การค้นหาแบบผสมและการจัดอันดับใหม่

When building advanced Retrieval Augmented Generation (RAG) systems, simply finding documents isn't enough. We need to find the most relevant documents efficiently.

Traditional keyword or semantic searches, while powerful, each have limitations. To overcome these, we can combine their strengths.

Keyword search (also known as sparse retrieval, e.g., using BM25 or TF-IDF) is excellent for finding exact matches and specific terms.

• Strengths: Great for precise queries, proper nouns, and when you know the exact wording.
• Weaknesses: Struggles with synonyms, different phrasing, or understanding conceptual meaning.

For example, searching 'Python list append' works well, but 'add element to Python array' might miss results.

Semantic search (dense retrieval, using embeddings) understands the meaning and context of your query and documents.

• Strengths: Handles synonyms, rephrased questions, and conceptual searches effectively.
• Weaknesses: Can struggle with very specific, rare terms or highly technical jargon if not well-represented in its embedding space.

It can understand 'add element to Python array' is similar to 'Python list append'.

Hybrid search combines the best of both worlds: the precision of keyword search and the contextual understanding of semantic search.

By running both types of retrieval and intelligently merging their results, hybrid search can significantly improve the relevance and completeness of retrieved documents for your RAG system.

A common method to combine results from multiple retrievers in hybrid search is Reciprocal Rank Fusion (RRF).

RRF assigns a score to each document based on its rank in each individual retriever's result list. Documents that appear high in multiple lists get a boosted score, leading to a more robust final ranking.

Try running this simplified example of RRF:

LangChain provides an EnsembleRetriever to easily implement hybrid search. It takes multiple retrievers (e.g., a keyword retriever and a vector store retriever) and combines their results, often using RRF by default.

This allows you to leverage both precise keyword matches and semantic understanding in one powerful retrieval step.

Even after hybrid search, the initial set of retrieved documents might contain some noise or documents that are not perfectly ordered by relevance.

Re-ranking is a crucial next step. It takes the top-k documents from the initial retrieval and re-evaluates their relevance to the query using a more sophisticated model.

A re-ranking model, often a cross-encoder, takes both the user query and each retrieved document as input.

Unlike embedding models that create separate embeddings, a cross-encoder jointly processes the query and document to generate a single relevance score. This allows for a more nuanced understanding of their interaction.

Various re-ranking models and services are available, such as Cohere's Re-rank API or open-source cross-encoders from libraries like sentence-transformers.

Integrating a re-ranker typically involves passing the initial retrieval results and the query to the re-ranker, which then returns the documents in a new, optimized order.

By combining hybrid search and re-ranking, you create a robust retrieval pipeline:

1. User Query
2. Hybrid Search: Combines keyword and semantic retrieval to get an initial set of relevant documents.
3. Re-ranking: A specialized model re-orders the top documents from hybrid search for maximum relevance.
4. Context for LLM: The highly relevant, re-ranked documents are passed to the LLM for generation.

This approach leads to more accurate and contextually rich answers from your RAG system.

Which of the following statements accurately describe the roles of Hybrid Search and Re-ranking in a RAG system?

We've explored advanced retrieval techniques to boost RAG performance:

• Hybrid Search: Combines keyword and semantic approaches for comprehensive initial retrieval.
• Reciprocal Rank Fusion (RRF): A method to merge and score results from different retrievers.
• Re-ranking: Uses specialized models (like cross-encoders) to refine the relevance order of retrieved documents.

These techniques ensure your RAG system provides the most accurate and contextually relevant information to the LLM.