ScreenMeet Documentation Assistant
Grounded answers from published product knowledge.
After ScreenMeet was quoted $199 per month for a chatbot, I built and production-hardened one instead. It answers product questions from published knowledge, links readers to its sources, supports follow-up questions, and declines when the available evidence is not strong enough.
The problem
The project began as a practical buy-versus-build decision. I used the quoted $199-per-month product as the baseline, built the narrower experience we needed, and integrated it into the existing Document360 site through custom HTML.
Documentation search makes readers translate a question into keywords, inspect several articles, and assemble an answer. A useful assistant should shorten that path without inventing product behavior or hiding where an answer came from.
I treated retrieval quality, citations, refusal behavior, privacy, and abuse resistance as parts of the product—not cleanup work after the chat interface was finished.
How it works
- Grounded retrieval: finds relevant published documentation and supplies only that evidence to the answer step.
- Source visibility: returns links to the supporting articles and filters unsafe or untrusted destinations.
- Conversation: keeps enough bounded history for useful follow-up questions without allowing the prompt to grow indefinitely.
- Honest limits: refuses when retrieval is empty or insufficient instead of presenting a plausible guess as product guidance.
Support knowledge model
I also analyzed historical ESD support tickets into a practical troubleshooting model: issue, observable signals, likely causes, diagnostic steps, and verified resolution. The structure makes a support path inspectable instead of treating a closed ticket as an unsearchable answer.
Customer documentation and internal diagnostic knowledge have different audiences and risk. I designed the retrieval architecture accordingly: Document360 remains the customer-facing source, while Jira and Confluence provide the internal support context that Rovo can retrieve in the agent workflow.
This separation keeps public answers grounded in published guidance while preserving the richer investigation detail that support engineers need to diagnose an escalation.
Production safeguards
The public surface uses reCAPTCHA v3 to verify the first message, rate-limits sessions atomically, bounds requests and conversation history, redacts sensitive prompt content, and applies restrictive browser security headers. Follow-up messages remain conversational without requiring another CAPTCHA check. Data access and retention are limited to what the assistant needs.
I verified the deployed system with real documentation questions, a follow-up conversation, direct abuse-path requests, dependency auditing, database checks, and production runtime monitoring.
Skills used
Applied AI and support knowledge: retrieval design, prompt boundaries, citations, refusal behavior, conversation context, and structured diagnostic evidence.
Security and reliability: bot controls, rate limits, privacy minimization, link validation, dependency review, and live verification.
Delivery: Next.js, Supabase, database migrations, Vercel configuration, and production testing.