In this tutorial, you will learn how the
Pipeline class acts as a connector between all the different
building blocks that are found in FARM. Whether you are using a Reader, Generator, Summarizer
or Retriever (or 2), the
Pipeline class will help you build a Directed Acyclic Graph (DAG) that
determines how to route the output of one component into the input of another.
Let's start by ensuring we have a GPU running to ensure decent speed in this tutorial. In Google colab, you can change to a GPU runtime in the menu:
- Runtime -> Change Runtime type -> Hardware accelerator -> GPU
# Make sure you have a GPU running !nvidia-smi
These lines are to install Haystack through pip
# Install the latest release of Haystack in your own environment #! pip install farm-haystack # Install the latest master of Haystack !pip install grpcio-tools==1.34.1 !pip install --upgrade git+https://github.com/deepset-ai/haystack.git # Install pygraphviz !apt install libgraphviz-dev !pip install pygraphviz
If running from Colab or a no Docker environment, you will want to start Elasticsearch from source
# In Colab / No Docker environments: Start Elasticsearch from source ! wget https://artifacts.elastic.co/downloads/elasticsearch/elasticsearch-7.9.2-linux-x86_64.tar.gz -q ! tar -xzf elasticsearch-7.9.2-linux-x86_64.tar.gz ! chown -R daemon:daemon elasticsearch-7.9.2 import os from subprocess import Popen, PIPE, STDOUT es_server = Popen(['elasticsearch-7.9.2/bin/elasticsearch'], stdout=PIPE, stderr=STDOUT, preexec_fn=lambda: os.setuid(1) # as daemon ) # wait until ES has started ! sleep 30
Here are some core imports
from haystack.utils import print_answers, print_documents
Then let's fetch some data (in this case, pages from the Game of Thrones wiki) and prepare it so that it can
be used indexed into our
from haystack.preprocessor.utils import fetch_archive_from_http, convert_files_to_dicts from haystack.preprocessor.cleaning import clean_wiki_text #Download and prepare data - 517 Wikipedia articles for Game of Thrones doc_dir = "data/article_txt_got" s3_url = "https://s3.eu-central-1.amazonaws.com/deepset.ai-farm-qa/datasets/documents/wiki_gameofthrones_txt.zip" fetch_archive_from_http(url=s3_url, output_dir=doc_dir) # convert files to dicts containing documents that can be indexed to our datastore got_dicts = convert_files_to_dicts( dir_path=doc_dir, clean_func=clean_wiki_text, split_paragraphs=True )
Here we initialize the core components that we will be gluing together using the
We have a
ElasticsearchRetriever and a
These can be combined to create a classic Retriever-Reader pipeline that is designed
to perform Open Domain Question Answering.
from haystack import Pipeline from haystack.utils import launch_es from haystack.document_store import ElasticsearchDocumentStore from haystack.retriever.sparse import ElasticsearchRetriever from haystack.retriever.dense import DensePassageRetriever from haystack.reader import FARMReader # Initialize DocumentStore and index documents launch_es() document_store = ElasticsearchDocumentStore() document_store.delete_all_documents() document_store.write_documents(got_dicts) # Initialize Sparse retriever es_retriever = ElasticsearchRetriever(document_store=document_store) # Initialize dense retriever dpr_retriever = DensePassageRetriever(document_store) document_store.update_embeddings(dpr_retriever, update_existing_embeddings=False) reader = FARMReader(model_name_or_path="deepset/roberta-base-squad2")
Haystack features many prebuilt pipelines that cover common tasks.
Here we have an
ExtractiveQAPipeline (the successor to the now deprecated
from haystack.pipeline import ExtractiveQAPipeline # Prebuilt pipeline p_extractive_premade = ExtractiveQAPipeline(reader=reader, retriever=es_retriever) res = p_extractive_premade.run( query="Who is the father of Arya Stark?", top_k_retriever=10, top_k_reader=5 ) print_answers(res, details="minimal")
If you want to just do the retrieval step, you can use a
from haystack.pipeline import DocumentSearchPipeline p_retrieval = DocumentSearchPipeline(es_retriever) res = p_retrieval.run( query="Who is the father of Arya Stark?", top_k_retriever=10 ) print_documents(res, max_text_len=200)
Or if you want to use a
Generator instead of a
you can initialize a
GenerativeQAPipeline like this:
from haystack.pipeline import GenerativeQAPipeline, FAQPipeline from haystack.generator import RAGenerator # We set this to True so that the document store returns document embeddings with each document # This is needed by the Generator document_store.return_embedding = True # Initialize generator rag_generator = RAGenerator() # Generative QA p_generator = GenerativeQAPipeline(generator=rag_generator, retriever=dpr_retriever) res = p_generator.run( query="Who is the father of Arya Stark?", top_k_retriever=10 ) print_answers(res, details="minimal") # We are setting this to False so that in later pipelines, # we get a cleaner printout document_store.return_embedding = False
Haystack features prebuilt pipelines to do:
- just document search (DocumentSearchPipeline),
- document search with summarization (SearchSummarizationPipeline)
- generative QA (GenerativeQAPipeline)
- FAQ style QA (FAQPipeline)
- translated search (TranslationWrapperPipeline) To find out more about these pipelines, have a look at our documentation
With any Pipeline, whether prebuilt or custom constructed, you can save a diagram showing how all the components are connected.
p_extractive_premade.draw("pipeline_extractive_premade.png") p_retrieval.draw("pipeline_retrieval.png") p_generator.draw("pipeline_generator.png")
Now we are going to rebuild the
ExtractiveQAPipelines using the generic Pipeline class.
We do this by adding the building blocks that we initialized as nodes in the graph.
