{"type":"doc","content":[{"type":"blockquote","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"A complete technical guide — from building a sequential multi-agent marketing workflow with AutoGen, through UserProxy interaction patterns, RAG chunking strategies, latency optimization, evaluation matrices, AWS Bedrock guardrails, and enterprise ERP integration."}]}]},{"type":"horizontalRule"},{"type":"heading","attrs":{"textAlign":null,"level":2},"content":[{"type":"text","text":"Table of Contents"}]},{"type":"orderedList","attrs":{"start":1,"type":null},"content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"link","attrs":{"href":"https://claude.ai/chat/06a2a95a-da8c-4c8c-b8f5-0889c1daa74d#1-sequential-agent-pipeline-core-architecture","target":"_blank","rel":"noopener noreferrer nofollow","class":"text-blue-600 underline"}}],"text":"Sequential Agent Pipeline: Core Architecture"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"link","attrs":{"href":"https://claude.ai/chat/06a2a95a-da8c-4c8c-b8f5-0889c1daa74d#2-building-the-autogen-marketing-workflow","target":"_blank","rel":"noopener noreferrer nofollow","class":"text-blue-600 underline"}}],"text":"Building the AutoGen Marketing Workflow"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"link","attrs":{"href":"https://claude.ai/chat/06a2a95a-da8c-4c8c-b8f5-0889c1daa74d#3-userproxy-interaction-flow--controls","target":"_blank","rel":"noopener noreferrer nofollow","class":"text-blue-600 underline"}}],"text":"UserProxy: Interaction Flow & Controls"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"link","attrs":{"href":"https://claude.ai/chat/06a2a95a-da8c-4c8c-b8f5-0889c1daa74d#4-initiate_chat-parameters-reference","target":"_blank","rel":"noopener noreferrer nofollow","class":"text-blue-600 underline"}}],"text":"initiate_chat Parameters Reference"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"link","attrs":{"href":"https://claude.ai/chat/06a2a95a-da8c-4c8c-b8f5-0889c1daa74d#5-content-iteration-cycle","target":"_blank","rel":"noopener noreferrer nofollow","class":"text-blue-600 underline"}}],"text":"Content Iteration Cycle"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"link","attrs":{"href":"https://claude.ai/chat/06a2a95a-da8c-4c8c-b8f5-0889c1daa74d#6-frontendbackend-integration","target":"_blank","rel":"noopener noreferrer nofollow","class":"text-blue-600 underline"}}],"text":"Frontend/Backend Integration"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"link","attrs":{"href":"https://claude.ai/chat/06a2a95a-da8c-4c8c-b8f5-0889c1daa74d#7-rag-chunking-strategies","target":"_blank","rel":"noopener noreferrer nofollow","class":"text-blue-600 underline"}}],"text":"RAG Chunking Strategies"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"link","attrs":{"href":"https://claude.ai/chat/06a2a95a-da8c-4c8c-b8f5-0889c1daa74d#8-rag-performance--latency-optimization","target":"_blank","rel":"noopener noreferrer nofollow","class":"text-blue-600 underline"}}],"text":"RAG Performance & Latency Optimization"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"link","attrs":{"href":"https://claude.ai/chat/06a2a95a-da8c-4c8c-b8f5-0889c1daa74d#9-evaluation-matrix-rag-vs-fine-tuned-models","target":"_blank","rel":"noopener noreferrer nofollow","class":"text-blue-600 underline"}}],"text":"Evaluation Matrix: RAG vs Fine-Tuned Models"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"link","attrs":{"href":"https://claude.ai/chat/06a2a95a-da8c-4c8c-b8f5-0889c1daa74d#10-aws-bedrock-guardrails","target":"_blank","rel":"noopener noreferrer nofollow","class":"text-blue-600 underline"}}],"text":"AWS Bedrock Guardrails"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"link","attrs":{"href":"https://claude.ai/chat/06a2a95a-da8c-4c8c-b8f5-0889c1daa74d#11-enterprise-erp-integration-with-gen-ai-pipelines","target":"_blank","rel":"noopener noreferrer nofollow","class":"text-blue-600 underline"}}],"text":"Enterprise ERP Integration with Gen-AI Pipelines"}]}]}]},{"type":"horizontalRule"},{"type":"heading","attrs":{"textAlign":null,"level":2},"content":[{"type":"text","text":"1. Sequential Agent Pipeline: Core Architecture"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"The central architectural pattern in this workflow is a "},{"type":"text","marks":[{"type":"bold"}],"text":"sequential agent chain"},{"type":"text","text":" mediated by a UserProxy. No agent talks directly to another — every message is routed through the proxy, enabling human oversight at each handoff."}]},{"type":"customImage","attrs":{"src":"https://67dc5f63b8b6a6f0ea39b99a.storage.fermion.app/public-cached/markdown-editor-file/file-unique-id/fkedrlqz7osq-mzpwxywjlztkrls0ezzhgvxg-n6ypqkkz077x7aocucuoyg_bbvexxc-e02u4owuuckyf5bqukeemk7v_nzrtap","alt":null,"width":600,"height":327}},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Figure: AutoGen's sequential pipeline — UserProxy routes messages between specialist agents and enforces human approval at each step."}]},{"type":"codeBlock","attrs":{"language":"javascript"},"content":[{"type":"text","text":"UserProxy → StrategyCreator → UserProxy → ContentCreator → UserProxy → Final Output\n ↑ | ↑ |\n feedback structured feedback channel-specific\n strategy copy\n"}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"1.1 Core Components"}]},{"type":"table","content":[{"type":"tableRow","content":[{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Component"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Description"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Agent"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"An autonomous LLM instance with a defined persona (system message)"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"LLM Configuration"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Dictates model (e.g., "},{"type":"text","marks":[{"type":"code"}],"text":"gpt-4o-mini"},{"type":"text","text":"), temperature, and other inference params"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Sequential Chat"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Agents exchange messages in a defined order (Agent1 → Agent2 → …)"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"UserProxy"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"The human or user agent that initiates conversation and routes every handoff"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Human-Input Mode"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Controls whether conversation stops for user confirmation ("},{"type":"text","marks":[{"type":"code"}],"text":"always"},{"type":"text","text":") or runs autonomously ("},{"type":"text","marks":[{"type":"code"}],"text":"never"},{"type":"text","text":")"}]}]}]}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"1.2 Cost & Efficiency Comparison"}]},{"type":"table","content":[{"type":"tableRow","content":[{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Aspect"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Human Team"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"AI Agent System"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Up-front cost"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"₹40–50 Lakh (salaries, consulting fees)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"< $200 (API usage)"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Time to deliver"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Weeks to months (analysis, approvals)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Hours (automated reasoning)"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Scalability"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Limited by headcount"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Unlimited parallel agents"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Decision latency"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Human approval required at each stage"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Configurable "},{"type":"text","marks":[{"type":"code"}],"text":"always"},{"type":"text","text":" / "},{"type":"text","marks":[{"type":"code"}],"text":"never"},{"type":"text","text":" mode"}]}]}]}]},{"type":"horizontalRule"},{"type":"heading","attrs":{"textAlign":null,"level":2},"content":[{"type":"text","text":"2. Building the AutoGen Marketing Workflow"}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"2.1 Problem Statement"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Goal"},{"type":"text","text":": Design a marketing strategy for a lightweight water-bottle product for school-age children (ages 5–16) and determine the most effective channels within a $2,000 budget."