กลยุทธ์การเรียกคืนจากหลายคำค้น

Sometimes, a single search query isn't enough to find all the relevant information. Multi-query retrieval is a technique that helps your RAG system cast a wider net.

It generates several different versions of your original question. This helps ensure you don't miss important context, leading to more comprehensive answers.

Imagine asking "What are the benefits of RAG?". A single search might only pick up documents directly matching those exact words. This can limit the context available to your LLM.

• It might miss documents using terms like "advantages of RAG" or "why use RAG".
• It could also miss related concepts that provide crucial background.

This limitation can lead to incomplete or less accurate answers from the LLM.

Before advanced tools, people would manually brainstorm related queries to ensure a broader search. For example:

• Original: "How does RAG improve LLM accuracy?"
• Expanded: "What are RAG benefits for LLMs?", "RAG's impact on factual correctness", "How RAG reduces hallucinations".

While effective, this manual process is tedious and hard to scale for complex systems. We need an automated solution!

Large Language Models (LLMs) are excellent at understanding context and generating variations of text. We can leverage an LLM to automatically create several alternative questions from an initial user query.

• This uses the LLM's natural language understanding abilities.
• It automates the query expansion process efficiently.

These diverse queries then provide multiple 'angles' for searching your knowledge base.

LangChain provides a powerful component called MultiQueryRetriever. This tool automates the entire multi-query process for you, making it easy to integrate into your RAG pipeline.

• It uses an LLM to generate diverse queries from your initial input.
• It then runs these multiple queries against your vector store.
• Finally, it combines the results into a single, comprehensive set of documents for the LLM.

Let's see how to set up MultiQueryRetriever in Python. You'll need an LLM and an existing retriever (e.g., from a vector store).

This example mocks a vector store for demonstration. In a real application, your vector store would be pre-populated with your documents.

Here's a simplified breakdown of what MultiQueryRetriever does behind the scenes:

1. Initial Query: You provide one query, e.g., "What are RAG's advantages?"
2. Query Generation: The LLM takes your query and generates 2-4 alternative queries (e.g., "Benefits of RAG", "Why use RAG?", "How RAG improves LLMs").
3. Parallel Retrieval: Each generated query is sent to your underlying retriever (e.g., vector store) simultaneously.
4. Result Combination: All retrieved documents from these multiple searches are collected and often de-duplicated.
5. Final Context: This combined set of documents is then passed to your main LLM for generating the final answer.

After multiple queries fetch documents, how do we best combine them to form the final context?

• Unique Documents: The simplest approach is to gather all documents and remove duplicates. This ensures variety without redundancy.
• Re-ranking: For more advanced control, you can apply a re-ranking model to the combined set. This model scores documents based on their relevance to the original query, ensuring the most important ones are prioritized (a topic for a future lesson!).

Implementing multi-query strategies offers significant benefits for your RAG system:

• Improved Recall: You're more likely to find all relevant pieces of information, even if they use different phrasing.
• Richer Context: The LLM receives a broader and more diverse set of documents, leading to more comprehensive and accurate answers.
• Reduced Hallucinations: With better context, the LLM is less likely to "make things up" due to lack of information.
• Handles Ambiguity: Helps when the user's initial query is slightly ambiguous or could have multiple interpretations.

In this lesson, we explored Multi-Query Retrieval. We learned that a single query can often miss valuable context, and how Large Language Models (LLMs) can generate multiple, diverse queries to overcome this.

LangChain's MultiQueryRetriever automates this entire process, significantly improving the recall and richness of the context provided to your RAG system. This ultimately leads to more comprehensive and accurate answers.