Adding RAG to text-generation pipeline

[lalitpagaria]

🚀 Feature request

Thank you for the awesome work. I am working on https://github.com/deepset-ai/haystack/issues/443 and just wanted to check whether any plan to add RAG into text-generation pipeline.

Motivation

text-generation already have other models, hence it I would be great to have it in there. And this will help keeping our code clean by not adding classes for each type of generators.

model = pipeline('text-generation', model="facebook/rag-token-nq", tokenizer=None, device=-1)
# ValueError: Unrecognized configuration class <class 'transformers.configuration_rag.RagConfig'> for this kind of AutoModel: AutoModelForCausalLM.
Model type should be one of CamembertConfig, XLMRobertaConfig, RobertaConfig, BertConfig, OpenAIGPTConfig, GPT2Config, TransfoXLConfig, XLNetConfig, XLMConfig, CTRLConfig, ReformerConfig, BertGenerationConfig.

Your contribution

If you guide me I am happy to help.

[patrickvonplaten]

Hey @lalitpagaria - RAG is quite different to other generation models so we don't have it on the short-term roadmap to add it to pipelines. We are still thinking about how to integrate retrieval augmented models to the pipelines.

[lalitpagaria]

Thanks @patrickvonplaten Yeah I totally agree with you. Please let me know whether I close this issue or keep it open for future reference.

[patrickvonplaten]

Leave it open - I'll put it under projects so that I don't forget it :-)

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[g-jing]

@patrickvonplaten Great work! I noticed that transformers included the implementation for DPR. But for RAG, I only find a demo. Is there a source code for RAG? Or do you know where is Facebook's source code for RAG?

[lhoestq]

transformers does include RAG. You can even find the documentation here: https://huggingface.co/transformers/model_doc/rag.html