}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"The traditional approach costs ₹40–50 Lakh in salaries for marketing strategists, content writers, and analysts. The agentic approach replaces this with two specialized agents + a proxy."}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"2.2 Step-by-Step Construction"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Step 1: Install and import"}]},{"type":"codeBlock","attrs":{"language":"bash"},"content":[{"type":"text","text":"pip install pyautogen\npip install --upgrade openai # ensure compatible version\n"}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"import autogen\nfrom autogen import AssistantAgent, UserProxyAgent\nfrom IPython.display import display, Markdown\nimport os\nfrom getpass import getpass\n\n# Securely set API key\nos.environ[\"OPENAI_API_KEY\"] = getpass(\"Enter OpenAI API key: \")\n"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Step 2: Configure the LLM"}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"llm_config = {\n \"config_list\": [\n {\n \"model\": \"gpt-4o-mini\", # cost-efficient, capable model\n \"temperature\": 0, # 0 = deterministic, reproducible outputs\n }\n ]\n}\n"}]},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Setting "},{"type":"text","marks":[{"type":"code"}],"text":"temperature=0"},{"type":"text","text":" ensures consistent, reproducible strategy outputs. For creative copy generation, increase to "},{"type":"text","marks":[{"type":"code"}],"text":"0.7–0.9"},{"type":"text","text":"."}]}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Step 3: Create Agent 1 — Strategy Creator"}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"strategy_agent = AssistantAgent(\n name=\"StrategyCreator\",\n system_message=(\n \"You are a logical, data-driven marketing manager. \"\n \"When given a product, target audience, and budget, you must: \"\n \"1. Identify the most effective marketing channels. \"\n \"2. Allocate the budget across those channels with clear rationale. \"\n \"3. Define the primary and secondary target audiences. \"\n \"4. Summarize key product features to emphasize in campaigns. \"\n \"Return a structured marketing plan.\"\n ),\n llm_config=llm_config,\n)\n"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Step 4: Create Agent 2 — Content Creator"}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"content_agent = AssistantAgent(\n name=\"ContentCreator\",\n system_message=(\n \"You are a creative AI marketing copywriter. \"\n \"Given a marketing strategy, produce compelling, channel-specific copy. \"\n \"For each channel (Instagram, email, school events), write a distinct piece of content. \"\n \"Use engaging tone, emojis where appropriate, and strong calls-to-action.\"\n ),\n llm_config=llm_config,\n)\n"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Step 5: Instantiate the UserProxy"}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"user_proxy = UserProxyAgent(\n name=\"User\",\n human_input_mode=\"always\", # pauses for human review after each agent response\n code_execution_config=False, # disable code execution for safety\n max_consecutive_auto_reply=3 # auto-reply limit before forcing human input\n)\n"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Step 6: Define the problem statement"}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"problem_statement = \"\"\"\nProduct: Reusable water bottle for school children aged 5–16.\nKey features: Durable, spill-proof, BPA-free, colorful, lightweight, optional personalization.\nTarget audience: Health-conscious parents in urban middle-to-upper-middle income brackets;\n secondary: school teachers and administrators.\nBudget: $2,000 total.\nTask: Create a channel-specific marketing strategy and generate copy for each channel.\n\"\"\"\n"}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"2.3 Strategy Agent Output"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"When the StrategyCreator receives the problem statement, it produces a structured plan. Here's a representative output:"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Budget Allocation ($2,000 total)"},{"type":"text","text":":"}]},{"type":"table","content":[{"type":"tableRow","content":[{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Channel"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Allocation"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Rationale"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Social-media ads (FB/IG)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"$800"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Target parents with video & carousel ads; boost posts in high-engagement regions"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Influencer marketing"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"$500"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Partner with local parenting influencers on Instagram/TikTok for reviews"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Content marketing (blogs)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"$400"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Publish hydration and sustainability posts; cross-post on school newsletters"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"School promotions (events)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"$200"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Host \"hydration challenge\" events; provide free bottles to teachers"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Email marketing"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"$100"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Campaign to school-collected parent emails"}]}]}]}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Revised budget"},{"type":"text","text":" (after proxy feedback):"}]},{"type":"table","content":[{"type":"tableRow","content":[{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Channel"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Original ($)"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Revised ($)"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Change"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Social-media ads"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"500"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"300"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"–200"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Influencer marketing"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"400"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"200"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"–200"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Content