README.md

---
language: en
tags:
- embedding
- transformers
- search
- e-commerce
- conversational-search
- semantic-search
license: mit
pipeline_tag: feature-extraction
---

# VectorPath SearchMap: Conversational E-commerce Search Embedding Model

## Model Description

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.

## Key Features

- 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

## Quick Start

Try out the model in our interactive [Colab Demo](https://colab.research.google.com/drive/1wUQlWgL5R65orhw6MFChxitabqTKIGRu?usp=sharing)!

## Model Details

- 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

## Usage

### Using Sentence Transformers

```python
# Install required packages
!pip install -U torch==2.5.1 transformers==4.44.2 sentence-transformers==2.7.0 xformers==0.0.28.post3

from sentence_transformers import SentenceTransformer

# Initialize the model
model = SentenceTransformer('vectopath/SearchMap_Preview', trust_remote_code=True)

# Encode queries
query = "A treat my dog and I can eat together"
query_embedding = model.encode(query)

# Encode products
product_description = "Organic peanut butter dog treats, safe for human consumption..."
product_embedding = model.encode(product_description)
```

### Using with FAISS for Vector Search

```python
import numpy as np
import faiss

# Create FAISS index
embedding_dimension = 1024  # or your chosen dimension
index = faiss.IndexFlatL2(embedding_dimension)

# Add product embeddings
product_embeddings = model.encode(product_descriptions, show_progress_bar=True)
index.add(np.array(product_embeddings).astype('float32'))

# Search
query_embedding = model.encode([query])
distances, indices = index.search(
    np.array(query_embedding).astype('float32'), 
    k=10
)
```

### Example Search Queries

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"

## Performance and Limitations

### Evaluation
The model's evaluation metrics are available on the [MTEB Leaderboard](https://huggingface.co/spaces/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](https://huggingface.co/datasets/SemRel/SemRel2024) 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

## Training Details

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

## Intended Use

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

## Citation

If you use this model in your research, please cite:

```bibtex
@misc{vectorpath2025searchmap,
  title={SearchMap: Conversational E-commerce Search Embedding Model},
  author={VectorPath Research Team},
  year={2025},
  publisher={Hugging Face},
  journal={HuggingFace Model Hub},
}
```

## Contact and Community

- Discord Community: [Join our Discord](https://discord.gg/gXvVfqGD)
- GitHub Issues: Report bugs and feature requests
- Interactive Demo: [Try it on Colab](https://colab.research.google.com/drive/1wUQlWgL5R65orhw6MFChxitabqTKIGRu?usp=sharing)

## License

This model is released under the MIT License. See the LICENSE file for more details.