하이브리드 검색: 벡터 + 키워드

Welcome to Hybrid Search! This lesson dives into a powerful technique that combines the best of two worlds: traditional keyword search and modern vector similarity search.

Pure keyword search can miss relevant results due to synonyms, while pure vector search might struggle with exact matches. Hybrid search aims to overcome these limitations.

Keyword search, often called lexical search, finds documents based on exact word matches or close variations. It uses inverted indexes to quickly locate terms.

• Strengths: Excellent for precise terms, names, or codes. Fast for exact matches.
• Weaknesses: Struggles with synonyms (e.g., 'car' vs. 'automobile'), contextual meaning, or queries phrased differently.

Vector search, or semantic search, operates on the meaning of words and phrases. It converts text into numerical vectors (embeddings) and finds semantically similar items.

• Strengths: Great for understanding intent, finding synonyms, and discovering conceptually related content.
• Weaknesses: Can miss exact keyword matches if the semantic meaning isn't strong. May struggle with very specific, rare terms.

Hybrid search bridges the gaps left by pure keyword or pure vector approaches. Imagine searching for 'best Italian restaurants'.

• Keyword search might find articles with 'Italian restaurants' but miss highly-rated places described differently.
• Vector search might find great restaurants but miss specific mentions of 'Italian' if the embedding doesn't heavily emphasize it.

Hybrid search combines both to give you the most comprehensive results.

In a nutshell, hybrid search performs both a vector similarity search and a keyword (lexical) search simultaneously or sequentially.

Each search method returns a list of results with associated relevance scores. The magic happens when these results and scores are combined into a single, unified ranking.

A common method for combining results is Reciprocal Rank Fusion (RRF). RRF takes the ranks of a document from different search results and calculates a combined score.

It's effective because it gives more weight to items that rank highly in multiple search lists, making it robust to individual search method biases.

While RRF is popular, you can also combine scores with a simple weighted sum, assuming scores are normalized. Here's a conceptual Python example:

Implementing hybrid search brings significant advantages to your applications:

• Improved Relevance: Users get more accurate and comprehensive results.
• Better Recall & Precision: You find more relevant items (recall) and fewer irrelevant ones (precision).
• Handles Diverse Queries: Effectively answers both specific keyword-driven queries and broad semantic questions.
• Robustness: Less susceptible to the limitations of a single search method.

Consider hybrid search when:

• Your data contains both highly specific terms and abstract concepts.
• User queries vary widely in their specificity and intent.
• You need to balance finding exact matches with understanding the overall meaning.
• Building RAG (Retrieval Augmented Generation) systems for LLMs, where precise context is crucial.

It's particularly useful in e-commerce, content recommendation, and knowledge base applications.

In this lesson, we explored Hybrid Search, a technique that intelligently combines traditional keyword search with modern vector similarity search.

By leveraging the strengths of both lexical and semantic matching, hybrid search delivers more relevant, robust, and comprehensive results, overcoming the individual limitations of each method. It's a crucial tool for building advanced information retrieval systems.