marketing"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"400"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"400"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"no change"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"School promotion events"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"200"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"800"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"+600"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Online contests"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"100"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"900"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"+800"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Email marketing"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"100"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"100"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"no change"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Total"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"2,000"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"2,000"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"balanced"}]}]}]}]},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"The proxy requested a shift toward direct-marketing channels (events, contests). Agent 1 regenerated the allocation — demonstrating the human-in-the-loop correction cycle."}]}]},{"type":"horizontalRule"},{"type":"heading","attrs":{"textAlign":null,"level":2},"content":[{"type":"text","text":"3. UserProxy: Interaction Flow & Controls"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"The UserProxy is the mediator at the center of every agent interaction. There is "},{"type":"text","marks":[{"type":"bold"}],"text":"no direct agent-to-agent communication"},{"type":"text","text":" — all messages pass through the proxy."}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"3.1 Sequential Chat Mechanics"}]},{"type":"codeBlock","attrs":{"language":"javascript"},"content":[{"type":"text","text":"Step 1: UserProxy → Agent1 (sends problem statement)\nStep 2: Agent1 → UserProxy (returns structured strategy)\n ↓\n UserProxy reviews:\n - \"please modify\" → Agent1 regenerates\n - carry_over → forward to Agent2\nStep 3: UserProxy → Agent2 (forwards strategy)\nStep 4: Agent2 → UserProxy (returns channel-specific copy)\n ↓\n Same approve/reject loop\nStep 5: proxy satisfied → \"summarize\" command → final output returned\n"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Key principle: The proxy stores context only if "},{"type":"text","marks":[{"type":"code"}],"text":"store_history=True"},{"type":"text","text":". Otherwise only the "},{"type":"text","marks":[{"type":"bold"}],"text":"final message"},{"type":"text","text":" is retained for the next agent."}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"3.2 Proxy Input Actions"}]},{"type":"table","content":[{"type":"tableRow","content":[{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"User Input"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"System Response"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Effect on Workflow"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"enter"},{"type":"text","text":" (blank)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Auto-reply triggers, forwarding last message to next agent"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Continues without additional feedback"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"exit"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Terminates the current run"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Stops all further agent activity"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Typed feedback (e.g., "},{"type":"text","marks":[{"type":"code"}],"text":"\"please make it more creative\""},{"type":"text","text":")"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Passed back to active agent → agent regenerates"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Enables iterative refinement"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"skip"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Proxy moves to next agent using "},{"type":"text","marks":[{"type":"code"}],"text":"carry_over=True"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Bypasses manual approval for that step"}]}]}]}]},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"⚠️ "},{"type":"text","marks":[{"type":"bold"}],"text":"Caution"},{"type":"text","text":": Pressing "},{"type":"text","marks":[{"type":"code"}],"text":"enter"},{"type":"text","text":" without additional text can inadvertently cause the system to reuse the previous message, potentially negating recent edits. Always use explicit feedback or "},{"type":"text","marks":[{"type":"code"}],"text":"exit"},{"type":"text","text":" to finalize."}]}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"3.3 human_input_mode Options"}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"# Option 1: Fully autonomous (no human pauses)\nuser_proxy = UserProxyAgent(\n name=\"User\",\n human_input_mode=\"never\", # runs end-to-end without stopping\n)\n\n# Option 2: Human approval at every step (default)\nuser_proxy = UserProxyAgent(\n name=\"User\",\n human_input_mode=\"always\", # pauses after each agent response\n)\n\n# Option 3: Terminate after first reply\nuser_proxy = UserProxyAgent(\n name=\"User\",\n human_input_mode=\"terminate\", # stops after the first response\n)\n"}]},{"type":"horizontalRule"},{"type":"heading","attrs":{"textAlign":null,"level":2},"content":[{"type":"text","text":"4. initiate_chat Parameters Reference"}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"4.1 Core Parameters"}]},{"type":"table","content":[{"type":"tableRow","content":[{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Parameter"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Type / Example"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Meaning"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"chat_queue"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"list of dicts"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Ordered list of sender → recipient specifications"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"sender"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"\"User\""}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Entity that starts the message (the UserProxy)"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"recipient"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"\"StrategyCreator\""}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Target agent for the current message"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"clear_history"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"True"},{"type":"text","text":" / "},{"type":"text","marks":[{"type":"code"}],"text":"False"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Whether to erase prior exchanges before sending"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"silent"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"False"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"True"},{"type":"text","text":" hides LLM responses; "},{"type":"text","marks":[{"type":"code"}],"text":"False"},{"type":"text","text":" shows them"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"carry_over"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"True"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Auto-passes current message to next agent after approval"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"finish_chat_index"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"-1"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Index of last message to retain (–1 = last message)"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"summary_method"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"\"last_message\""}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"How the final result is compiled"}]}]}]}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"4.2 Typical Initiation Call"}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"chat_result = autogen.initiate_chats(\n chat_queue=[\n {\n \"sender\": user_proxy,\n \"recipient\": strategy_agent,\n \"message\": problem_statement,\n \"clear_history\": True,\n \"silent\": False,\n \"carry_over\": True,\n \"summary_method\": \"last_message\"\n },\n {\n \"sender\": user_proxy,\n \"recipient\": content_agent,\n \"message\": \"Based on the strategy above, generate channel-specific marketing copy.