Confusion about return types of retrievers

[wxywb]

when we set a rm in dspy, and we call the dspy.Retrieve, the following lines will be executed

dspy/dsp/primitives/search.py

Lines 8 to 24 in b05394b

	def retrieve(query: str, k: int, **kwargs) -> list[str]:
	"""Retrieves passages from the RM for the query and returns the top k passages."""
	if not dsp.settings.rm:
	raise AssertionError("No RM is loaded.")
	passages = dsp.settings.rm(query, k=k, **kwargs)
	if not isinstance(passages, Iterable):
	# it's not an iterable yet; make it one.
	# TODO: we should unify the type signatures of dspy.Retriever
	passages = [passages]
	passages = [psg.long_text for psg in passages]
	if dsp.settings.reranker:
	passages_cs_scores = dsp.settings.reranker(query, passages)
	passages_cs_scores_sorted = np.argsort(passages_cs_scores)[::-1]
	passages = [passages[idx] for idx in passages_cs_scores_sorted]
	return passages

so we can see in L12, the passages return by retrievers should be a list of dotdict which has long_text attribute.

But we can find the base class of retrievers should return Prediction

dspy/dspy/retrieve/retrieve.py

Line 32 in b05394b

def forward(self, query_or_queries: Union[str, List[str]], k: Optional[int] = None,**kwargs) -> Prediction:

the official example which uses dspy.ColBERTv2 returns List[dotdict] instead of Prediction (although it is not a dspy.Retrieve)

dspy/dsp/modules/colbertv2.py

Line 35 in b05394b

return [dotdict(psg) for psg in topk]

and we now have two kinds of retrievers, one return List[dotdict]

• azureaisearch_rm.py
• chromadb_rm.py
• clarifai_rm.py
• vectara_rm.py

another one return dspy.Prediction

• databricks_rm.py
• mongodb_atlas_rm.py
• pinecone_rm.py
• ragatouille_rm.py