クエリ変換の技術
ユーザーのクエリを言い換えたり拡張したりして、ベクトルデータベースからより効果的に検索する方法を学びます。
「クエリ変換の技術」はCoddyKit上の無料Vector Databases: Pinecone, Weaviate & pgvectorレッスンです。 これはレッスン1/4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Vector Databases: Pinecone, Weaviate & pgvector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。 Vector Databases: Pinecone, Weaviate & pgvector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。
What's Query Transformation?
In Retrieval Augmented Generation (RAG), the quality of the information retrieved from your vector database directly impacts the LLM's response. Sometimes, the raw user query isn't ideal for retrieval.
Query transformation is the process of modifying a user's original query to make it more effective for searching your vector database.
Why Original Queries Fall Short
User queries can be:
- Too short or vague: Lacking enough detail for precise retrieval.
- Ambiguous: Open to multiple interpretations.
- Colloquial: Using informal language that doesn't match document embeddings.
- Complex: Asking multiple questions at once.
These issues lead to irrelevant context being retrieved, impacting RAG quality.
Bridging the Semantic Gap
Your vector database stores information as numerical vectors (embeddings). For effective retrieval, the query's embedding needs to be semantically similar to the embeddings of relevant documents.
Query transformation helps by:
- Adding more context.
- Clarifying intent.
- Aligning query terms with document vocabulary.
This "bridges the semantic gap" for better matches.
Technique 1: Query Expansion
Query expansion involves adding more terms or phrases to the original query. This makes the search broader and increases the chances of hitting relevant documents.
Common methods include:
- Adding synonyms.
- Including related concepts.
- Using domain-specific jargon.
This technique is especially useful for short, vague queries.
Simple Keyword Expansion
Here's a basic Python example where we manually expand a query with related terms. In real systems, an LLM might generate these expansions automatically.
def expand_query_keywords(query):
expansion_map = {
"AI": ["artificial intelligence", "machine learning", "deep learning"],
"vector DB": ["vector database", "embedding store", "Pinecone", "Weaviate"]
}
expanded_terms = []
for keyword, additions in expansion_map.items():
if keyword.lower() in query.lower():
expanded_terms.extend(additions)
return query + " " + " ".join(expanded_terms)
# Example usage
user_query = "What is AI?"
transformed_query = expand_query_keywords(user_query)
print(f"Original: {user_query}")
print(f"Transformed: {transformed_query}")Technique 2: Query Rewriting/Rephrasing
Query rewriting (or rephrasing) uses an LLM to generate a completely new query that is clearer, more specific, or better aligned with the expected content of your documents.
This is particularly effective for:
- Ambiguous questions.
- Conversational queries.
- Queries that implicitly refer to past turns in a conversation.
Rewriting with an LLM
This Python snippet demonstrates how you might use an LLM to rephrase a user query. In a real application, llm_api_call would interact with models like OpenAI GPT or a local LLM.
import os
# Placeholder for an actual LLM API call
def llm_api_call(prompt):
# For demonstration, we'll simulate a response.
if "rephrase" in prompt.lower() and "vector databases" in prompt.lower():
return "Explain the core components and function of vector databases."
return "Could not rephrase query effectively."
def rephrase_query_with_llm(original_query):
prompt = f"Rephrase the following user query to be more effective for searching a technical documentation database about vector databases: '{original_query}'"
rephrased_query = llm_api_call(prompt)
return rephrased_query
# Example usage
user_query = "Tell me about vector DBs."
transformed_query = rephrase_query_with_llm(user_query)
print(f"Original: {user_query}")
print(f"Transformed: {transformed_query}")Technique 3: Sub-Query Generation
When a user asks a complex question that involves multiple aspects, a single query might not retrieve all necessary information. Sub-query generation breaks down a complex query into several simpler, more focused queries.
Each sub-query can then be used to retrieve specific pieces of context, which are then combined for the LLM.
Generating Sub-Queries
Here's a conceptual Python example for generating sub-queries. An LLM is often used to parse the original query and identify distinct questions.
import os
# Placeholder for an actual LLM API call
def llm_api_call_sub_query(prompt):
if "break down" in prompt.lower() and "Pinecone" in prompt.lower():
return ["What is Pinecone?", "How do I upsert data into Pinecone?", "What is a Pinecone index?"]
return ["Could not break down query."]
def generate_sub_queries(original_query):
prompt = f"Break down the following complex user query into a list of simpler, distinct questions for retrieving information from a vector database: '{original_query}'"
sub_queries = llm_api_call_sub_query(prompt)
return sub_queries
# Example usage
user_query = "How do I use Pinecone for storing and querying embeddings?"
transformed_queries = generate_sub_queries(user_query)
print(f"Original: {user_query}")
print(f"Sub-queries: {transformed_queries}")Hybrid & Contextual Transformations
Effective RAG often combines these techniques. You might first rephrase a query, then expand it. Also, consider the conversational context.
- Multi-turn awareness: Use previous turns to enrich the current query.
- Hybrid approaches: Combine transformations with traditional keyword search.
Experimentation is key to finding what works best for your data!
Understanding Query Transformation
Which of the following scenarios would MOST benefit from using query expansion as a transformation technique?
Recap: Transforming Queries
You've learned that query transformation is crucial for improving RAG performance by making user queries more effective for vector database retrieval.
- Query Expansion: Adds terms to broaden search.
- Query Rewriting: Rephrases for clarity and specificity.
- Sub-Query Generation: Breaks down complex queries.
Mastering these techniques will significantly enhance the quality of your RAG applications!
よくある質問
「クエリ変換の技術」レッスンは無料ですか？
はい。「クエリ変換の技術」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）、Vector Databases: Pinecone, Weaviate & pgvector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。 Vector Databases: Pinecone, Weaviate & pgvectorコースには全4レッスンが含まれています。
「クエリ変換の技術」で何を学びますか？
ユーザーのクエリを言い換えたり拡張したりして、ベクトルデータベースからより効果的に検索する方法を学びます。 ブラウザで直接実行するハンズオンコードでVector Databases: Pinecone, Weaviate & pgvector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。
Vector Databases: Pinecone, Weaviate & pgvector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Vector Databases: Pinecone, Weaviate & pgvector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1/4です。
「クエリ変換の技術」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Vector Databases: Pinecone, Weaviate & pgvector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Vector Databases: Pinecone, Weaviate & pgvector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。