\",\n \"clear_history\": False, # retain strategy context\n \"silent\": False,\n \"carry_over\": True,\n \"summary_method\": \"last_message\"\n }\n ]\n)\n\n# Access final output\nprint(chat_result[-1].summary)\n"}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"4.3 Parameter Tuning Tips"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"clear_history=False"},{"type":"text","text":" when you need context from earlier turns (e.g., multi-step campaigns where Agent 2 needs Agent 1's output)"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"silent=False"},{"type":"text","text":" during development to monitor each LLM response; switch to "},{"type":"text","marks":[{"type":"code"}],"text":"True"},{"type":"text","text":" in production to reduce console noise"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"carry_over=True"},{"type":"text","text":" streamlines handoffs but the proxy must still approve each intermediate result"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"finish_chat_index=-1"},{"type":"text","text":" to capture only the final message (the most common production setting)"}]}]}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"4.4 Custom Termination Condition"}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"# Lambda-based termination: stops when agent says \"terminate\"\nterminate_condition = lambda msg: msg.get(\"content\", \"\").strip().lower().endswith(\"terminate\")\n\nuser_proxy = UserProxyAgent(\n name=\"User\",\n human_input_mode=\"never\",\n is_termination_msg=terminate_condition,\n max_consecutive_auto_reply=10 # safety cap\n)\n"}]},{"type":"horizontalRule"},{"type":"heading","attrs":{"textAlign":null,"level":2},"content":[{"type":"text","text":"5. Content Iteration Cycle"}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"5.1 Cycle Overview"}]},{"type":"codeBlock","attrs":{"language":"javascript"},"content":[{"type":"text","text":"Strategy Agent → supplies revised budget and high-level plan\n ↓\nContent Agent → drafts channel-specific copy\n ↓\nUserProxy reviews:\n satisfied? → carry_over to next stage\n not satisfied? → typed feedback → agent regenerates\n ↓\nCycle repeats until proxy issues 'exit'\n"}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"5.2 Example: Generated Social Media Copy"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Draft 1"},{"type":"text","text":" (initial output):"}]},{"type":"codeBlock","attrs":{"language":"javascript"},"content":[{"type":"text","text":"\"Attention all parents! Ready to transform hydration time into fun time? \nPicture your child sipping from a vibrant, spill-proof bottle designed just \nfor them. No spills. No fuss. Pure joy with every sip. 🌟 \n#HydrationHeroes #KidFriendly #EcoSafe\"\n"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Revised Draft"},{"type":"text","text":" (after proxy feedback: \"make it more creative, add plastic-free highlight\"):"}]},{"type":"codeBlock","attrs":{"language":"javascript"},"content":[{"type":"text","text":"\"🚀 Hydration heroes, unite! 🌈 Give your child the coolest, cleanest \nhydration companion – 100% plastic-free, smell-free, and built to survive \nany playground adventure. Join the #HydrationChallenge and watch them \nthrive! #EcoKids #StayHydrated\"\n"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"The proxy can continue requesting tweaks (e.g., "},{"type":"text","marks":[{"type":"code"}],"text":"\"add a giveaway alert\""},{"type":"text","text":", "},{"type":"text","marks":[{"type":"code"}],"text":"\"highlight BPA-free feature\""},{"type":"text","text":") — the content agent updates accordingly until the proxy is satisfied."}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"5.3 Full End-to-End Code"}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"import autogen\nimport os\nfrom getpass import getpass\n\nos.environ[\"OPENAI_API_KEY\"] = getpass(\"OpenAI API Key: \")\n\nllm_config = {\"config_list\": [{\"model\": \"gpt-4o-mini\", \"temperature\": 0}]}\n\n# Define agents\nstrategy_agent = autogen.AssistantAgent(\n name=\"StrategyCreator\",\n system_message=\"You are a data-driven marketing manager. Create a structured channel strategy with budget allocation.\",\n llm_config=llm_config,\n)\n\ncontent_agent = autogen.AssistantAgent(\n name=\"ContentCreator\",\n system_message=\"You are a creative copywriter. Given a strategy, produce engaging channel-specific marketing copy.\",\n llm_config=llm_config,\n)\n\nuser_proxy = autogen.UserProxyAgent(\n name=\"User\",\n human_input_mode=\"always\",\n code_execution_config=False,\n is_termination_msg=lambda msg: \"DONE\" in msg.get(\"content\", \"\").upper()\n)\n\n# Run the sequential pipeline\nproblem = \"\"\"\nProduct: Reusable water bottle for school children (ages 5-16).\nFeatures: Durable, spill-proof, BPA-free, colorful, lightweight.\nAudience: Health-conscious parents; secondary: teachers/administrators.\nBudget: $2000.\nDeliver: 1) Marketing strategy 2) Channel-specific copy.\n\"\"\"\n\n# Initiate pipeline: proxy → strategy agent → content agent\nuser_proxy.initiate_chats([\n {\"recipient\": strategy_agent, \"message\": problem, \"carry_over\": True},\n {\"recipient\": content_agent, \"message\": \"Based on the strategy above, write copy for each channel.\", \"carry_over\": True}\n])\n"}]},{"type":"horizontalRule"},{"type":"heading","attrs":{"textAlign":null,"level":2},"content":[{"type":"text","text":"6. Frontend/Backend Integration"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"To expose the AutoGen workflow as a web application:"}]},{"type":"customImage","attrs":{"src":"https://67dc5f63b8b6a6f0ea39b99a.storage.fermion.app/public-cached/markdown-editor-file/file-unique-id/yifd_8a8bnhtni4solxegqkivlt2e7vqtaselibaoczlvgaapeqvuteo3z0ay0el-2wd7ktzkcakztzz8bzialbbgdszbt0moxsl","alt":null,"width":600,"height":327}},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Figure: FastAPI hosts the agent pipeline; React renders the UI and maps user actions to proxy commands."}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"6.1 Backend (FastAPI)"}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"from fastapi import FastAPI\nfrom pydantic import BaseModel\nimport autogen\nimport os\n\napp = FastAPI()\n\nllm_config = {\"config_list\": [{\"model\": \"gpt-4o-mini\", \"temperature\": 0.7}]}\n\nclass ChatRequest(BaseModel):\n message: str\n action: str = \"feedback\" # \"feedback\", \"skip\", \"exit\"\n\n@app.post(\"/chat\")\nasync def chat(request: ChatRequest):\n if request.action == \"exit\":\n return {\"status\": \"terminated\", \"content\": \"Session ended.\"}\n\n # Run agent pipeline with user's message as proxy input\n strategy_agent = autogen.AssistantAgent(\n name=\"StrategyCreator\",\n system_message=\"You are a marketing manager. Create a strategy.