Building a Production-Grade Retrieval Augmented Generation (RAG) Pipeline

Nutan Sahoo Advitya Gemawat Description: Most standard RAG tutorials rely on a naive pattern: splitting text, embedding chunks, and injecting the top matches directly into an LLM prompt. While simple to implement, this approach frequently fails in real-world scenarios, retrieving semantically related but factually irrelevant context, generating unverified assertions, etc. In this hands-on workshop, we move beyond naive RAG to build a robust, production-ready RAG pipeline. Participants will learn how to design resilient architectures that handle complex, domain-specific document. Key Topics & Implementations Covered: Structural Document Preprocessing & Chunking: Parsing hierarchical documents into semantically coherent units rather than arbitrary fixed-character splits. Hybrid Search & Retrieval: Combining sparse keyword matching (BM25) with dense semantic embeddings to capture both exact terminology and conceptual intent. Adaptive Feedback Learning: Integrating dynamic feedback for improving retrieval. Sentence-Level In-Line Citations: Adding verifiable citation tags, so every fact links back to its source document. Cross-Session Persistence: Extracting and storing high-signal user profile facts to maintain continuity across distinct user sessions.