L39- AWS RAG Chatbot Architecture | Knowledge Base, S3, OpenSearch, Guardrails & Evaluation

Lecture 39 of the AI for Software Engineers series — Bipin Kumar designs and sets up Project 2: a production RAG chatbot deployed entirely on AWS. This class covers the complete chatbot architecture, Bedrock Knowledge Base setup from scratch, Guardrails configuration, and the built-in evaluation module. 🧠 What's Covered: Chatbot vs Document Classifier — Different Architecture: The doc classifier from the previous project was asynchronous — user uploads an image, S3 fires a trigger, Lambda processes it in the background, and the result is stored. The chatbot is synchronous — the user asks a question and waits for the answer immediately. This changes the architecture. Instead of an S3 trigger, we use API Gateway directly calling Lambda, which then calls the Knowledge Base and returns the answer in real time. AWS RAG Chatbot Architecture: User query → API Gateway (authentication, throttling, public URL) → Lambda (business logic) → Bedrock Knowledge Base (retrieval + generation) → OpenSearch (vector similarity search) → Lambda returns answer → API Gateway passes it back to the user. Knowledge Base — The Core of Project 2: Bedrock Knowledge Base is AWS's fully managed RAG pipeline. You configure it once via the console and it handles chunking, embedding, and vector storage automatically — no code needed for the pipeline itself. Knowledge Base key components: Data Source — an S3 bucket containing your PDF documents, Word files, or other supported formats. Everything in this bucket becomes searchable. Parser — two options. Default parser handles text-based documents (Word, Excel, CSV, Markdown, TXT, HTML). Foundation Model as Parser uses a multimodal LLM to read each page as an image — handles charts, tables, and complex PDF layouts that text extraction misses. Chunking Strategy — default, fixed size, hierarchical, or semantic. Choose based on your document type. Semantic chunking groups related content together by meaning. Embedding Model — converts each chunk into a vector. Options include Amazon Titan Text Embedding V2 (1024 dimensions, general purpose) and Cohere Embed V4 (multilingual — supports queries in Hindi even if documents are in English). Vector Database — Amazon OpenSearch Serverless is the recommended choice. Fully managed, auto-scales as document volume and query load grows. Aurora PostgreSQL is the alternative for teams already on relational databases. Sync — Keeping the Knowledge Base Updated: After uploading new documents to S3, you must trigger a Sync operation inside the Knowledge Base console. AWS then re-processes the new files — chunking, embedding, and storing in OpenSearch. Without syncing, new documents remain invisible to the chatbot. Guardrails — Controlling What the Chatbot Can Discuss: Guardrails are a virtual filter applied to every query and response. You configure which content categories to block: hate speech, insults, sexual content, violence, misconduct, and prompt injection attacks. Each category has an independently configurable threshold — allow, flag, or block. Guardrails also protect sensitive data. Configure it to detect and block personal information like phone numbers, Aadhaar numbers, bank account numbers, and email addresses from appearing in responses. Guardrails can be attached at the Knowledge Base level (applied to retrieval) or at the model level (applied to generation). For maximum protection, attach at both points. Evaluation — Measuring Chatbot Quality: Bedrock has a built-in evaluation module that scores your RAG chatbot across four metrics without needing an external library. Correctness — does the generated answer match the expected ground truth answer? Faithfulness — is every part of the answer grounded in the retrieved context? Extra information not from the documents reduces this score. Completeness — does the answer cover all parts of the question? Relevancy — does the answer actually address what was asked? Upload a test dataset with queries and expected answers, connect your Knowledge Base as the retrieval source, and run evaluation. AWS generates a score per metric and flags low-performing queries for review. ⏭️ Next Lecture (Lecture 40): 👉 Hands-on: Lambda function for chatbot + API Gateway integration + live testing 💬 Questions about Knowledge Base setup or Guardrails configuration? Drop them in the comments — Bipin replies! 📌 Subscribe so you never miss a class. #AWSBedrock #KnowledgeBase #RAGChatbot #Guardrails #OpenSearch #AWSEvaluation #Lambda #APIGateway #AIforEngineers #BipinKumar #GenAI #RAGPipeline #S3 #EmbeddingModel #AIInterview