Example of Retrieval Augmented Generation with a private dataset.
We have demonstrated three different ways to utilise RAG Implementations over the document for Question/Answering and Parsing.
1 - Original MetaAI RAG Paper Implementation for user dataset.
2 - Llama-Index, LangChain and OpenAI RAG Implementation for user dataset.
3 - Minimilistic Implementation of LaBSE + OpenAI RAG Implementation for user dataset.
Note:
1 - Implementation 2 and Implementation 3 has metrics result for the test dataset on WER (Word Error Rate used for Transcription) and RoUGE Scores (used for Translation)
2 - These also have implementation of Multilingual Input/Output because of LID from NLLB (MetaAI) with ChatGPT and LID + LaBSE with ChatGPT.
All the notebook implementations are mostly self-explanatory.
It consists of usually 5 sections -
- Package Installation and Setup
- Data Cleaning and Preprocessing
- Main CodeBlock with Indexing and Usage.
- Inference Example and WrapUp for Productionizing.
- Evaluation Metrics.
I mostly used the evaluation metrics which made sense to me and usually towards established and statistical metrics.
But for further work,
BertScore- https://huggingface.co/spaces/evaluate-metric/bertscoreColBERT- https://github.com/stanford-futuredata/ColBERTRAGAS- https://github.com/explodinggradients/ragas
Gradio Application is an Inference Code for Implementation 2 - LlamaIndex/LangChain/OpenAI.
It uses gpt-3.5-turbo model with temperature 0 (zero) for answer generation.
You need put OpenAI Key in line 22 for Gradio Application and similarly in the notebook instance.
os.environ['OPENAI_API_KEY'] = <openai-api-key>
Then spin up the gradio application with given configuration, change question examples if using for different dataset.
python gradio_app.py
Audio Implementation can also be added to these implementations via OpenAI/Whisper and Microsoft/speecht5_tts but has not been implemented.
Can be added later based on request.