gh-giuseppe-trisciuoglio-developer-kit-developer-kit/skills/rag/references/retrieval-strategies.md

Advanced Retrieval Strategies

Overview

Different retrieval approaches for finding relevant documents in RAG systems, each with specific strengths and use cases.

Retrieval Approaches

1. Dense Retrieval

Method: Semantic similarity via embeddings Use Case: Understanding meaning and context Example: Finding documents about "machine learning" when query is "AI algorithms"

from langchain.vectorstores import Chroma

vectorstore = Chroma.from_documents(chunks, embeddings)
results = vectorstore.similarity_search("query", k=5)

2. Sparse Retrieval

Method: Keyword matching (BM25, TF-IDF) Use Case: Exact term matching and keyword-specific queries Example: Finding documents containing specific technical terms

from langchain.retrievers import BM25Retriever

bm25_retriever = BM25Retriever.from_documents(chunks)
bm25_retriever.k = 5
results = bm25_retriever.get_relevant_documents("query")

3. Hybrid Search

Method: Combine dense + sparse retrieval Use Case: Balance between semantic understanding and keyword matching

from langchain.retrievers import BM25Retriever, EnsembleRetriever

# Sparse retriever (BM25)
bm25_retriever = BM25Retriever.from_documents(chunks)
bm25_retriever.k = 5

# Dense retriever (embeddings)
embedding_retriever = vectorstore.as_retriever(search_kwargs={"k": 5})

# Combine with weights
ensemble_retriever = EnsembleRetriever(
    retrievers=[bm25_retriever, embedding_retriever],
    weights=[0.3, 0.7]
)

4. Multi-Query Retrieval

Method: Generate multiple query variations Use Case: Complex queries that can be interpreted in multiple ways

from langchain.retrievers.multi_query import MultiQueryRetriever

# Generate multiple query perspectives
retriever = MultiQueryRetriever.from_llm(
    retriever=vectorstore.as_retriever(),
    llm=OpenAI()
)

# Single query → multiple variations → combined results
results = retriever.get_relevant_documents("What is the main topic?")

5. HyDE (Hypothetical Document Embeddings)

Method: Generate hypothetical documents for better retrieval Use Case: When queries are very different from document style

# Generate hypothetical document based on query
hypothetical_doc = llm.generate(f"Write a document about: {query}")
# Use hypothetical doc for retrieval
results = vectorstore.similarity_search(hypothetical_doc, k=5)

Advanced Retrieval Patterns

Contextual Compression

Compress retrieved documents to only include relevant parts

from langchain.retrievers import ContextualCompressionRetriever
from langchain.retrievers.document_compressors import LLMChainExtractor

compressor = LLMChainExtractor.from_llm(llm)
compression_retriever = ContextualCompressionRetriever(
    base_compressor=compressor,
    base_retriever=vectorstore.as_retriever()
)

Parent Document Retriever

Store small chunks for retrieval, return larger chunks for context

from langchain.retrievers import ParentDocumentRetriever
from langchain.storage import InMemoryStore

store = InMemoryStore()
child_splitter = RecursiveCharacterTextSplitter(chunk_size=400)
parent_splitter = RecursiveCharacterTextSplitter(chunk_size=2000)

retriever = ParentDocumentRetriever(
    vectorstore=vectorstore,
    docstore=store,
    child_splitter=child_splitter,
    parent_splitter=parent_splitter
)

Retrieval Optimization Techniques

1. Metadata Filtering

Filter results based on document metadata

results = vectorstore.similarity_search(
    "query",
    filter={"category": "technical", "date": {"$gte": "2023-01-01"}},
    k=5
)

2. Maximal Marginal Relevance (MMR)

Balance relevance with diversity

results = vectorstore.max_marginal_relevance_search(
    "query",
    k=5,
    fetch_k=20,
    lambda_mult=0.5  # 0=max diversity, 1=max relevance
)

3. Reranking

Improve top results with cross-encoder

from sentence_transformers import CrossEncoder

reranker = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
candidates = vectorstore.similarity_search("query", k=20)
pairs = [[query, doc.page_content] for doc in candidates]
scores = reranker.predict(pairs)
reranked = sorted(zip(candidates, scores), key=lambda x: x[1], reverse=True)[:5]

Selection Guidelines

1. Query Type: Choose strategy based on typical query patterns
2. Document Type: Consider document structure and content
3. Performance Requirements: Balance quality vs speed
4. Domain Knowledge: Leverage domain-specific patterns
5. User Expectations: Match retrieval behavior to user expectations