Asymmetric Search
A search paradigm where queries and documents are encoded differently, optimized for scenarios where queries are short and documents are long. Common in information retrieval and modern embedding models designed specifically for search.
About this tool
Overview
Asymmetric search recognizes that queries and documents have different characteristics: queries are typically short (3-10 words) while documents are long (100-1000s of words). Some embedding models optimize for this asymmetry.
Query vs Document
Query: "best pizza recipe"
- Short
- Keyword-like
- User intent
- Natural language questions
Document: Full recipe with ingredients, instructions, tips
- Long
- Complete information
- Structured content
Asymmetric Models
Models like Sentence-BERT can be trained with different encoders or prompts for queries vs documents:
# Some models use prefixes
query_embedding = model.encode("query: best pizza recipe")
doc_embedding = model.encode("passage: [full recipe text]")
Benefits
- Better matching of short queries to long documents
- Optimized for search use case
- Improved recall
Contrast with Symmetric
Symmetric: Document-to-document similarity (clustering) Asymmetric: Query-to-document search (retrieval)
Model Examples
- INSTRUCTOR embeddings (instruction-based)
- Cohere embed-english-v3 (search-optimized)
- Some Sentence-BERT configurations
Pricing
Not applicable (search paradigm).
Loading more......
Information
Categories
Tags
Similar Products
6 result(s)