How to implement a chain of multiple retrievers?

[nasrin-taghizadeh]

Question
I need to utilize multiple retrievers in a chain, i.e. the first one retrieves some documents from data store and pass them to the second one. Haystack provided JoinDocuments node which concatenates or merges output of multiple retrievers but it doesn't fit to my need. Any idea is appreciated.

[masci]

Hi @nasrin-taghizadeh

retrievers only support Documentstores as their input and we don't have plans to change this, but depending on your use case, you might find helpful using a ranker, specifically the SentenceTransformersRanker. This node can take the input from (e.g.) a BM25 retriever and use a transformer model to sort its results. We have an example in our docs (I'm linking the code because the documentation website has a problem rendering this specific snippet):

    retriever = BM25Retriever(document_store=document_store)
    ranker = SentenceTransformersRanker(model_name_or_path="cross-encoder/ms-marco-MiniLM-L-12-v2")
    p = Pipeline()
    p.add_node(component=retriever, name="ESRetriever", inputs=["Query"])
    p.add_node(component=ranker, name="Ranker", inputs=["ESRetriever"])

[nasrin-taghizadeh]

Thank you very much.

[nasrin-taghizadeh]

Dear Sir,

For sage of performance, Is it possible to use SentenceTransformersRanker with pre-saved vectors in DocumentStore? Of course for documents.

[tstadel]

Hey @nasrin-taghizadeh,
SentenceTransformersRanker uses Cross-Encoders, which always need Document + Query as an input. So you cannot precompute document vectors unless you know the query in advance. This is fundamentally different to Bi-Encoders (e.g. used by EmbeddingRetriever or DensePassageRetriever) which calculate Query and Document vectors separately and use a similarity operator like cosine or dot product on top to obtain the relevance score. See https://www.sbert.net/examples/applications/cross-encoder/README.html for more infos.

[nasrin-taghizadeh]

I was very thankful you are considering my problem.