Skip to main content

​
Overview

RetrievalConfig controls how documents are retrieved and processed during the RAG query pipeline.

​
Definition

from dataclasses import dataclass

@dataclass
class RetrievalConfig:
    """Configuration for retrieval settings."""
    top_k: int = 5
    rerank_top_k: int = 3
    use_query_rewriting: bool = True
    use_reranking: bool = True
    use_hybrid_search: bool = False
    rrf_k: int = 60

​
Fields

​
top_k
int
default:"5"
Number of chunks to retrieve initially from the vector store
​
rerank_top_k
int
default:"3"
Number of chunks to keep after reranking
​
use_query_rewriting
bool
default:"True"
Whether to generate query variations for better retrieval
​
use_reranking
bool
default:"True"
Whether to rerank retrieved chunks for relevance
​
use_hybrid_search
bool
default:"False"
Whether to use hybrid search (semantic + BM25 keyword search)
​
rrf_k
int
default:"60"
RRF (Reciprocal Rank Fusion) constant for hybrid search result combination

​
Usage

​
Basic Usage

from mini import AgenticRAG, RetrievalConfig

rag = AgenticRAG(
    vector_store=vector_store,
    embedding_model=embedding_model,
    retrieval_config=RetrievalConfig(
        top_k=10,
        rerank_top_k=3
    )
)

​
Complete Configuration

retrieval_config = RetrievalConfig(
    top_k=15,
    rerank_top_k=5,
    use_query_rewriting=True,
    use_reranking=True,
    use_hybrid_search=True,
    rrf_k=60
)

rag = AgenticRAG(
    vector_store=vector_store,
    embedding_model=embedding_model,
    retrieval_config=retrieval_config
)

​
Configuration Strategies

​
Fast and Focused

For quick, focused answers: 
retrieval_config = RetrievalConfig(
    top_k=5,
    rerank_top_k=2,
    use_query_rewriting=False,
    use_reranking=True
)

​
Comprehensive

For thorough, well-sourced answers: 
retrieval_config = RetrievalConfig(
    top_k=20,
    rerank_top_k=5,
    use_query_rewriting=True,
    use_reranking=True
)
For queries with specific keywords or technical terms: 
retrieval_config = RetrievalConfig(
    top_k=10,
    rerank_top_k=3,
    use_hybrid_search=True,
    use_query_rewriting=True,
    use_reranking=True
)

​
Semantic Only

For purely conceptual queries: 
retrieval_config = RetrievalConfig(
    top_k=10,
    rerank_top_k=3,
    use_hybrid_search=False,
    use_query_rewriting=True
)

​
Parameter Guidelines

​
top_k

Number of chunks to retrieve initially:
  • 5-10: Fast, focused answers
  • 10-15: Balanced (recommended default)
  • 15-25: Comprehensive, thorough answers
  • >25: May include noise, slower

​
rerank_top_k

Number of chunks to keep after reranking:
  • 2-3: Concise answers
  • 3-5: Balanced (recommended)
  • 5-10: Detailed answers with multiple sources
  • Should always be ≤ top_k

​
use_query_rewriting

Generate query variations for better retrieval:
  • True: Better recall, finds more relevant content (recommended)
  • False: Faster, uses only original query

​
use_reranking

Rerank retrieved chunks by relevance:
  • True: Higher quality results (recommended)
  • False: Faster, relies on initial ranking
Combine semantic and keyword search:
  • True: Better for specific terms, technical content
  • False: Faster, semantic-only search (default)

​
rrf_k

RRF fusion constant for hybrid search:
  • 20-40: More weight to top results
  • 60: Balanced (default)
  • 70-80: More even distribution

​
Runtime Override

You can override config values when querying: 
# Default configuration
rag = AgenticRAG(
    vector_store=vector_store,
    embedding_model=embedding_model,
    retrieval_config=RetrievalConfig(top_k=10, rerank_top_k=3)
)

# Override for specific query
response = rag.query(
    "What is this about?",
    top_k=20,        # Override default
    rerank_top_k=5   # Override default
)

​
Performance Trade-offs

ConfigurationSpeedQualityCost
top_k=5, no rewriting⚡⚡⚡⭐⭐💰
top_k=10, with reranking⚡⚡⭐⭐⭐⭐💰💰
top_k=20, full features⚡⭐⭐⭐⭐⭐💰💰💰

​
Complete Example

from mini import AgenticRAG, LLMConfig, RetrievalConfig, RerankerConfig

# Production configuration
rag = AgenticRAG(
    vector_store=vector_store,
    embedding_model=embedding_model,
    llm_config=LLMConfig(
        model="gpt-4o-mini",
        temperature=0.7
    ),
    retrieval_config=RetrievalConfig(
        top_k=15,
        rerank_top_k=5,
        use_query_rewriting=True,
        use_reranking=True,
        use_hybrid_search=True,
        rrf_k=60
    ),
    reranker_config=RerankerConfig(type="cohere")
)

# Query with default config
response = rag.query("What are the key findings?")

# Query with custom retrieval
response = rag.query(
    "Specific technical question?",
    top_k=25,
    rerank_top_k=10
)

​
See Also

LLMConfig

Configure language model

RerankerConfig

Configure reranking

Hybrid Search

Learn about hybrid search