[{"data":1,"prerenderedAt":190},["ShallowReactive",2],{"profile":3,"work-enterprise-rag-conversational-ai":14},{"name":4,"role":5,"tagline":6,"location":7,"contact":8},"Sankar Vema","AI Builder & Architect of Agentic Systems","I help enterprise leaders turn AI ambition into capability that actually ships.","India · open to global advisory engagements",{"email":9,"linkedin":10,"github":11,"blog":12,"twitter":13},"sankar.vema@gmail.com","https:\u002F\u002Fwww.linkedin.com\u002Fin\u002Fsankarvema\u002F","http:\u002F\u002Fsankarvema.github.io\u002F","http:\u002F\u002Fsankarvema.blogspot.com\u002F","https:\u002F\u002Ftwitter.com\u002Fsansvema",{"id":15,"title":16,"body":17,"context":160,"description":151,"domain":161,"draft":162,"extension":163,"hero":164,"impact":165,"meta":166,"navigation":167,"path":168,"role":169,"scale":170,"seo":171,"slug":172,"stack":173,"status":181,"stem":182,"tags":183,"year":188,"__hash__":189},"work\u002Fwork\u002Fenterprise-rag-conversational-ai.md","Enterprise RAG and Conversational AI",{"type":18,"value":19,"toc":150},"minimark",[20,25,29,33,36,40,69,73,129,133,136,139],[21,22,24],"h2",{"id":23},"the-mandate","The mandate",[26,27,28],"p",{},"I architected a set of retrieval-augmented and conversational AI systems across\nenterprise engagements, several of them in finance and healthcare. The common thread\nwas not the model. It was building systems that could be trusted in domains where a\nconfident wrong answer carries real consequences.",[21,30,32],{"id":31},"the-problem","The problem",[26,34,35],{},"A raw LLM in a regulated enterprise is a fluent liability. It answers from a frozen,\ngeneric memory, it cannot cite where an answer came from, and it will invent detail\nrather than admit a gap. In finance and healthcare that is not a rough edge, it is a\nreason not to ship. The work is turning a general model into a system that answers from\nthe enterprise's own current knowledge and can show its sources.",[21,37,39],{"id":38},"the-system","The system",[41,42,43,51,57,63],"ul",{},[44,45,46,50],"li",{},[47,48,49],"strong",{},"Retrieval-augmented generation."," Answers grounded in vector-indexed enterprise\nknowledge, retrieved at query time, so responses track the current corpus rather than\nthe model's training snapshot.",[44,52,53,56],{},[47,54,55],{},"Citations by construction."," Every answer carries the sources it was built from,\nwhich is what makes the output auditable rather than merely plausible.",[44,58,59,62],{},[47,60,61],{},"Voice and chat surfaces."," Both voicebot and chatbot front ends over the same\nretrieval core, so the channel is a choice and the grounding is constant.",[44,64,65,68],{},[47,66,67],{},"Guardrails for regulated use."," Refusal and escalation paths for out-of-scope or\nlow-confidence queries, because in these domains a careful no beats a fluent guess.",[21,70,72],{"id":71},"key-decisions-and-what-they-cost","Key decisions and what they cost",[74,75,76,92],"table",{},[77,78,79],"thead",{},[80,81,82,86,89],"tr",{},[83,84,85],"th",{},"Decision",[83,87,88],{},"Why",[83,90,91],{},"What it traded",[93,94,95,107,118],"tbody",{},[80,96,97,101,104],{},[98,99,100],"td",{},"RAG over fine-tuning as the default",[98,102,103],{},"Freshness and citable sources",[98,105,106],{},"A retrieval and indexing pipeline to own and keep current",[80,108,109,112,115],{},[98,110,111],{},"Citations required, not optional",[98,113,114],{},"Auditability in regulated domains",[98,116,117],{},"Extra plumbing on every answer path",[80,119,120,123,126],{},[98,121,122],{},"Explicit refusal and escalation",[98,124,125],{},"A careful no is safer than a confident guess",[98,127,128],{},"Some coverage traded for trust",[21,130,132],{"id":131},"what-it-proves","What it proves",[26,134,135],{},"Shipping conversational AI into finance and healthcare is a governance problem as much\nas a model problem. These systems are the pattern I keep reusing: ground the answer, cite\nthe source, and design the failure mode on purpose.",[137,138],"hr",{},[26,140,141,144,145],{},[47,142,143],{},"Related writing:"," ",[146,147,149],"a",{"href":148},"\u002Fwriting\u002Fdata-problem-in-costume","The data problem in a costume",{"title":151,"searchDepth":152,"depth":152,"links":153},"",3,[154,156,157,158,159],{"id":23,"depth":155,"text":24},2,{"id":31,"depth":155,"text":32},{"id":38,"depth":155,"text":39},{"id":71,"depth":155,"text":72},{"id":131,"depth":155,"text":132},"Multiple enterprise engagements","ai-products",false,"md","Retrieval-augmented assistants and voice and chat agents built for finance and healthcare, where an ungrounded answer is not a bug, it is a liability.","Grounded answers with citations, in finance and healthcare",{},true,"\u002Fwork\u002Fenterprise-rag-conversational-ai","Architect","Cross-industry, regulated domains",{"title":16,"description":151},"enterprise-rag-conversational-ai",[174,175,176,177,178,179,180],"RAG","Vector DB","LLM","Voicebot","Chatbot","Retrieval","Guardrails","shipped","work\u002Fenterprise-rag-conversational-ai",[184,185,186,187],"build","llm","rag","conversational","2023-2024","N3GXI3a4hiMHe9Y_z-eqnjvK2WALdX5_5F-nYsonfWg",1790601541920]