Helvia
Most support chatbots are a black box: a user asks a question, and an answer appears with no visibility into whether the bot actually understood the question or just pattern-matched to the nearest FAQ entry. I wanted to explore what happens when you make that reasoning visible, surfacing the agent's inferred intent, sentiment, and retrieved sources in real time instead of hiding them behind the chat bubble.
The technical challenge was wiring three live surfaces to the same conversation: a retrieval-augmented knowledge base lookup with match-confidence scoring, a running sentiment and context classification per turn, and the actual chat, all updating in sync as a single message streams in, rather than bolting on an after-the-fact analytics dashboard.
Black-Box Chatbots
Standard support bots give an answer with no visibility into what the model actually understood or retrieved.
Three Live Surfaces, One Turn
Reasoning, sentiment, and retrieved sources all need to update in sync as a single message streams in.
Retrieval Confidence
Every knowledge base match needed a visible confidence score, not just a silently-injected snippet of context.
I built and ran this locally by adapting Anthropic's open-source "Customer Support Agent" quickstart, a Next.js app that pairs a Claude-powered chat with two live side panels. The left panel surfaces the assistant's inferred intent and sentiment for each turn; the right panel shows the knowledge base sources it retrieved, each with a match-confidence percentage, sourced via Retrieval-Augmented Generation. Nothing is hidden behind the chat bubble, every inference the agent makes is visible as it happens.
Live Reasoning Panel
Every turn surfaces the assistant's inferred sentiment and context classification, tagged neutral, positive, or negative alongside a confidence indicator.
RAG Knowledge Retrieval
Answers are grounded in retrieved knowledge base entries, each shown with its match percentage instead of a black-box citation.
User Mood Detection
The agent tracks conversation sentiment turn by turn and can redirect to a different flow when it detects frustration.
Three-Pane Transparency Layout
Reasoning, chat, and sources live side by side in one view, so nothing about the agent's decision-making is hidden after the fact.
A Closer Look
Live Agent Interface (Empty State)
Agent Reasoning, RAG Retrieval & Sentiment in Action
As a personal exploration rather than a client build, Helvia demonstrated what a genuinely transparent support agent looks like: every retrieved source, every sentiment read, and every context decision visible in real time rather than buried in logs. It's a working template for building trust into an AI agent's interface, not just its answers.
“I wanted to see what a support agent looks like when you stop hiding its reasoning. Watching the sentiment and retrieval panels update turn by turn, in sync with the actual chat, made it obvious how much trust a black-box chatbot throws away by hiding that exact information.”