Retrieval

Retrieval finds relevant chunks for a given user query. Search Toolkit supports vector (semantic), keyword (BM25), and hybrid search strategies that can be combined for optimal results.

An optional query preprocessor transforms the raw query before search — rewriting it for clarity or expanding it into multiple variants.

One or more Retrievers execute the search against the index. They can run in parallel and their results are merged.

An optional Reranker re-scores the merged results using a more precise scoring strategy — an LLM, a cross-encoder, or rank fusion across multiple result sets.

QueryEngine orchestrates the retrieval pipeline. It accepts one or more retrievers, optional query preprocessing, and optional rerankers:

from mistralai.search.toolkit.retrieval import QueryEngine
from mistralai.search.toolkit.retrieval.retrievers import VectorRetriever
from mistralai.search.toolkit.retrieval.rerankers import LLMReRanker
from mistralai.search.toolkit.retrieval.pre_processors import LLMQueryRewriter

query_engine = QueryEngine(
    retriever=vector_retriever,       # Also accepts a list of retrievers
    query_rewriter=query_rewriter,    # Optional
    rerankers=[llm_reranker],         # Optional, supports ReRanker and GroupedRanker
)

result = await query_engine.search(
    query="What is RAG?",
    top_k=10,
    include_metadata=True,
    include_content=True,
)

print(f"Original query: {result.original_query}")
print(f"Results: {len(result.results)}")

from mistralai.search.toolkit.retrieval import QueryEngine
from mistralai.search.toolkit.retrieval.retrievers import VectorRetriever
from mistralai.search.toolkit.retrieval.rerankers import LLMReRanker
from mistralai.search.toolkit.retrieval.pre_processors import LLMQueryRewriter

query_engine = QueryEngine(
    retriever=vector_retriever,       # Also accepts a list of retrievers
    query_rewriter=query_rewriter,    # Optional
    rerankers=[llm_reranker],         # Optional, supports ReRanker and GroupedRanker
)

result = await query_engine.search(
    query="What is RAG?",
    top_k=10,
    include_metadata=True,
    include_content=True,
)

print(f"Original query: {result.original_query}")
print(f"Results: {len(result.results)}")

Each retrieval component is documented in detail with examples and best practices:

• Retrievers — VectorRetriever, KeywordRetriever, and hybrid search patterns
• Rerankers — LLMReRanker, CrossEncoderReRanker, RRF fusion, and custom rerankers
• Query preprocessing — Query rewriting and expansion for better retrieval quality
• Semantic cache — Cache results by query similarity to reduce latency