rag

 1. BERT lai cosine similarity ko basis ma embeddings generate garna train garne. [SentenceTransformers using Siamese BERT  Networks](https://arxiv.org/abs/1908.10084)
2. documents lai pahila chunks ma divide garerw tesko embeddings haru nikalxa, tespaxi tyo embeddings haru lai vector database ma store garxa. jun embedding model documents ko embeddings generate garna use gareko xah tesaile query ko pani embeddings generate garinxa ani tyo embedding rw vector database ma vako embeddings haru bich cosine similarity nikalxa. ani jun ko cosine similarity high xah (matlab angle between two vector small xah) tyo embeddings ko original text document dekhi retrieve garerw ani teslai user ko query sanga attach garerw prompt sanga llm lai dinxa, ani llm le contextual response provide garxa.
3. Tarw aba embeddings chai token level ma ani sentence level ma generate garna milxa, jastai BERT le words words ko token-level ma generate garxa vane Sentence-BERT le chai sentence ko direct embedding create garxa jun chai aru sentences haru sanga compare garna kaam lagxa RAG pipeline ma. BERT lai chai cosine similarity ko basis ma embeddings generate garna (masked transformer) train garinxa jasle garda sentences haru ko embeddings pani (semantic ani contextual meanings) majjale intact rahanxa jasle garda accurate retrieval garna maddat milxa. Tyo embeddings ko lagi train garne bert ma chai different tarika le model train garinxa (layers), kosaile [cls] token use garxa, kosaile maxpooling use garxa vane production ma chai mean pooling use hunxa josle k garxa vanda euta sentence lai single vector (something dimension ko) ma compress garxa.