Skip to Content
SearchSemantic Search

Semantic Search

Virza’s search system combines keyword matching with semantic understanding in every query. There is no separate “semantic search mode” to enable.

For the full search architecture, including how queries are classified, how results are ranked across seven signals, and how the three search tabs work, see How Search Works.

What semantic search does

Semantic search uses AI embeddings (vector representations of meaning) to find documents that are conceptually related to your query, even when they don’t share the same words:

  • “climate impact on migration” finds papers about climate refugees, displacement patterns, and environmental asylum
  • “neural network architectures” finds papers about transformers, CNNs, RNNs, and attention mechanisms
  • “side effects of treatment” finds papers discussing adverse reactions, complications, and safety profiles

Virza does not show keyword and semantic results separately. Both signals feed into a unified ranking system:

  1. Your query is matched against the full-text index (keyword matching)
  2. Your query is embedded and compared against document embeddings (semantic matching)
  3. A cross-encoder neural model reranks results by reading your query alongside each candidate
  4. Seven ranking signals are combined into a final relevance score

The relative weight of keyword vs. semantic matching adjusts automatically based on query type. Short keyword queries lean toward lexical matching, while natural language questions boost semantic similarity.

When semantic search helps most

ScenarioExample
Exploring a concept you can’t name precisely”methods for measuring social isolation in elderly populations”
Finding related work across disciplines”computational approaches to protein folding”
Searching with synonyms or paraphrases”machine learning for medical images” finds “deep learning for radiology”
Asking questions”How does attention improve NLP performance?”

For known-item searches (specific author, DOI, or exact title), keyword matching is already optimal and activates automatically.

Further reading

Last updated on