Posts

Simple understanding of Pega Knowledge Buddy functionality

Image
Pega Knowledge Buddy (KB) converts documents into searchable semantic chunks, stores their embeddings in a vector database, retrieves the most relevant chunks for a user's question, and gives those chunks to the LLM through RAG to generate a grounded answer. Step1: Ingestion — Build the Knowledge Base Document → Chunks → Embeddings → Vector Storage Source content Pega Knowledge articles PDFs/documents SharePoint content, etc. Chunking Knowledge Buddy breaks large content into smaller pieces called chunks . For example, the default SIZE approach uses 1,000 characters with 200-character overlap. Pega also supports TITLE and NONE chunking methods. Semantic/Embedding processing Each chunk is sent through the Pega GenAI Gateway . An embedding is generated — a numerical representation of the meaning of that chunk. Additional attributes/metadata can also be associated with the chunk, improving semantic retrieval. Vector storage The embeddings and associated chunk information are store...

Build Pega Applications with GitHub Copilot CLI, Pega Infinity AI Plugin & MCP — Branch-Based Agentic Authoring

Image
To work on this experiment, we need the Github Copilot CLI, Pega Infinity Instance, and Pega Infinity AI Plugins ( https://github.com/pegasystems/infinity-ai-plugins ) . In my case, I have used the Pega Infinity V26.1 Community Edition and here is the Github repository to upload it in the Github Copilot plugin marketplace. For this entire exercise to work well, we need Java 17 or above.  Windows Command Prompt (Agent-Driven Terminal): Represents the Github Copilot CLI. It shows the active terminal execution where commands like pega-setup, list-skills, list-available-applications etc. MCP (Model Context Protocol) Orchestrator: Acts as the bridge and mediator between the AI terminal environment (GitHub Copilot) and the Pega platform. It translates natural language directives and agent commands into structured operations and API calls to manage application configurations. Here the name of the MCP server is pega--infinity-authoring Workflow Directives: Represents...

Stop treating MCP, RAG, and Skills like alternatives.

Image
The confusion stems from the fact that all three "give AI capabilities." But in reality, they solve three completely different parts of the architecture: Data Access , Context Retrieval , and Action Execution . Here’s a breakdown of how it works: 1️⃣ MCP Connection (Real-Time Data) Before making any decision, the AI connects via Model Context Protocol (MCP) to fetch live data from internal databases, custom APIs, or CRM systems (like Pega). 2️⃣ RAG Knowledge Base (Context & Rules) The agent uses Retrieval-Augmented Generation (RAG) to pull company policies, legal regulations, and standard operating procedures (SOPs). This grounds the AI in enterprise context so it doesn't hallucinate. 3️⃣ Skills Execution (Automated Actions) Once the AI synthesizes live data + policy guidelines to make a decision, it triggers concrete Skills —such as updating case statuses, routing tickets, or dispatching real-time email notifications. 💡 Key Takeaway: AI Agents aren't...

My First YOLO Object Detection Experiment using Python with Multiple Images

Image
I recently tried a simple YOLO (You Only Look Once) object detection exercise using Python and the Ultralytics YOLO11 model. The goal was straightforward: instead of processing just one image, I wanted to provide multiple images , run object detection on all of them, and generate a single combined output image . What I Used Python Ultralytics YOLO YOLO11n pretrained model PIL/Pillow for image processing Multiple JPG images The Basic Approach First, I loaded the pretrained YOLO11 model: from ultralytics import YOLO model = YOLO("yolo11n.pt") I then provided multiple images: image_paths = [ r"D:\Python\images\dog.jpg", r"D:\Python\images\MyPic.jpg" ] results = model(image_paths) YOLO automatically performed object detection on each image and identified objects such as people, animals, vehicles, and other objects that the pretrained model recognizes. The interesting part was combining all the annotated images into one output image . This makes it ...

I Built a Team of AI "Employees" to Process Insurance Claims — In One Afternoon, For Free

Image
  The Problem With "One Big AI" Picture a health insurance claim landing on someone's desk. Before it gets approved or rejected, a human needs to: Check that all the paperwork is filled in correctly Confirm the treatment is actually covered by the policy Make sure nothing about the claim looks suspicious Understand the medical context well enough to judge if it makes sense Finally, make the call: approve, reject, or send it for closer review That's five different jobs, each needing a different kind of expertise. Now imagine trying to train one person to be equally good at all five. They'd be okay at everything and great at nothing. This is exactly the problem with asking a single AI prompt to "handle insurance claims." You get a jack-of-all-trades that misses details a specialist would catch. The fix? Don't hire one generalist. Hire a team. Meet the Team Here's what I actually built — a little assembly line of AI "employees,...