\",\n llm_config=llm_config\n )\n\n content_agent = autogen.AssistantAgent(\n name=\"ContentCreator\",\n system_message=\"You are a copywriter. Generate channel-specific copy.\",\n llm_config=llm_config\n )\n\n user_proxy = autogen.UserProxyAgent(\n name=\"User\",\n human_input_mode=\"never\", # automated in API context\n code_execution_config=False\n )\n\n # Inject user's message into the pipeline\n results = user_proxy.initiate_chats([\n {\"recipient\": strategy_agent, \"message\": request.message},\n {\"recipient\": content_agent, \"message\": \"Generate copy based on the strategy above.\"}\n ])\n\n return {\"status\": \"success\", \"content\": results[-1].summary}\n"}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"6.2 Frontend (React — Conceptual)"}]},{"type":"codeBlock","attrs":{"language":"jsx"},"content":[{"type":"text","text":"// Conceptual React component — maps UI actions to proxy commands\nfunction MarketingAgentUI() {\n const [draft, setDraft] = useState(\"\");\n const [feedback, setFeedback] = useState(\"\");\n\n const sendFeedback = async (action) => {\n const response = await fetch(\"/chat\", {\n method: \"POST\",\n headers: { \"Content-Type\": \"application/json\" },\n body: JSON.stringify({ message: feedback, action })\n });\n const data = await response.json();\n setDraft(data.content);\n };\n\n return (\n

\n      <input\n        placeholder=\"Your feedback (or leave blank to approve)\"\n        value={feedback}\n        onChange={(e) => setFeedback(e.target.value)}\n      />\n      <button onClick={() => sendFeedback(\"feedback\")}>Submit Feedback</button>\n      <button onClick={() => sendFeedback(\"skip\")}>Approve & Continue</button>\n      <button onClick={() => sendFeedback(\"exit\")}>Exit</button>\n    </div>\n  );\n}\n"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"The three buttons map directly to UserProxy actions:"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Submit Feedback"},{"type":"text","text":" → typed feedback sent to active agent → regenerates"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Approve & Continue"},{"type":"text","text":" → "},{"type":"text","marks":[{"type":"code"}],"text":"carry_over=True"},{"type":"text","text":" → advances to next agent"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Exit"},{"type":"text","text":" → terminates the session"}]}]}]},{"type":"horizontalRule"},{"type":"heading","attrs":{"textAlign":null,"level":2},"content":[{"type":"text","text":"7. RAG Chunking Strategies"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Chunking is the process of splitting raw documents into smaller pieces before indexing. Chunk size and overlap directly affect both retrieval quality and latency."}]},{"type":"customImage","attrs":{"src":"https://67dc5f63b8b6a6f0ea39b99a.storage.fermion.app/public-cached/markdown-editor-file/file-unique-id/ocqe0nzem6rcigxjqeyymnvuzvw3ruumk8w--gp3diva3mi8yrynnhfxdwoss7ihahy0wwhahkoaiok6d2sbmxitmsrcu8eoklld","alt":null,"width":600,"height":327}},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Figure: Different chunking methods suit different document types — recursive for structured docs, semantic for topic coherence."}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"7.1 Chunk Size Trade-offs"}]},{"type":"table","content":[{"type":"tableRow","content":[{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Chunk Size"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Speed"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Quality"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Risk"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Large (e.g., 5,000 tokens)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"❌ Slow (LLM processes more per chunk)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"✅ More context per chunk"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Context window overflow"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Small (e.g., 2–3 sentences)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"✅ Fast"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"❌ Limited context; may miss key info"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Answer quality degradation"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Moderate (200–500 tokens)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"✅ Balanced"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"✅ Balanced"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Recommended default"}]}]}]}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"7.2 Overlap"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Adding overlapping tokens between consecutive chunks preserves context across boundaries — particularly important for questions that span two adjacent chunks."}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"# Example: 300-token chunks with 50-token overlap\nfrom langchain.text_splitter import RecursiveCharacterTextSplitter\n\nsplitter = RecursiveCharacterTextSplitter(\n    chunk_size=300,\n    chunk_overlap=50,       # 50 tokens shared between adjacent chunks\n    length_function=len,\n    separators=[\"\\n\\n\", \"\\n\", \". \", \" \", \"\"]  # recursive splitting hierarchy\n)\n\nchunks = splitter.split_text(raw_document)\nprint(f\"Total chunks: {len(chunks)}\")\nprint(f\"First chunk (first 100 chars): {chunks[0][:100]}\")\n"}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"7.3 Chunking Methods"}]},{"type":"table","content":[{"type":"tableRow","content":[{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Method"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Best For"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Notes"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Recursive split"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Hierarchical documents (chapters → sections)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Maintains document structure"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Fixed-size sliding window"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Uniform text streams"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Simple; may create redundant data at boundaries"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Semantic split"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Content-driven boundaries (sentence embeddings)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Preserves topical coherence; computationally heavier"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Custom heuristics"},{"type":"text","text":" (headings, markdown)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Domain-specific docs"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Requires domain knowledge; very effective for structured docs"}]}]}]}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"7.4 What to Report When Asked About Chunking"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"When explaining your RAG chunking approach (interviews or documentation):"}]},{"type":"orderedList","attrs":{"start":1,"type":null},"content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Chunk size chosen"},{"type":"text","text":" — e.g., 500 tokens"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Overlap amount"},{"type":"text","text":" — e.g., 100 tokens"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Method used"},{"type":"text","text":" — e.g., recursive character splitter"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Rationale"},{"type":"text","text":" — why this combination suits your document type and latency requirements"}]}]}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"# Semantic chunking example (preserves topic boundaries)\nfrom