RAG on Não Inviabilize podcast transcriptions

RAG Pipeline

1. Data Webcrawling

Extract episodes transcriptions from Não Inviabilize podcast website.

2. QA synthetization

LLaMA 3 70B: generate questions and answers based on transcriptions.

3. Multi-index pre-processing

Chunknize episodes transcriptions:

• Window of sentences splitter
• ...

4. Retrieval of passages

1. BM25 vs Semantic Embeddings
2. MonoPTT5 Reranking

5. Generation

1. Naive approach
2. Agentic approach (ReAct, ...)

6. Pipeline evaluation

RAGAs:

• Faithfulness Score
• Answer Relevance
• Context Relevancy