04. Multi-Strategy Retrieval

Production retrieval rarely relies on a single method. Different query types demand different strategies, and combining approaches improves recall.

Query classification determines which retrieval paths to activate. Transactional queries ("what is X") need precise definitions. Exploratory queries ("how does Y work") benefit from broad context. Factual queries ("when did Z happen") require exact matches.

Strategy routing assigns queries to retrieval methods:

• Dense retrieval handles semantic similarity and paraphrases
• Sparse retrieval handles exact terminology and technical terms
• Keyword matching handles product names, dates, identifiers

from enum import Enum
from typing import List, Callable

class QueryType(Enum):
    SEMANTIC = "semantic"      # Conceptual, how/why questions
    FACTUAL = "factual"        # Specific data, dates, names
    DEFINITIONAL = "definitional"  # What is X, definitions
    PROCEDURAL = "procedural"  # How to, steps, tutorials

class MultiStrategyRetriever:
    def __init__(self, retrievers: dict, classifier: Callable):
        self.retrievers = retrievers  # {strategy_name: retriever}
        self.classifier = classifier
    
    def retrieve(self, query: str, top_k: int = 10) -> List[dict]:
        query_type = self.classifier(query)
        
        # Activate relevant strategies
        strategies = self._get_strategies_for_type(query_type)
        
        results = []
        for strategy in strategies:
            retrieved = self.retrievers[strategy].search(query, top_k * 2)
            results.extend(retrieved)
        
        # Deduplicate and return combined results
        return self._deduplicate_and_rank(results, top_k)
    
    def _get_strategies_for_type(self, query_type: QueryType) -> List[str]:
        strategy_map = {
            QueryType.SEMANTIC: ["dense", "sparse"],
            QueryType.FACTUAL: ["sparse", "dense"],
            QueryType.DEFINITIONAL: ["dense"],
            QueryType.PROCEDURAL: ["dense", "sparse", "keyword"]
        }
        return strategy_map.get(query_type, ["dense"])

Local verification checkpoint

Run the smallest example from this chapter in a local workspace and record the package version, runtime, data path, and observed output. If the result depends on model size, vector count, CPU/GPU backend, or available memory, note that constraint beside the exercise so the lesson remains reproducible.

Local verification checkpoint

Run the smallest example from this chapter in a local workspace and record the package version, runtime, data path, and observed output. If the result depends on model size, vector count, CPU/GPU backend, or available memory, note that constraint beside the exercise so the lesson remains reproducible.