Implementing GliNER in Haystack v2 Pipeline

[chaalia]

I wanted to reach out to see if you could provide any guidance or recommendations on the following:

Existing Support: Is there any existing support or upcoming feature within Haystack v2 that facilitates the integration of GliNER or similar NER models?

Best Practices: Could you advise on the best practices for integrating custom NER models like GliNER into a Haystack pipeline? Specifically, any tips on performance optimization and maintaining compatibility with other Haystack components would be greatly appreciated.

Documentation or Examples: If there are any documentation, examples, or community discussions that you can point me to regarding custom model integration, that would be very helpful.

Thank you for your time and assistance. I look forward to your guidance on this matter.

[julian-risch]

Hi @chaalia out of the box Haystack only supports NER with the NamedEntityExtractor component with possible backends two NER backends: Hugging Face and spaCy. These two backends work with any HF or spaCy model that supports token classification or NER. https://docs.haystack.deepset.ai/docs/namedentityextractor
If you want to use GliNER, I suggest to create a custom component. We have instructions for how to do that here:
https://docs.haystack.deepset.ai/docs/custom-components

You can use the implementation of https://github.com/deepset-ai/haystack/blob/main/haystack/components/extractors/named_entity_extractor.py or https://github.com/deepset-ai/haystack/blob/main/haystack/components/generators/hugging_face_local.py as an example.
As you want to load a model as part of the custom component, you should add a warm_up method. This warm_up method then should include something like model = GLiNER.from_pretrained("urchade/gliner_mediumv2.1"). Haystack will run the warm_up method if the component is part of a pipeline automatically. If you are using the component without a pipeline, you need to run warm_up manually.

Here is some draft code to get you started with your custom component:

from haystack import component

@component
class GliNERNamedEntityExtractor:
  """
  A component extracting entities with GliNER
  """
    @component.output_types(documents=List[Document])
    def run(self, documents: List[Document]) -> Dict[str, Any]:
      labels = ["Person", "Award", "Date", "Competitions", "Teams"] # you'll want to add a parameter for that
      for doc in documents:
            entities = model.predict_entities(text, labels, threshold=0.5)
            doc.meta[self._METADATA_KEY] = entities # you'll need to make sure entities are in a format that can be stored in meta
       return {"documents": documents}

  def __init__(self, model_name: str = "urchade/gliner_mediumv2.1")
      self.model_name = model_name

  def warm_up(self):
        """
        Initializes the component.
        """
        if self.model is None:
            self.model = GLiNER.from_pretrained("urchade/gliner_mediumv2.1")

[chaalia]

Hello @julian-risch thanks a lot for this detailed answer.I really appreciate that.
I will test this custom component.