search

deepset-ai / haystack

:mag: AI orchestration framework to build customizable, production-ready LLM applications. Connect components (models, vector DBs, file converters) to pipelines or agents that can interact with your data. With advanced retrieval methods, it's best suited for building RAG, question answering, semantic search or conversational agent chatbots.
https://haystack.deepset.ai
Apache License 2.0
17.31k stars 1.89k forks source link

feat: Create `TransformersTextRouter` based on HF TextClassification pipeline #7753

Closed sjrl closed 4 months ago

sjrl commented 5 months ago

Is your feature request related to a problem? Please describe. Recently we half ported over this component from v1 with this PR. By this I mean previously the TransformerQueryClassifier worked both the the ZeroShotClassification pipeline from HF and the TextClassification pipeline from HF.

However, to follow the philosophy of Haystack v2 we decided to split up this old component into two new ones, the TransformersZeroShotTextRouter (based on the ZeroShotClassification pipeline) and the yet to be made TransformersTextRouter (based on the TextClassification pipeline).

Describe the solution you'd like A new Router called TransformersTextRouter based on the TextClassification pipeline.

Additional context This would allow a query routing based on text classification which the Sol team actively uses to build pipelines in Haystack V1 for clients (e.g. prompt injection classification). Therefore, to ease the transition to v2 for clients it would be great to add this component.

mrm1001 commented 5 months ago

Hi, there is some demand for a DocumentClassifier component (https://github.com/deepset-ai/haystack/issues/7669). So instead of having a TransformersTextRouter, another approach with a slightly longer pipeline would be having a DocumentClassifier connected to a ConditionalRouter. Did I understand correctly?

sjrl commented 5 months ago

Not quite, since the goal of the DocumentClassifier is to add classification labels to Documents in their metadata. And then index the Documents with that metadata into to the DocStore to then be used later by a query pipeline (e.g. through filters).

While technically it would be possible to use the DocumentClassifier with the ConditionalRouter as you outlined you would also have to convert the query into a Document (so some DocumentBuilder component), then pass that to the DocumentClassifier and then the ConditionalRouter, which is somewhat clunky.

I think it makes sense to have a component that can directly work on Text to perform this routing so we can skip the Document intermediate. I think this is similar reasoning as to why we have separate TextEmbedder and DocumentEmbedder components that use the same underlying models.

What do you think?

mrm1001 commented 5 months ago

Makes sense! Thanks @sjrl

julian-risch commented 4 months ago

Adding a new component TransformersTextRouter based on the TextClassification pipeline is a good idea! 👍 @sjrl We won't have capacity to work on it in the next sprint. Would you like to add the component in a PR? Sounds like you already have a clear understanding of what the implementation should look like and you could use the TransformersZeroShotTextRouter as a blue print.

sjrl commented 4 months ago

No worries, this isn't something we need urgently, but within the next quarter. I'd be happy to open PR once I have some free time to work on this :)