การค้นคืนข้อมูลตามบริบท

In a Retrieval-Augmented Generation (RAG) system, the Large Language Model (LLM) needs relevant information to generate accurate responses.

• Contextual Information Retrieval is the process of finding and fetching this relevant data from your vector database.
• It's the bridge that connects the user's query to the knowledge stored in your specialized data.

When a user asks a question, several steps happen to get the right context:

1. The user's question (query) is converted into a vector embedding.
2. This query embedding is sent to the vector database.
3. The vector database searches for stored document embeddings that are most similar to the query embedding.
4. The text chunks associated with these similar embeddings are retrieved and sent to the LLM.

The core of retrieval is vector similarity. Your vector database calculates how 'close' your query vector is to all the stored document vectors.

• Closer vectors mean higher semantic similarity.
• Common similarity metrics include cosine similarity or Euclidean distance.
• The database efficiently finds the closest vectors, usually using specialized indexing techniques.

The most straightforward retrieval strategy is Top-K Retrieval.

• You simply ask the vector database to return the K most similar document chunks to your query.
• K is a number you choose (e.g., 3, 5, or 10), representing how many pieces of context you want to provide to the LLM.
• While simple, choosing the right K is crucial for balancing relevance and LLM token limits.

This simple Python code simulates a vector store and demonstrates how top_k retrieval works. Try changing the top_k value!

The quality of your retrieval heavily depends on how your original documents were broken down into chunks before being embedded.

• Chunk Size: Too small, and context might be lost. Too large, and irrelevant information might be included.
• Overlap: Adding overlap between chunks helps ensure that important information isn't split across boundaries.
• Good chunking ensures each retrieved piece of context is meaningful and self-contained.

Sometimes, simple Top-K retrieval isn't enough. The most 'similar' vectors aren't always the most 'relevant' in context.

• Re-ranking is an optional but powerful step performed after initial retrieval.
• It takes the top K chunks from the vector database and uses a smaller, more specialized model to score their relevance more deeply.
• This secondary scoring helps filter out less useful chunks and prioritize truly pertinent information.

Why do we need re-ranking?

• Vector similarity can sometimes be fooled by superficial semantic closeness.
• A re-ranker, often a smaller transformer model, can better understand the nuanced relationship between the query and the retrieved document chunks.
• It helps ensure the context provided to the LLM is not only similar but also highly relevant and useful for answering the user's specific question.

Beyond basic Top-K and re-ranking, advanced strategies can further improve retrieval:

• Query Expansion: Rewriting or adding terms to the user's original query to improve search results.
• Hybrid Search: Combining traditional keyword search (like full-text search) with vector similarity for a more comprehensive retrieval.
• These methods aim to make the initial retrieval even more robust before context is sent to the LLM.

You've learned how to retrieve relevant context for RAG systems!

• Contextual retrieval bridges user queries and stored knowledge.
• It involves converting queries to embeddings, querying the vector DB for similar vectors, and fetching associated text.
• Top-K retrieval is the basic method, but good chunking is vital.
• Re-ranking can further refine results by applying a secondary relevance filter.
• Advanced techniques like query expansion and hybrid search offer even more sophisticated retrieval.