OnePiece at Shopee: How LLM-style reasoning enters a ranking system

Introduction

When hearing “LLM-style reasoning”, many people immediately think of chatbots. But Shopee’s OnePiece paper shows another direction: bringing context engineering and reasoning ideas into an industrial ranking/recommendation system.

This is quite interesting, because ranking systems have traditionally been viewed as a pipeline of retrieval → ranking → reranking, optimizing CTR, CVR, GMV, or some business metric.

OnePiece asks the question: if LLMs get stronger thanks to context and reasoning, can ranking models learn a part of that mindset?

1. A ranking system is not just model scoring

In e-commerce, ranking is not simply “sorting products from good to bad”.

A ranking system must consider:

• who the user is;
• what they just searched for;
• what they clicked/bought before;
• what the current context is;
• which products are potentially a good fit;
• which business metric is being optimized.

The pipeline is usually a cascade:

Candidate Retrieval → Ranking → Reranking → Final List

Each layer filters out the number of items and increases precision.

2. OnePiece brings context engineering to ranking

According to the paper, OnePiece brings two ideas from LLMs into retrieval/ranking:

• structured context engineering: enriching the input with interaction history, preferences, and scenario signals;
• block-wise latent reasoning: allowing the model to have steps to refine representations in a more multi-step manner;
• progressive multi-task training: using user feedback sequences to supervise the learning process.

Simply put: instead of just feeding raw features into the model, the system tries to “tell the right context” so the model understands the user and the situation better.

3. An easy-to-understand example

Suppose a user searches for “rain jacket”.

If only looking at the query, the system might return raincoats, waterproof jackets, and windbreakers.

But if context is added:

User previously bought trekking gear
Currently in a rainy area
Usually chooses medium-priced products
Recently clicked on outdoor gear

Ranking might prioritize outdoor waterproof products instead of cheap plastic raincoats.

This is the spirit of context engineering: the model doesn’t naturally understand everything if we don’t provide enough structured context.

4. Lessons for CRM search

A CRM also has a ranking problem.

When searching for a customer or lead, results should consider:

• name/contact match;
• open deals;
• latest interaction;
• priority level;
• deal value;
• quote status;
• follow-up history.

Just keyword matching is not enough. A hot lead must be prioritized over an old, inactive contact, even if the text match is the same.

5. Conclusion

OnePiece is a good example that “LLM thinking” doesn’t just belong in chatbots. Context engineering and reasoning can become a pattern for many other systems, especially search/ranking/recommendation.

The important thing is not to drag LLMs into everything, but to learn the useful parts: providing better context, designing better representations, and evaluating with real metrics.

References

• OnePiece: Bringing Context Engineering and Reasoning to Industrial Cascade Ranking System: https://arxiv.org/abs/2509.18091
• Hugging Face Papers — OnePiece: https://huggingface.co/papers/2509.18091