Retrieval Pipelines

Algorithms that take a query and return relevant documents.

Available Pipelines

Pipeline	Algorithm	Modality
BM25	Sparse (term frequency)	Text
Hybrid	RRF / Convex Combination	Text
GQR Hybrid	Guided Query Refinement over hybrid candidates	Text
Vector Search	Dense (vector similarity)	Text
HyDE	Dense (hypothetical document embeddings)	Text
Query Rewrite	Rewrite query text before retrieval	Text
RETRO*	Rubric-based LLM reranking over retrieved candidates	Text
Power of Noise	Retrieval wrapper + seeded noise injection	Text

Base Class

All retrieval pipelines extend BaseRetrievalPipeline:

from autorag_research.pipelines.retrieval import BaseRetrievalPipeline


class MyRetrievalPipeline(BaseRetrievalPipeline):
    def _get_retrieval_func(self):
        def retrieve(queries: list[str], top_k: int) -> list[list[dict]]:
            # Return list of results per query
            # Each result: {"doc_id": ..., "score": ...}
            pass

        return retrieve

    def _get_pipeline_config(self):
        return {"type": "my_pipeline"}

Methods

Method	Description
retrieve(query, top_k)	Single query retrieval
run(top_k, batch_size)	Batch retrieval for all queries