benchen4395

benchen4395 / onesearch-family

Public

The corresponding codes and dataset for OneSearch series

18
2
69% credibility
Found Mar 30, 2026 at 18 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

A set of machine learning scripts for optimizing product search and recommendation systems through reward modeling, preference training, embedding quantization, and self-distillation techniques.

How It Works

1
🔍 Discover OneSearch tools

You find a helpful collection of tools to make search recommendations smarter and more accurate for products.

2
📦 Gather your search data

You collect lists of customer queries, product details, and their smart descriptions to prepare everything.

3
🔗 Combine queries and products

You mix your queries with product info so the tools can learn connections between what people search and what they like.

4
🗂️ Build smart categories

The tools automatically group similar items into efficient categories, making searches faster and more precise.

5
⭐ Define what makes a good match

You set up rewards that celebrate relevant, clickable, and popular products to guide the learning.

6
🧠 Train the preference learner

The system learns from good and bad examples to prefer top-quality search results every time.

7
✨ Refine with self-teaching

The tools improve themselves by learning from their own best outputs, getting even sharper.

🎉 Enjoy better searches

Your search now delivers spot-on product recommendations that customers love and click on more.

Sign up to see the full architecture

6 more

Sign Up Free

Star Growth

See how this repo grew from 18 to 18 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is onesearch-family?

onesearch-family delivers Python codes and datasets for the OneSearch series, tackling e-commerce search by training LLMs to generate hierarchical item descriptors from queries. It equips you with RLHF tools using listwise DPO and custom rewards blending relevance, CTR, clicks, and orders—plus efficient embedding quantization via residual and OPQ methods for fast retrieval. Users get prepped datasets and scripts to align models on real conversion signals, skipping generic RLHF setups.

Why is it gaining traction?

It stands out with domain-specific rewards for search (like click-order hierarchies) and token-position advantages in GRPO, outperforming vanilla DPO on structured outputs. The RQ-OPQ pipeline compresses embeddings while preserving recall, ideal for scaling retrieval without FAISS bloat. Developers hook on the self-distillation integration with LLaMA-Factory, distilling privileged info from full prompts into student models.

Who should use this?

ML engineers at e-commerce platforms tuning LLMs for query-to-item ranking, especially those handling Daikin-style catalogs with nested attributes. RecSys teams needing RLHF on proprietary signals like posterior CTR or purchase SIDs, or anyone quantizing query-item embeddings for production ANN search.

Verdict

Grab it if you're in search alignment—solid Python toolkit despite 18 stars signaling early maturity and thin docs. Credibility score of 0.699999988079071% flags caution on stability, but corresponding dataset and series codes make it worth forking for custom tweaks.

(187 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.