VisualText Search

Bridging the gap between words and visuals in product discovery — for search that understands what users see, say, and mean.

Combining text and images

In e-commerce, customers rely on both product descriptions and images to make purchasing decisions. Traditional search systems often focus solely on text, missing the rich information conveyed by visuals. VisualText Search bridges this gap by integrating textual and visual data, enhancing the relevance and accuracy of search results.

How it works

  1. Multimodal Representation:
    • Combines product titles and images into a unified representation, capturing nuanced product attributes that text alone might miss.
  2. Dual-Encoder Architecture:
    • Employs separate encoders for text and images, allowing efficient processing and retrieval of multimodal data.
  3. Improved Retrieval Performance:
    • Demonstrates enhanced purchase recall and relevance accuracy compared to text-only models, leading to better user satisfaction.
  4. Exclusive Match Analysis:
    • Provides insights into unique matches retrieved by the multimodal model, highlighting its ability to uncover relevant products missed by traditional methods.

Why it matters

• More accurate ranking: Combines the visual semantics of product appearance with descriptive language for better matches.
• Improves discovery: Helps users find relevant products even when they use limited or ambiguous text.
• Supports zero-shot search: Enables matching to unseen or rare product types by leveraging visual similarity.
• No extra training cost: Efficient retrieval thanks to precomputed multimodal embeddings and shared vector space.

Proven results

VisualText Search has been benchmarked across large-scale product catalogs and public datasets. Results show:
↑ 9.1% improvement in purchase recall vs. text-only baselines
↑ 8.7% improvement in relevance accuracy (human-rated)
↑ +5% exclusive matches — items found only through multimodal search

These gains translate to better engagement, reduced bounce rates, and more satisfied users — especially in fashion, furniture, beauty, and other visual-heavy categories.

Use cases

How VisualText Search creates real value
• Fashion: understands style, silhouette, and color beyond keywords
• Home Decor: matches based on texture, form, and visual theme
• Beauty: ranks products by packaging, branding, and finish
• Second-hand marketplaces: improves matching for listings with poor text quality

Get started with Nibelung

Ready to elevate your on-site search and navigation? Contact us today to learn more about our AI Search Audit Agent module and how it can benefit your business.


hello@nibelung.ai

Vector search

Vector search is a machine learning technology for AI search. Now anyone can add AI-powered search to their site or app in minutes to deliver a better experience for their customers.

Re-ranking

Rank your search results based on your business needs, time of the year, scheduled campaigns, special offers, and KPIs. Boost the most profitable products and help them get noticed by your visitors first.

Personalization

For each user, the search experience becomes more interesting with the addition of personalised preferences. Interactive search is strengthened by personalisation.

Search insights

Search analytics offer insights into what your users search and which results they select through metrics such as popular searches, no results, and click-through rates.

Autocomplete

Autocomplete refers to a feature that helps users find information more quickly by predicting and suggesting possible completions for the query they are typing into a search box.