langchain_experimental.text_splitter import SemanticChunker\nfrom langchain.embeddings import HuggingFaceEmbeddings\n\nembed_model = HuggingFaceEmbeddings(model_name=\"sentence-transformers/all-MiniLM-L6-v2\")\n\nsemantic_splitter = SemanticChunker(\n    embeddings=embed_model,\n    breakpoint_threshold_type=\"percentile\",\n    breakpoint_threshold_amount=95    # split when similarity drops below 95th percentile\n)\n\nsemantic_chunks = semantic_splitter.split_text(raw_document)\nprint(f\"Semantic chunks: {len(semantic_chunks)}\")\n"}]},{"type":"horizontalRule"},{"type":"heading","attrs":{"textAlign":null,"level":2},"content":[{"type":"text","text":"8. RAG Performance & Latency Optimization"}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"8.1 Identifying the Bottleneck"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"RAG latency can originate from two places:"}]},{"type":"codeBlock","attrs":{"language":"javascript"},"content":[{"type":"text","text":"Total response time = Retrieval time + Generation time\n\nRetrieval time: vector DB similarity search\nGeneration time: LLM inference on retrieved context\n"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Diagnostic steps"},{"type":"text","text":":"}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"import time\n\ndef timed_rag_query(query: str, retriever, llm) -> dict:\n    \"\"\"Instrument RAG pipeline to identify latency source.\"\"\"\n    # Measure retrieval\n    t0 = time.time()\n    docs = retriever.get_relevant_documents(query)\n    retrieval_time = time.time() - t0\n\n    # Measure generation\n    t1 = time.time()\n    context = \"\\n\".join([d.page_content for d in docs])\n    answer = llm.predict(f\"Context:\\n{context}\\n\\nQuestion: {query}\")\n    generation_time = time.time() - t1\n\n    return {\n        \"answer\": answer,\n        \"retrieval_ms\": round(retrieval_time * 1000, 2),\n        \"generation_ms\": round(generation_time * 1000, 2),\n        \"total_ms\": round((retrieval_time + generation_time) * 1000, 2),\n        \"chunks_retrieved\": len(docs)\n    }\n\nresult = timed_rag_query(\"What are the key product features?\", retriever, llm)\nprint(f\"Retrieval: {result['retrieval_ms']}ms | Generation: {result['generation_ms']}ms\")\n"}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"8.2 Top-k Impact on Latency"}]},{"type":"customImage","attrs":{"src":"https://67dc5f63b8b6a6f0ea39b99a.storage.fermion.app/public-cached/markdown-editor-file/file-unique-id/ubqoe8tncp5vijcqym2q-rqgitxmzsvcn9ixnxac1hsma0h_ek3nnkcmbjvu_r-mfcwqfsvpkw3l9kx7b7d8eeiiw6vnfbkqkodz","alt":null,"width":600,"height":327}},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Figure: Increasing k raises retrieval coverage but also increases the LLM's context processing time."}]},{"type":"table","content":[{"type":"tableRow","content":[{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"k value"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Retrieval speed"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Answer quality"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"LLM processing"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"k = 2–3"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Fast"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"May miss relevant info"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Minimal context"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"k = 4–5"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Balanced"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Good for most cases"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Manageable"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"k = 10+"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Slower"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Higher recall"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Significantly more tokens"}]}]}]}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"# Test different k values\nfrom langchain.vectorstores import FAISS\nfrom langchain.embeddings import HuggingFaceEmbeddings\n\nembed_model = HuggingFaceEmbeddings(model_name=\"all-MiniLM-L6-v2\")\nvectorstore = FAISS.load_local(\"./index\", embed_model)\n\nfor k in [2, 4, 6, 10]:\n    retriever = vectorstore.as_retriever(search_kwargs={\"k\": k})\n    result = timed_rag_query(\"What is the marketing budget?\", retriever, llm)\n    print(f\"k={k}: {result['total_ms']}ms | chunks={result['chunks_retrieved']}\")\n"}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"8.3 Caching Frequent Queries"}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"from functools import lru_cache\nimport hashlib\n\n# In-memory cache for common queries\nquery_cache = {}\n\ndef cached_rag_query(query: str) -> str:\n    query_hash = hashlib.md5(query.encode()).hexdigest()\n\n    if query_hash in query_cache:\n        print(\"Cache hit!\")\n        return query_cache[query_hash]\n\n    result = run_rag_pipeline(query)\n    query_cache[query_hash] = result\n    return result\n"}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"8.4 Latency Troubleshooting Checklist"}]},{"type":"codeBlock","attrs":{"language":"javascript"},"content":[{"type":"text","text":"Check logs → Identify bottleneck (retrieval vs generation)\n      ↓\nIf retrieval is slow:\n  → Reduce k value (e.g., 5 → 3)\n  → Reduce chunk size (smaller chunks = faster similarity search)\n  → Re-index with optimized embeddings\n      ↓\nIf generation is slow:\n  → Switch to lower-latency model (gpt-4o-mini vs gpt-4o)\n  → Enable parallel inference\n  → Reduce context window (fewer chunks passed to LLM)\n      ↓\nApply caching for frequent queries\n      ↓\nRe-test latency; compare against baseline\n"}]},{"type":"horizontalRule"},{"type":"heading","attrs":{"textAlign":null,"level":2},"content":[{"type":"text","text":"9. Evaluation Matrix: RAG vs Fine-Tuned Models"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"When a fine-tuned model underperforms the base RAG system, a structured evaluation matrix reveals where the gap lies."}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"9.1 Evaluation Dimensions"}]},{"type":"table","content":[{"type":"tableRow","content":[{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Dimension"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Description"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"How to Measure"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Relevance"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Does the answer address the query?"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Human rating (1–5) or embedding similarity to query"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Correctness"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Factual accuracy of the content"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Compare against ground-truth answers"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Fluency"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Naturalness of language"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"BLEU/ROUGE or human evaluation"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Latency"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Time to produce the answer"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"time.time()"},{"type":"text","text":" instrumentation"}]}]}]}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"9.2 Building the Evaluation Matrix"}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"import pandas as pd\n\ndef evaluate_system(system_name: str, rag_pipeline, test_questions: list, ground_truths: list) -> pd.DataFrame:\n    \"\"\"Run evaluation across test questions and record scores.