How to filter answers from reader?

[SanjuktaK]

After training the basic QA pipeline it gives top_k answers and I want it to filter all the possible answers from all the documents that are used. How can I do that?

[julian-risch]

Hi @SanjuktaK there are three parameters of the FARMReader (same for the TransformersReader) that can help you. When you initialize a Reader you can set top_k, which is the maximum number of answers to return. Similarly, top_k_per_candidate and top_k_per_sample can be set to increase the number of answers that are returned at maximum from each document. Here is the docstring:

haystack/haystack/nodes/reader/farm.py

Line 86 in 509944f

:param top_k_per_candidate: How many answers to extract for each candidate doc that is coming from the retriever (might be a long text).

If you increase these values, for example to 20, more answers will be returned but the runtime will increase. As the default top_k is 10, you can expect the runtime to double if you set top_k to 20.

Note that more answers can be generated if the retriever returns more documents that are then passed to the reader. You might want to increase also the top_k of the retriever.