Interpretation of the TenderHack 2026 Champion Solution: Architecture of an AI Intelligent Search System Based on Multi-Strategy Fusion

This article interprets smart-search-system, the champion project of TenderHack2026 Hackathon—developed by the Ir0n Chain team, it is a multi-strategy AI search system for e-commerce procurement scenarios. The system solves three major challenges: spelling error tolerance, synonym understanding, and personalized result ranking. It adopts a microservice architecture to support local deployment, integrates multiple search technologies, and provides a complete technical reference for enterprise-level semantic search platforms.

The core goal of the system is to solve three major e-commerce search challenges: spelling error tolerance, synonym understanding, and personalized ranking. The team analyzed real business data and found: the median number of signed contracts per platform user is about 200, with 160 independent SKUs, and the repeat purchase rate is only 18%. Users are "list-type buyers" with clear procurement needs. Based on this, the system design focuses on context awareness, understanding the user's business domain and historical procurement patterns to achieve personalized search.

The project uses a microservice architecture and can be deployed with one click via Docker Compose. Core components include: front-end service (port 3000), Go Fiber backend API (port 8080), ML inference service (port 8000), PostgreSQL database, MinIO object storage, Qdrant vector database, and Ollama local LLM. All services can be deployed locally without relying on external APIs, making it suitable for data-sensitive enterprises. The tech stack includes Go Fiber, PostgreSQL, MinIO, Qdrant, etc., each serving its own purpose.

The system integrates five complementary technologies:

1. Spelling correction: Uses the Levenshtein distance algorithm to correct input errors in real time;
2. Semantic search: Encodes queries and products into vectors, and achieves semantic matching via Qdrant vector database;
3. Morphological search: Combines lemmatization with TF-IDF to handle languages with rich morphology (e.g., Russian);
4. User domain matching: Builds user profile vectors and calculates the matching degree with product domains;
5. Historical purchase similarity: Based on collaborative filtering, increases the ranking weight of products similar to the user's historical purchases.

The system supports three re-ranking configuration modes:

• Pure text scenario: BM25 initial screening + CatBoostRanker re-ranking, with features including TF-IDF, BM25 scores, etc.;
• Text + vector scenario: Dual-tower retrieval recall + CatBoost/LightGBM re-ranking, with features extended to vector similarity, price, etc.;
• Full profile scenario: Integrates user vector features, uses cosine similarity, dot product, etc., and supports DIN/DIEN sequence models when data is sufficient.

The system has an interpretability design that can show users the reasons for product relevance, improving trust and debugging efficiency. In addition, it implements a session-level deduplication mechanism, so that products ignored in the same session are not displayed repeatedly, optimizing the user experience.

Deployment aspects: Provides Docker Compose configurations, supports three modes (production, ML, simulation), manages configurations via .env files, has an independent database migration service, and configures health checks for each service. Summary insights: Modern search requires multi-strategy fusion; personalization should be deeply integrated into ranking; interpretability and local deployment are important; it is recommended to evolve incrementally: start with a BM25+CatBoost baseline, gradually introduce vector search, and finally use deep sequence models.