Optimization

Speeding up Reader

In many pipelines, the Reader will be the most computationally expensive component. If this is a step that you would like to speed up, you can opt for a smaller Reader model that can process more passages in the same amount of time.

On Benchmarks, you will find a comparison of many of the common model architectures. While our default recommendation is RoBERTa, MiniLM offers much faster processing for only a minimal drop in accuracy. If you are interested in compressing your large models into smaller ones similar to TinyBERT, you can also take a look at our Model Distillation guide. You can find the models that we've trained on the HuggingFace Model Hub.

GPU acceleration

The transformer based models used in Haystack are designed to be run on a GPU enabled machine. The design of these models means that they greatly benefit from the parallel processing capabilities of graphics cards. If Haystack has successfully detected a graphics card, you should see these lines in your console output.

INFO - farm.utils -   Using device: CUDA
INFO - farm.utils -   Number of GPUs: 1

You can track the work load on your CUDA enabled Nvidia GPU by tracking the output of nvidia-smi -l on the command line while your Haystack program is running.

Document Length

Document length has a very direct impact on the speed of the Reader which is why we recommend using the PreProcessor class to clean and split your documents. If you halve the length of your Documents, you will halve the workload placed onto your Reader.

For sparse Retrievers, very long documents pose a challenge since the signal of the relevant section of text can get washed out by the rest of the Document. To get a good balance between Reader speed and Retriever performance, we recommend splitting documents to a maximum of 500 words. If there is no Reader in the pipeline following the Retriever, we recommend that Documents be no longer than 10,000 words.

Dense Retrievers are limited in the length of text that they can read in one pass. As such, it is important that Documents are not longer than the dense Retriever's maximum input length. By default, Haystack's DensePassageRetriever model has a maximum length of 256 tokens. As such, we recommend that Documents contain significantly less words. We have found decent performance with Documents around 100 words long.

Respecting Sentence Boundaries

When splitting Documents, it is generally not a good idea to let document boundaries fall in the middle of sentences. Doing so means that each document will contain incomplete sentence fragments which may be hard for both Retriever and Reader to interpret. It is therefore recommended to set split_respect_sentence_boundary=True when initializing your PreProcessor.

Choosing the Right top-k Values

The top-k parameter in both the Retriever and Reader determine how many results they return. More specifically, Retriever top-k dictates how many retrieved Documents are passed on to the next stage, while Reader top-k determines how many Answer candidates to show.

In our experiments, we have found that Retriever top_k=10 gives decent overall performance and so we have set this as the default in Haystack.

The choice of Retriever top-k is a trade-off between speed and accuracy, especially when there is a Reader in the pipeline. Setting it higher means passing more documents to the Reader, thus reducing the chance that the answer-containing passage is missed. However, passing more documents to the Reader will create a larger workload for the component.

To set the top_k of your Retriever and Reader in a pipeline, run:

answers = pipeline.run(
    query="What did Einstein work on?", 
    params={
        "retriever": {"top_k": 10}, 
        "reader": {"top_k": 5}
    }
)

If you are directly calling the Retriever, run:

retrieved_docs = retriever.retrieve(top_k=10)