README.md

---
language: en
license: apache-2.0
tags:
- sentence-transformers
- feature-extraction
- saas
- product-recommendation
- semantic-search
pipeline_tag: sentence-similarity
library_name: sentence-transformers
datasets:
- ComparEdge/saas-market-intelligence
---

# ComparEdge SaaS Product Matcher

Semantic search model for SaaS product recommendation. Fine-tuned on 331 product descriptions from [ComparEdge](https://comparedge.com) — a live SaaS comparison platform covering dozens of categories.

Given a natural-language query, this model returns the most relevant SaaS tools from the ComparEdge database.

## Quick Start

```python
from sentence_transformers import SentenceTransformer
from sentence_transformers.util import cos_sim
from huggingface_hub import hf_hub_download
import numpy as np, json, torch

model = SentenceTransformer("ComparEdge/saas-product-matcher")

# Load pre-computed embeddings (hundreds of products, no re-encoding needed)
emb_path = hf_hub_download("ComparEdge/saas-product-matcher", "product_embeddings.npy")
idx_path  = hf_hub_download("ComparEdge/saas-product-matcher", "products_index.json")

embeddings = np.load(emb_path)
with open(idx_path) as f:
    products = json.load(f)

query  = "I need a CRM for a small startup"
q_emb  = model.encode(query, normalize_embeddings=True)
scores = cos_sim(torch.tensor(q_emb), torch.tensor(embeddings))[0]
top_idx = scores.argsort(descending=True)[:5]

for idx in top_idx:
    p = products[idx]
    print(f"{p['name']} ({p['category']}): {scores[idx]:.3f}")
    print(f"  → https://comparedge.com/tools/{p['slug']}")
```

## Repository Files

| File | Description |
|---|---|
| `product_embeddings.npy` | Pre-computed normalized embeddings for all hundreds of products (shape: 331×384) |
| `products_index.json` | Metadata index: slug, name, category, description |
| `example_search.py` | Standalone CLI search script |

## Training

**Base model:** `sentence-transformers/all-MiniLM-L6-v2` (384-dim, 22M params)

**Loss:** `MultipleNegativesRankingLoss` — treats every other item in the batch as a hard negative, which works well for retrieval tasks without manual negative mining.

**Training data:** ~4,000 (query, product) pairs generated from 331 SaaS products across dozens of categories:

- Natural-language queries from 16 templates per product ("best X tool", "cheap X software", etc.)
- Product descriptions, long-form reviews, and feature lists
- Use-case titles extracted from each product
- Pricing signals (free-plan queries, under-$N queries)

**Coverage:** 28 SaaS categories from [comparedge.com](https://comparedge.com):

| Category | Example products |
|---|---|
| project-management | Notion, Asana, Linear, ClickUp |
| crm | HubSpot, Pipedrive, Salesforce |
| email-marketing | Mailchimp, ActiveCampaign, Brevo |
| video-conferencing | Zoom, Google Meet, Whereby |
| ai-writing | Jasper, Copy.ai, Writesonic |
| design-tools | Figma, Canva, Adobe XD |
| password-managers | 1Password, Bitwarden, Dashlane |
| vpn | NordVPN, ExpressVPN, Surfshark |
| … 20 more | … |

## Performance

Evaluated on held-out queries not seen during training:

| Metric | Score |
|---|---|
| Top-1 accuracy | ~78% |
| Top-5 accuracy | ~94% |
| Mean Reciprocal Rank | 0.85 |

## Links

- 🌐 [ComparEdge](https://comparedge.com) — Live SaaS comparison and product discovery platform
- 📊 [Dataset](https://huggingface.co/datasets/ComparEdge/saas-market-intelligence) — Raw product data on HuggingFace
- 🔗 [API](https://comparedge-api.up.railway.app/docs) — REST API for programmatic access