Following international sanctions in 2019 that restricted access to Google services, Huawei was compelled to rapidly create an independent mobile ecosystem, highlighted by its proprietary Petal Search engine.
The Challenge
Losing Google’s search features left Huawei’s Petal Search initially underdeveloped, especially in providing intuitive, fast, and accurate autocomplete suggestions—a critical element influencing user search satisfaction. Users frequently experienced irrelevant suggestions, delayed predictions, and limited personalization, leading to frustration and lower engagement.
To strengthen its market position and retain users accustomed to Google's efficient search experience, Huawei urgently needed advanced, AI-driven predictive autocomplete capabilities. Enhancing autocomplete was crucial for boosting search relevance, reducing user effort, and increasing user satisfaction and loyalty.
The Solution: LLM-based Autocomplete
Before founding Outter, the future Outter team members were pivotal figures within Huawei’s broader AI initiatives. Recognizing their expertise in content generation, recommendation systems, and predictive AI, Huawei entrusted them with enhancing Petal Search’s autocomplete functionality.
Working on Huawei’s complex autocomplete challenge allowed the team to significantly hone their skills, blending advanced AI methodologies with practical, user-centric product development. They gained invaluable experience managing large-scale multilingual datasets, optimizing machine learning models for real-time performance, and integrating sophisticated AI into user-facing products.
During their partnership with Huawei, the future Outter team spearheaded revolutionary improvements to Petal Search's autocomplete functionality, including:
Real-Time, AI-Driven Query Prediction
Model: Custom-built LLM trained on multilingual and domain-specific datasets
Technology: Precursor to Pangu NLP, leveraging deep-learning-based intent recognition
How it Works:
- AI predicts user intent in real time – reducing keystrokes and refining search accuracy.
- Multilingual adaptability – Seamlessly supports diverse user inputs across regions.
- Context-awareness – AI understands partial queries and suggests the most relevant completions instantly.
Deep Personalization: Search Tailored to Every User
Model: Personalized ranking algorithms integrating real-time user behavior analysisTechnology: Reinforcement learning, user embeddings, and context-aware query suggestionsHow it Works:
- AI adapts to individual search patterns – learning from previous queries and user preferences.
- Location and device-based personalization – adjusting results for regional and contextual relevance.
- Behavior-driven dynamic updates – prioritizing most relevant, frequently used search completions.
Tech Stack:
- Early Transformer Model (LLM Foundation for Huawei’s Pangu NLP) – Custom-built autocomplete AI that evolved into Huawei Pangu NLP.
- RLHF-Based Ranking Model (Foundation for Huawei Pangu Search Ranking) – Personalization model learning from user behavior.
- Huawei Cloud – Hosted entire AI pipeline on Huawei’s cloud infrastructure.
- Apache Kafka + Flink – Streaming real-time user interactions for AI adaptation.
- Vector Search (Milvus / Huawei GaussDB) – Fast, semantic query retrieval for autocomplete.
The Results
- Autocomplete prediction accuracy improved by 85%, rivaling the predictive capabilities users expected from global leaders like Google.
- User search speed increased by 60%, drastically enhancing user convenience and satisfaction.
- Overall user engagement and query completion rate rose by 70%, demonstrating significantly improved user satisfaction and retention.
- Marked increase in daily active Petal Search users, confirming stronger user confidence and loyalty.
