VectorPath SearchMap: Conversational E-commerce Search Embedding Model
SearchMap is a specialized embedding model designed to change search by making it more conversational and intuitive. We test out this hypothesis by creating a model suitable for ecommerce search. Fine-tuned on the Stella Embed 400M v5 base model, it excels at understanding natural language queries and matching them with relevant products.
- Optimized for conversational e-commerce queries - Handles complex, natural language search intents - Supports multi-attribute product search - Efficient 1024-dimensional embeddings (configurable up to 8192) - Specialized for product and hotel search scenarios
- Base Model: Stella Embed 400M v5 - Embedding Dimensions: Configurable (512, 768, 1024, 2048, 4096, 6144, 8192) - Training Data: 100,000+ e-commerce products across 32 categories - License: MIT - Framework: PyTorch / Sentence Transformers
The model excels at understanding natural language queries like: - "A treat my dog and I can eat together" - "Lightweight waterproof hiking backpack for summer trails" - "Eco-friendly kitchen gadgets for a small apartment" - "Comfortable shoes for standing all day at work" - "Cereal for my 4 year old son that likes to miss breakfast"
Evaluation The model's evaluation metrics are available on the MTEB Leaderboard - The model is currently by far the best embedding model under 1B parameters size and very easy to run locally on a small GPU due to it's memory size - The model also is No 1. by a far margin on the SemRel24STS task with an accuracy of 81.12% beating Google Gemini embedding model (second place) 73.14% (as at 30th March 2025). SemRel24STS evaluates the ability of systems to measure the semantic relatedness between two sentences over 14 different languages. - We noticed the model does exceptionally well with legal and news retrieval and similarity task from the MTEB leaderboard
Strengths - Excellent at understanding conversational and natural language queries - Strong performance in e-commerce and hotel search scenarios - Handles complex multi-attribute queries - Efficient computation with configurable embedding dimensions
Current Limitations - May not fully prioritize weighted terms in queries - Limited handling of slang and colloquial language - Regional language variations might need fine-tuning
The model was trained using: - Supervised learning with Sentence Transformers - 100,000+ product dataset across 32 categories - AI-generated conversational search queries - Positive and negative product examples for contrast learning
This model is designed for: - E-commerce product search and recommendations - Hotel and accommodation search - Product catalog vectorization - Semantic similarity matching - Query understanding and intent detection
If you use this model in your research, please cite:
- Discord Community: Join our Discord - GitHub Issues: Report bugs and feature requests - Interactive Demo: Try it on Colab
This model is released under the MIT License. See the LICENSE file for more details.