distribution. To address this challenge, we introduce \textbf{CardRewriter}, an LLM-based framework that incorporates domain-specific knowledge to enhance long-tail query rewriting. For each query, our method aggregates multi-source knowledge relevant to the query and summarizes it into an informative and query-relevant knowledge card. This card then guides the LLM to better capture user intent and produce more effective query rewrites. We optimize CardRewriter using a two-stage training pipeline: supervised fine-tuning followed by group relative policy optimization, with a tailored reward system balancing query relevance and retrieval effectiveness. Offline experiments show that CardRewriter substantially improves rewriting quality for queries targeting proprietary content. Online A/B testing further confirms significant gains in long-view rate (LVR) and click-through rate (CTR), along with a notable reduction in initiative query reformulation rate (IQRR). Since September 2025, CardRewriter has been deployed on Kuaishou, one of China's largest short-video platforms, serving hundreds of millions of users daily.

Figures (7)

• Figure 1: An example of query rewriting on the short-video platform. (a) Original Query: Fails to retrieve relevant results; (b) LLM Rewrite: Misinterprets semantics and yields an ineffective rewrite; (c) CardRewriter: Leverages the generated knowledge card as search context to produce an accurate rewrite.
• Figure 2: The overall workflow of CardRewriter. Given an input query, CardRewriter retrieves multi-source knowledge from short-video platforms, summarizes it into a concise knowledge card relevant to the query, and then leverages this card to better interpret user intent and refine the query.
• Figure 3: Development Strategy.
• Figure 4: Effectiveness of different rewards.
• Figure 5: Effectiveness of different card generation methods.