\"\"\"\n    results = []\n\n    for q, gt in zip(test_questions, ground_truths):\n        answer = rag_pipeline(q)\n\n        # Simple keyword-overlap relevance (replace with LLM judge for production)\n        relevance = len(set(q.lower().split()) & set(answer.lower().split())) / len(q.split())\n\n        results.append({\n            \"system\": system_name,\n            \"question\": q,\n            \"answer\": answer[:100] + \"...\",\n            \"relevance\": round(relevance, 3),\n            \"latency_ms\": rag_pipeline.last_latency_ms\n        })\n\n    return pd.DataFrame(results)\n\n# Compare base RAG vs fine-tuned\nbase_df = evaluate_system(\"Base RAG\", base_rag_pipeline, test_q, ground_truth)\nft_df = evaluate_system(\"Fine-tuned\", ft_pipeline, test_q, ground_truth)\n\ncomparison = pd.concat([base_df, ft_df])\nprint(comparison.groupby(\"system\")[[\"relevance\", \"latency_ms\"]].mean())\n"}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"9.3 Remedial Actions When Fine-Tuned Model Underperforms"}]},{"type":"table","content":[{"type":"tableRow","content":[{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Gap Identified"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Remedial Action"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Missing domain terminology"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Increase fine-tuning data with domain-specific examples"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Over-fitting"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Reduce epochs; add regularization; augment with synthetic examples"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Insufficient training data"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Augment with synthetic examples generated by a stronger LLM"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Model too small for the task"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Try a larger base model or switch adaptation technique"}]}]}]}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Adaptation technique alternatives"},{"type":"text","text":":"}]},{"type":"table","content":[{"type":"tableRow","content":[{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Technique"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Core Idea"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Memory Usage"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"LoRA"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Low-rank matrices added to frozen weights"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Low"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"QLoRA"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"4-bit quantized backbone + LoRA"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Very low"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"IA³"},{"type":"text","text":" (Infused Adapter)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Modifies only bias/scale terms"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Minimal"}]}]}]}]},{"type":"horizontalRule"},{"type":"heading","attrs":{"textAlign":null,"level":2},"content":[{"type":"text","text":"10. AWS Bedrock Guardrails"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"AWS Bedrock Guardrails provide content moderation and safety policy enforcement for LLM-powered applications."}]},{"type":"customImage","attrs":{"src":"https://67dc5f63b8b6a6f0ea39b99a.storage.fermion.app/public-cached/markdown-editor-file/file-unique-id/acojndhnhdmq_qxsl6wpo3ys1dt5hw4x9idnddgv0hcz9xeosxbvv9fximuq9wzxit5evyak4qpmdcrcrll6bcgqr2eyqhsxtiaa","alt":null,"width":600,"height":327}},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Figure: Bedrock Guardrails intercept both user inputs and model outputs, applying configurable content policies."}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"10.1 What Guardrails Do"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Content moderation thresholds"},{"type":"text","text":" — flag or block profanity, hate speech, and disallowed topics"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Custom topic policies"},{"type":"text","text":" — define allowed/blocked domains (e.g., \"only answer questions about hydration products\")"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Input filtering"},{"type":"text","text":" — intercept prompt injection attempts before they reach the LLM"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Output rewriting"},{"type":"text","text":" — replace harmful outputs with safe fallback responses"}]}]}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"10.2 Guardrail Diagnostics"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"When RAG outputs seem filtered or altered unexpectedly:"}]},{"type":"orderedList","attrs":{"start":1,"type":null},"content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Go to "},{"type":"text","marks":[{"type":"bold"}],"text":"AWS Bedrock Console → Guardrails"},{"type":"text","text":" → inspect active policies"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Review the "},{"type":"text","marks":[{"type":"bold"}],"text":"trace logs"},{"type":"text","text":" to see which policy triggered"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Adjust thresholds — sensitivity levels are configurable per category (None → Low → Medium → High)"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Test with adversarial inputs to verify guardrail behavior before production"}]}]}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"import boto3\n\nbedrock_client = boto3.client(\"bedrock-runtime\", region_name=\"us-east-1\")\n\ndef invoke_with_guardrail(prompt: str, guardrail_id: str) -> str:\n    response = bedrock_client.invoke_model(\n        modelId=\"anthropic.claude-3-sonnet-20240229-v1:0\",\n        guardrailIdentifier=guardrail_id,\n        guardrailVersion=\"DRAFT\",\n        body={\n            \"anthropic_version\": \"bedrock-2023-05-31\",\n            \"max_tokens\": 512,\n            \"messages\": [{\"role\": \"user\", \"content\": prompt}]\n        }\n    )\n\n    result = response[\"body\"].read()\n    # If guardrail triggers: result contains intervention notice, not model output\n    return result\n"}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"10.3 Guardrail Pre-Deployment Checklist"}]},{"type":"codeBlock","attrs":{"language":"javascript"},"content":[{"type":"text","text":"1. Define allowed topics and blocked categories in the Bedrock console\n2. Test with adversarial prompts (prompt injection, jailbreak attempts)\n3. Verify fallback messages for blocked content\n4. Review trace logs for false positives (over-blocking legitimate queries)\n5. Adjust sensitivity thresholds based on test results\n6. Document guardrail version in deployment notes\n"}]},{"type":"horizontalRule"},{"type":"heading","attrs":{"textAlign":null,"level":2},"content":[{"type":"text","text":"11. Enterprise ERP Integration with Gen-AI Pipelines"}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"11.1 Why ERP Integration is Complex"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"ERP systems (e.g., SAP, Oracle) store structured business data in complex, highly normalized relational schemas. Feeding this directly into a vector DB is inefficient and often incorrect — vector DBs return similarity scores, not exact rows."