Search every document in doc store for the ques vs the top semantic matches

[hepbc]

Hi Guys,

Thanks for your help with this lib! I am just finding my way around it. I want to get the the (top k) responses to the question from every document in the document store instead of the retriever taking the top instances across all documents. My use case is more in the nature of data retrieval from periodic documents.

Would appreciate some pointers around this. Thanks!
BC

[lalitpagaria]

Tagging @tholor @brandenchan @oryx1729

[brandenchan]

Hi @hepbc, cool to hear you are using Haystack this way! This Information extraction style of QA is something we see a lot of potential in!

So in this style of QA, you obviously want to iterate over each Q in your set of Qs. In terms of Docs, you also want to iterate over each Doc but make sure you're performing the query on just that single document and not your full set of Docs!

To do this, you need to perform some metadata filtering (see this for a basic intro) . The idea is that you attach a unique piece of meta data to Document, and at query time, you isolate a single document to perform the query on by setting a filter.

In a very rough sketch, you will want something like this

# Document preparation and indexing
docs = [
    {
        'text': "some text here...",
        'meta': {'doc_id': 12345, ...}
    }, ...
]
document_store.write_documents(docs)

# Query time
for q in queries:
    for doc_id in doc_ids:
        result = pipeline.run(
            query=q,
            filters={"doc_id":[doc_id]}
        )

Let me know how this goes for you!

All the best,
From a fellow BC

[hepbc]

Thanks BC 👍 Will try this and get back.

[hepbc]

Hi @brandenchan
Have been trying to follow the steps you outlined above:

##create the doc store
##docs have meta_data = name
document_store = InMemoryDocumentStore()
document_store.write_documents(docs)

##create pipeline
retriever = TfidfRetriever(document_store=document_store)
reader = FARMReader(model_name_or_path="deepset/roberta-base-squad2", use_gpu=False)
pipe = ExtractiveQAPipeline(reader, retriever)

##defined names - list of docs I want to run query on and query
for name in names:
      prediction = pipe.run(
                            query=question,
                            filters={"name":[name]})
      print_answers(prediction, details="minimal")

However this gives me an error: NotImplementedError: Filters are not implemented in TfidfRetriever.
Tried doing it with the BaseRetriever class as well, but that does not work too.

What am I missing here? Appreciate your help.
Thanks

[brandenchan]

Hey, so currently not all Retrievers support the filtering. The ElasticsearchDocumentStore + ElasticsearchRetriever are designed to perform this filtering, and so can the WeaviateDocumentStore. Are either of these options for you?