ハイブリッド検索と再ランキング
キーワード検索とセマンティック検索を組み合わせ、再ランキングモデルで最も関連性の高いドキュメントを優先します。
「ハイブリッド検索と再ランキング」はCoddyKit上の無料LangChain / RAG / Vector DBsレッスンです。 これはレッスン3/4です。 下記で完全なレッスンを無料で読むことができます。その後、ブラウザ内の組み込みコードエディタと24時間対応のAIチューターでハンズオン演習できます。 これはLangChain / RAG / Vector DBs学習パスの一部であり、ウェブとCoddyKitアプリ全体で進捗が同期されます。 LangChain / RAG / Vector DBsコースには全4レッスンが含まれています。
このレッスンの一部はまだ翻訳されておらず、英語で表示されています。
Beyond Basic Retrieval
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: Specificity
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: Understanding Meaning
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'.
Introducing Hybrid Search
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.
Merging Results: Reciprocal Rank Fusion
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:
def reciprocal_rank_fusion(rank_lists, k=60):
fused_scores = {}
for rank_list in rank_lists:
for rank, doc_id in enumerate(rank_list):
if doc_id not in fused_scores:
fused_scores[doc_id] = 0.0
fused_scores[doc_id] += 1.0 / (k + rank + 1)
sorted_docs = sorted(fused_scores.items(), key=lambda item: item[1], reverse=True)
return [doc_id for doc_id, score in sorted_docs]
if __name__ == "__main__":
# Simulate results from two retrievers
keyword_results = ["docA", "docC", "docB", "docE"]
semantic_results = ["docB", "docA", "docD", "docC"]
fused_order = reciprocal_rank_fusion([keyword_results, semantic_results])
print("Fused Order:", fused_order)Hybrid Search with LangChain
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.
# from langchain.retrievers import EnsembleRetriever
# from langchain_community.retrievers import BM25Retriever
# from langchain_community.vectorstores import FAISS
# from langchain_openai import OpenAIEmbeddings
# # Assume you have a BM25 retriever and a vector store retriever
# bm25_retriever = BM25Retriever.from_documents(docs)
# vectorstore = FAISS.from_documents(docs, OpenAIEmbeddings())
# vectorstore_retriever = vectorstore.as_retriever()
# ensemble_retriever = EnsembleRetriever(retrievers=[
# bm25_retriever,
# vectorstore_retriever
# ], weights=[0.5, 0.5])
# # query = "What are the capital cities of Europe?"
# # docs = ensemble_retriever.invoke(query)The Need for Re-ranking
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.
How Re-ranking Works
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.
Integrating Re-rankers
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.
# from langchain.retrievers.document_compressors import CohereRerank
# from langchain.retrievers import ContextualCompressionRetriever
# # Assume you have an existing base_retriever (e.g., your EnsembleRetriever)
# # base_retriever = ensemble_retriever
# # Initialize the Cohere Rerank compressor
# # cohere_re_ranker = CohereRerank(top_n=5, cohere_api_key="YOUR_COHERE_API_KEY")
# # Create a compression retriever that uses the re-ranker
# # compression_retriever = ContextualCompressionRetriever(
# # base_compressor=cohere_re_ranker,
# # base_retriever=base_retriever
# # )
# # query = "What is the capital of France?"
# # compressed_docs = compression_retriever.invoke(query)The Full Advanced Retrieval Pipeline
By combining hybrid search and re-ranking, you create a robust retrieval pipeline:
- User Query
- Hybrid Search: Combines keyword and semantic retrieval to get an initial set of relevant documents.
- Re-ranking: A specialized model re-orders the top documents from hybrid search for maximum relevance.
- 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.
Quick Check: Retrieval Steps
Which of the following statements accurately describe the roles of Hybrid Search and Re-ranking in a RAG system?
Recap: Advanced Retrieval
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.
よくある質問
「ハイブリッド検索と再ランキング」レッスンは無料ですか？
はい。「ハイブリッド検索と再ランキング」の完全なテキストはこのウェブで無料で読めます。インタラクティブに演習し（組み込みコードエディタと24時間対応のAIチューター）、LangChain / RAG / Vector DBsコースの残りをアンロックするには、CoddyKit PROにアップグレードしてください。 LangChain / RAG / Vector DBsコースには全4レッスンが含まれています。
「ハイブリッド検索と再ランキング」で何を学びますか？
キーワード検索とセマンティック検索を組み合わせ、再ランキングモデルで最も関連性の高いドキュメントを優先します。 ブラウザで直接実行するハンズオンコードでLangChain / RAG / Vector DBsを演習し、24時間対応のAIチューターがレッスンを進める中での質問に答えます。
LangChain / RAG / Vector DBsを始めるのに経験は必要ですか？
事前経験は必要ありません。CoddyKitのLangChain / RAG / Vector DBsは初級者から上級者向けに構成されているため、ここから始めるか最初から始めて、自分のペースで進むことができます。 これはレッスン3/4です。
「ハイブリッド検索と再ランキング」レッスンにはどのくらい時間がかかりますか？
ほとんどのCoddyKitレッスンは約5～10分かかります。各レッスンはコンパクトでインタラクティブなので、着実に進歩し、ウェブとアプリ全体で正確に前回の場所から再開できます。
このLangChain / RAG / Vector DBsレッスンでコードを書いて実行できますか？
はい。すべてのLangChain / RAG / Vector DBsレッスンに組み込みコードエディタが含まれているため、ブラウザでリアルコードを書いて実行し、即座のAIフィードバックを取得できます。ローカル設定は不要です。
このコースのすべてのレッスン
- マルチクエリ検索戦略
- LLMによるコンテキスト圧縮
- ハイブリッド検索と再ランキング
- 親ドキュメントと文ウィンドウによる検索