# Custom built extractive QA pipeline p_extractive = Pipeline() p_extractive.add_node(component=es_retriever, name="Retriever", inputs=["Query"]) p_extractive.add_node(component=reader, name="Reader", inputs=["Retriever"]) # Now we can run it res = p_extractive.run( query="Who is the father of Arya Stark?", top_k_retriever=10, top_k_reader=5 ) print_answers(res, details="minimal") p_extractive.draw("pipeline_extractive.png")
Pipelines offer a very simple way to ensemble together different components.
In this example, we are going to combine the power of a
with the keyword based
See our documentation to understand why
we might want to combine a dense and sparse retriever.
Here we use a
JoinDocuments node so that the predictions from each retriever can be merged together.
from haystack.pipeline import JoinDocuments # Create ensembled pipeline p_ensemble = Pipeline() p_ensemble.add_node(component=es_retriever, name="ESRetriever", inputs=["Query"]) p_ensemble.add_node(component=dpr_retriever, name="DPRRetriever", inputs=["Query"]) p_ensemble.add_node(component=JoinDocuments(join_mode="concatenate"), name="JoinResults", inputs=["ESRetriever", "DPRRetriever"]) p_ensemble.add_node(component=reader, name="Reader", inputs=["JoinResults"]) p_ensemble.draw("pipeline_ensemble.png") # Run pipeline res = p_ensemble.run( query="Who is the father of Arya Stark?", top_k_retriever=5 #This is top_k per retriever ) print_answers(res, details="minimal")
Nodes are relatively simple objects and we encourage our users to design their own if they don't see on that fits their use case
The only requirements are:
- Add a method run(self, **kwargs) to your class. **kwargs will contain the output from the previous node in your graph.
- Do whatever you want within run() (e.g. reformatting the query)
- Return a tuple that contains your output data (for the next node) and the name of the outgoing edge (by default "output_1" for nodes that have one output)
- Add a class attribute outgoing_edges = 1 that defines the number of output options from your node. You only need a higher number here if you have a decision node (see below).
Here we have a template for a Node:
class NodeTemplate(): outgoing_edges = 1 def run(self, **kwargs): # Insert code here to manipulate the variables in kwarg return (kwargs, "output_1")
Decision Nodes help you route your data so that only certain branches of your
Pipeline are run.
One popular use case for such query classifiers is routing keyword queries to Elasticsearch and questions to DPR + Reader.
With this approach you keep optimal speed and simplicity for keywords while going deep with transformers when it's most helpful.
Though this looks very similar to the ensembled pipeline shown above, the key difference is that only one of the retrievers is run for each request. By contrast both retrievers are always run in the ensembled approach.
Below, we define a very naive
QueryClassifier and show how to use it:
class QueryClassifier(): outgoing_edges = 2 def run(self, **kwargs): if "?" in kwargs["query"]: return (kwargs, "output_2") else: return (kwargs, "output_1") # Here we build the pipeline p_classifier = Pipeline() p_classifier.add_node(component=QueryClassifier(), name="QueryClassifier", inputs=["Query"]) p_classifier.add_node(component=es_retriever, name="ESRetriever", inputs=["QueryClassifier.output_1"]) p_classifier.add_node(component=dpr_retriever, name="DPRRetriever", inputs=["QueryClassifier.output_2"]) p_classifier.add_node(component=reader, name="QAReader", inputs=["ESRetriever", "DPRRetriever"]) p_classifier.draw("pipeline_classifier.png") # Run only the dense retriever on the full sentence query res_1 = p_classifier.run( query="Who is the father of Arya Stark?", top_k_retriever=10 ) print("DPR Results" + "\n" + "="*15) print_answers(res_1) # Run only the sparse retriever on a keyword based query res_2 = p_classifier.run( query="Arya Stark father", top_k_retriever=10 ) print("ES Results" + "\n" + "="*15) print_answers(res_2)
We have also designed a set of nodes that can be used to evaluate the performance of a system. Have a look at our tutorial to get hands on with the code and learn more about Evaluation Nodes!
Pipeline can be defined in a YAML file and simply loaded.
Having your pipeline available in a YAML is particularly useful
when you move between experimentation and production environments.
Just export the YAML from your notebook / IDE and import it into your production environment.
It also helps with version control of pipelines,
allows you to share your pipeline easily with colleagues,
and simplifies the configuration of pipeline parameters in production.
It consists of two main sections: you define all objects (e.g. a reader) in components and then stick them together to a pipeline in pipelines. You can also set one component to be multiple nodes of a pipeline or to be a node across multiple pipelines. It will be loaded just once in memory and therefore doesn't hurt your resources more than actually needed.
The contents of a YAML file should look something like this:
version: '0.7' components: # define all the building-blocks for Pipeline - name: MyReader # custom-name for the component; helpful for visualization & debugging type: FARMReader # Haystack Class name for the component params: no_ans_boost: -10 model_name_or_path: deepset/roberta-base-squad2 - name: MyESRetriever type: ElasticsearchRetriever params: document_store: MyDocumentStore # params can reference other components defined in the YAML custom_query: null - name: MyDocumentStore type: ElasticsearchDocumentStore params: index: haystack_test pipelines: # multiple Pipelines can be defined using the components from above - name: my_query_pipeline # a simple extractive-qa Pipeline nodes: - name: MyESRetriever inputs: [Query] - name: MyReader inputs: [MyESRetriever]
To load, simply call:
The possibilities are endless with the
Pipeline class and we hope that this tutorial will inspire you
to build custom pipeplines that really work for your use case!
We bring NLP to the industry via open source!
Our focus: Industry specific language models & large scale QA systems.
Some of our other work:
By the way: we're hiring!