}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Correct architecture"},{"type":"text","text":":"}]},{"type":"codeBlock","attrs":{"language":"javascript"},"content":[{"type":"text","text":"ERP (SAP/Oracle)\n      │\n      ▼ (ODBC/JDBC connectors or SAP APIs)\nExtract tables → Transform to plain text / JSON\n      │\n      ▼\nChunk + Embed metadata → Store in Vector DB\n      │\n      ▼ (at query time)\nUser Query → Vector DB retrieves relevant metadata\n      │\n      ▼\nLLM generates SQL from metadata context\n      │\n      ▼\nSQL executed against original ERP/relational DB → Exact rows returned\n      │\n      ▼\nLLM synthesizes final answer from exact rows\n"}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"11.2 Step-by-Step ERP Integration"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Step 1: Extract data via connectors"}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"import pyodbc  # ODBC connector for SAP/Oracle\n\nconnection_string = (\n    \"DRIVER={SAP HANA};SERVER=hana-host;PORT=30015;\"\n    \"UID=your_user;PWD=your_password;DATABASE=DEMO\"\n)\nconn = pyodbc.connect(connection_string)\n\n# Extract relevant columns as metadata (NOT full tables)\ncursor = conn.cursor()\ncursor.execute(\"\"\"\n    SELECT product_id, product_name, category, description, price\n    FROM sap_products\n    WHERE category = 'WaterBottles'\n\"\"\")\nrows = cursor.fetchall()\n"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Step 2: Transform to text for chunking"}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"def row_to_text(row) -> str:\n    \"\"\"Convert a structured row to a text description for embedding.\"\"\"\n    return (\n        f\"Product: {row.product_name} | \"\n        f\"Category: {row.category} | \"\n        f\"Price: ${row.price} | \"\n        f\"Description: {row.description}\"\n    )\n\nmetadata_texts = [row_to_text(r) for r in rows]\n"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Step 3: Embed and store metadata"}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"from sentence_transformers import SentenceTransformer\nimport chromadb\n\nembed_model = SentenceTransformer('all-MiniLM-L6-v2')\nchroma_client = chromadb.PersistentClient(path=\"./erp_metadata\")\ncollection = chroma_client.get_or_create_collection(\"sap_products\")\n\nfor i, (row, text) in enumerate(zip(rows, metadata_texts)):\n    embedding = embed_model.encode(text).tolist()\n    collection.add(\n        ids=[str(row.product_id)],\n        embeddings=[embedding],\n        documents=[text],\n        metadatas=[{\"product_id\": str(row.product_id), \"table\": \"sap_products\"}]\n    )\n"}]},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Step 4: Query — generate SQL and execute"}]},{"type":"codeBlock","attrs":{"language":"python"},"content":[{"type":"text","text":"from openai import OpenAI\n\nclient = OpenAI()\n\ndef erp_rag_query(user_question: str) -> str:\n    # Step 4a: Retrieve relevant metadata\n    query_emb = embed_model.encode(user_question).tolist()\n    results = collection.query(query_embeddings=[query_emb], n_results=3)\n    metadata_context = \"\\n\".join(results[\"documents\"][0])\n\n    # Step 4b: LLM generates SQL\n    sql_prompt = f\"\"\"\n    The user asked: \"{user_question}\"\n    Relevant metadata from SAP:\n    {metadata_context}\n\n    Write a SQL query to answer this question. Use table: sap_products.\n    Return ONLY the SQL, no explanation.\n    \"\"\"\n    sql_response = client.chat.completions.create(\n        model=\"gpt-4o-mini\",\n        messages=[{\"role\": \"user\", \"content\": sql_prompt}]\n    )\n    generated_sql = sql_response.choices[0].message.content.strip()\n\n    # Step 4c: Execute SQL against original ERP database\n    cursor.execute(generated_sql)\n    exact_rows = cursor.fetchall()\n\n    # Step 4d: Synthesize final answer\n    synthesis_prompt = f\"\"\"\n    User question: {user_question}\n    Database results: {exact_rows}\n    Write a clear, concise answer.\n    \"\"\"\n    final = client.chat.completions.create(\n        model=\"gpt-4o-mini\",\n        messages=[{\"role\": \"user\", \"content\": synthesis_prompt}]\n    )\n    return final.choices[0].message.content\n\n\nanswer = erp_rag_query(\"What water bottle products are available under $25?\")\nprint(answer)\n"}]},{"type":"heading","attrs":{"textAlign":null,"level":3},"content":[{"type":"text","text":"11.3 Data Governance Considerations"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Ensure compliance with corporate data policies before feeding ERP data into any LLM"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Implement role-based access: only authorized columns/tables should be accessible to the AI pipeline"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Some ERP vendors (e.g., SAP) provide Gen-AI connectors — verify compatibility before building custom integrations"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"One-to-one schema mapping may be required for complex ERP schemas before text conversion"}]}]}]},{"type":"horizontalRule"},{"type":"heading","attrs":{"textAlign":null,"level":2},"content":[{"type":"text","text":"Summary"}]},{"type":"table","content":[{"type":"tableRow","content":[{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Topic"}]}]},{"type":"tableHeader","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Key Takeaway"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Sequential agent design"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Split complex tasks into specialized agents; proxy mediates every handoff"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"UserProxy"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Central mediator — handles approval loops, carry-over, and termination"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"initiate_chat params"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"code"}],"text":"clear_history"},{"type":"text","text":", "},{"type":"text","marks":[{"type":"code"}],"text":"silent"},{"type":"text","text":", "},{"type":"text","marks":[{"type":"code"}],"text":"carry_over"},{"type":"text","text":", "},{"type":"text","marks":[{"type":"code"}],"text":"finish_chat_index"},{"type":"text","text":", "},{"type":"text","marks":[{"type":"code"}],"text":"summary_method"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Content iteration"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Proxy feedback → agent regenerates → repeat until satisfied"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Frontend integration"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"FastAPI backend + React frontend; 3 buttons map to proxy actions"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Chunking"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"200–500 tokens, 50-token overlap; method depends on document type"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Latency optimization"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Instrument retrieval vs generation; tune k and chunk size; add caching"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Evaluation matrix"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Compare relevance, correctness, fluency, latency across RAG vs fine-tuned"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Bedrock Guardrails"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Content moderation + custom topic policies; test adversarially pre-deployment"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"ERP integration"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null},"content":[{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","text":"Metadata in vector DB → LLM-generated SQL → exact rows from relational DB"}]}]}]}]},{"type":"horizontalRule"},{"type":"paragraph","attrs":{"textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Generated from session notes on AI Marketing Automation — AutoGen Pipelines, RAG Optimization & ERP Integration."}]}]}</article></body></html>