主题演讲 - 企业级代理 RAG：从数据统一到值得信赖的洞察.pdf

1、Enterprise-grade Agentic RAG:From Data Unification to Trustworthy Insights1A blueprint for scalable,auditable AI knowledge systemsViktor BotevCTO&Co-Founder2The Pilot Trap95%AI Pilots FailThe issue:messy data,weak evaluation,no governance.3The Real Enterprise ChallengeDisconnected silos across depar

2、tments.No unified context for reasoning.High cost of hallucinations&non-compliance.AI without knowledge integrity cannot be trusted.4From Retrieval to Agentic RAGRetrieval RAG Agentic RAG Enterprise Agentic RAG.SEARCH5Enterprise Agentic RAG ArchitectureThere is no progress without proper Evaluation.

3、Agentic RAG Develop domain-specific LLM agents to identify the best data sources align to every questions&answer with the right reasoningBuilt-in Governance Monitor with fine-grained access policies,real-time quality&efficiency metrics tracking for continuous improvementLLM EvaluationOptimize the qu

4、ality of entire agentic workflow and ability to evaluate LLMs continuously.Unified Data FabricConnect to any structured or unstructured data source and delivers agent-ready data without rebuilding pipelines,even PDF!(Heavy Indexing)6Enterprise Agentic RAG ArchitectureCONNECT Your Data(Extract,OCR,In

5、dex,Ontology)BUILD-RAG Core(Retriever,Assembler,Generator)ORCHESTRATE-Specialized AgentsEVALUATE(ConSens,WISDM)DEPLOY(Monitoring,Cost Control)There is no progress without proper Evaluation.7Data Unification Multimodal extraction(text,tables,images,schematics).Linking entities and context into reusab

6、le knowledge assets.Asset an Entity connected to specs,drawings,and databases.Data is Enterprises Gold.It needs to be treated as such.8Orchestrating Specialized AgentsAbstraction and Generalization gives control.Multi-AgentArchitecture9Evaluation&Governance The Trust LayerMetrics