RAG_Presentation_on_transformers_in.pptx

about rag

1. Retrieval-Augmented Generation (RAG) Understanding how large language models use external knowledge to enhance responses.

2. 1. What is RAG? RAG stands for Retrieval-Augmented Generation. It combines information retrieval with text generation. Developed to allow models to access external knowledge sources dynamically.

3. 2. Why do we need RAG? Language models have limited training data and may hallucinate facts. RAG enables them to fetch up-to-date, factual information. Improves reliability, factual accuracy, and domain adaptability.

4. 3. Basic Components of RAG Retriever – fetches relevant documents from a knowledge base. Generator – uses these documents to generate final answers. Knowledge Base – collection of documents or embeddings.

5. 4. How RAG Works – Step by Step 1. User asks a question. 2. Retriever searches relevant documents. 3. Retrieved info is combined with the query. 4. Generator produces the final output using both.

6. 5. Example: RAG in QA Systems Question: 'Who discovered penicillin?' Retriever fetches: Document about Alexander Fleming (1928). Generator forms: 'Penicillin was discovered by Alexander Fleming in 1928.'

7. 6. RAG Architecture Diagram RAG = Retriever + Generator + Knowledge Base Retriever finds context → Generator uses it to generate output. (Visual representation: arrows showing query → retriever → docs → generator → answer)

8. 7. Types of Retrievers Sparse retrievers: TF-IDF, BM25. Dense retrievers: DPR (Dense Passage Retrieval), Sentence-BERT. Hybrid retrievers: Combine both for better recall.

9. 8. Knowledge Sources Structured databases (SQL, vector DBs like FAISS, Pinecone). Unstructured text (Wikipedia, PDFs, docs). Company-specific or domain-specific datasets.

10. 9. Advantages of RAG Factual and up-to-date responses. Requires less retraining compared to fine-tuning. Easier to customize using private data sources.

11. 10. Limitations of RAG Retriever quality affects overall performance. Latency – two-step process can be slower. Hard to handle ambiguous or multi-hop queries.

12. 11. RAG vs Fine-tuning Fine-tuning: Model learns fixed knowledge. RAG: Retrieves dynamic knowledge at inference time. RAG is more adaptable and efficient for fast-changing domains.

13. 12. Applications of RAG Chatbots and assistants (ChatGPT with retrieval). Search engines and recommendation systems. Medical, legal, and academic question answering.

14. 13. Research Directions Improving retriever efficiency. Better integration of retrieved context. Reducing hallucinations and improving interpretability.

15. 14. Future Scope Integration with real-time web data. Use in multi-modal models (text + images). Personalized knowledge retrieval for adaptive AI.

16. 15. References Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”, NeurIPS 2020. Izacard & Grave, “Distilling Knowledge from Reader to Retriever”, ICLR 2021. Karpukhin et al., “Dense Passage Retrieval for Open-Domain Question Answering”, EMNLP 2020.

17. 16. Summary RAG enhances LLMs with retrieval capabilities. Bridges gap between static knowledge and dynamic facts. Future of intelligent, fact-grounded AI systems.