The ElasticsearchDocumentStore and MilvusDocumentStore rely on Elasticsearch and Milvus backend services which persist after your Python script has finished running. If you rerun your script without deleting documents, you could end up with duplicate copies of your documents in your database. The easiest way to avoid this is to call
DocumentStore.delete_documents() after initialization to ensure that you are working with an empty DocumentStore.
DocumentStores also have a
duplicate_documents argument in their
write_documents methods where you can define whether you'd like skip writing duplicates, overwrite existing duplicates or raise an error when there are duplicates.
You will want to ensure that a CUDA enabled GPU is being engaged when Haystack is running (you can check by running
nvidia-smi -l on your command line). Components which can be sped up by GPU have a
use_gpu argument in their constructor which you will want to set to
There are many different ways to speed up the performance of your Haystack system.
The Reader is usually the most computationally expensive component in a pipeline and you can often speed up your system by using a smaller model, like
deepset/minilm-uncased-squad2 (see Benchmarks). This usually comes with a small trade-off in accuracy.
You can reduce the work load on the Reader by instructing the Retriever to pass on less documents. This is done by setting the
top_k_retriever parameter to a lower value.
Making sure that your documents are shorter can also increase the speed of your system. You can split your documents into smaller chunks by using the
PreProcessor (see How to Preprocess Documents).
The components in Haystack, such as the
Retriever or the
Reader, are designed in a language agnostic way. However you may have to set certain parameters or load models pretrained for your language in order to get good performance out of Haystack. See Languages Other Than English for more details.
When providing your documents in the input format you can provide metadata information as a dictionary under the
meta key. At query time, you can provide a
filters argument (most likely through
Pipeline.run()) that specifies the accepted values for a certain metadata field (see Metadata Filtering).
To find out more about how to debug a pipeline, have a look at Returning Debugging Information.
Haystack models can be wrapped in a REST API. For basic details on how to set this up, have a look at REST API.
The confidence scores are in the range of 0 and 1 and reflect how confident the model is in each prediction that it makes. Having a confidence score is particularly useful in cases where you need Haystack to work with a certain accuracy threshold. Many of our users have built systems where predictions below a certain confidence value are routed on to a fallback system.
For more information on model confidence and how to tune it, please refer to Confidence Scores.
My Documents aren't showing up in my DocumentStore even though I've called
When indexing, retrieving or querying for documents from a DocumentStore, you can specify an
index on which to perform this action. This can be specified in almost all methods of
DocumentStore as well as
Retriever.retrieve(). Ensure that you are performing these operations on the one index! Note that this also applies at evaluation where labels are written into their own separate DocumentStore index.
In short, the FARMReader uses a QA pipeline implementation that comes from our own FARM framework that we can more easily update and also optimize for performance. By contrast, the TransformersReader uses a QA pipeline implementation that comes from HuggingFace's Transformers. See Deeper Dive: FARM vs Transformers for a more details about their differences!
The default logging level in Haystack is set to WARNING. This means messages on the INFO level are not logged. For example, if you call
FARMReader.train(), the evaluation reports won't be shown in the console output by default. To see logs at INFO level, run the following code before your first import of Haystack:
import logging logging.basicConfig(format="%(levelname)s - %(name)s - %(message)s", level=logging.WARNING) logging.getLogger("haystack").setLevel(logging.INFO)
Updated 2 months ago