Key Outcomes
- Employees now get answers in seconds instead of waiting on calls or messages
- Centralized, searchable knowledge base for policies, measurements, and training
- Reduced support burden on senior staff and managers
- Faster onboarding for new employees
Background
Glass Doctor operates with extensive internal documentation, policies, procedures, training materials, and operational references critical to day-to-day work.
Before this project, employees relied on calls and internal messages to get answers. Knowledge lived with a handful of experienced staff, creating constant interruptions, delays, and inconsistent responses.
Leadership needed a secure, internal AI system that could turn existing documents into a reliable, always-available source of truth without exposing data externally or requiring complex workflows.
The Challenge
Business Challenges
- Support overload from repetitive internal questions
- Slow response times blocking employee productivity
- Difficult onboarding for new hires
- Knowledge sprawl across static documents with no effective search
Operational Pain Points
- Employees interrupted others for basic policy or measurement questions
- Inconsistent answers depending on who was contacted
- No visibility into what questions were being asked most often
Technical Constraints
- Internal-only system with strict access control
- AI responses must stay grounded in approved documents
- OCR needed for scanned PDFs
- Context window limits with large document sets
The Solution
Stackup Solutions designed and built Glass Doctor Chatbot, a secure, single-tenant, web-based AI knowledge assistant powered by Retrieval-Augmented Generation (RAG).
The platform converts internal documents into an intelligent, conversational interface while giving administrators full control over users, content, and AI behavior.
Architecture & Technical Decisions
Core Architecture
- Frontend: Next.js for fast, responsive UI
- Backend: Node.js for scalable API handling
- Database: MySQL for users, roles, and chat logs
- Vector Database: Pinecone for high-performance semantic search
- AI Layer: ChatGPT with strict RAG constraints
- Auth: Passport.js + JWT with role-based access
RAG Training Pipeline
- Admin uploads approved documents (PDF, DOC/DOCX)
- OCR extracts text from scanned files when required
- Content is chunked into controlled, context-safe segments
- Chunks are embedded and stored in Pinecone
- Documents become AI-ready automatically after upload
Query & Response Flow
- User asks a question in the chat interface
- Query is converted into an embedding
- Pinecone retrieves the most relevant document chunks
- Retrieved context is injected into a guarded system prompt
- ChatGPT generates an answer only from provided context
- If no relevant content exists, the AI safely declines to guess
This approach prevents hallucinations and ensures every response stays aligned with internal documentation.
Key Features
Admin Panel
- User management (Super Admin, Admin, Employees)
- Secure password resets and account control
- Document & media library with validation and upload progress
- AI training automation on document upload
- Read-only chat logs for auditing and improvement
- Usage metrics dashboard
Employee Experience
- Clean chat interface with conversation history
- Instant answers for policies, measurements, and procedures
- In-app document viewer for PDFs, DOCs, and images
- Secure authentication with profile management
Implementation Process
Discovery & Strategy
- Identified high-friction knowledge workflows
- Defined AI guardrails and security boundaries
Architecture & RAG Design
- Chunking strategy optimized for long operational documents
- OCR tuning to handle inconsistent scans
- Context window constraints carefully managed
Platform Build
- Full admin + user system with role-based access
- Automated AI training pipeline
- Secure logging and audit trails
Launch & Iteration
- Live production deployment
- Prompt refinement using real chat logs
- Continuous improvement without retraining the entire system
Results & Impact
While exact metrics were not formally tracked, the impact was immediate and clear:
- Significant reduction in internal messages and calls
- Employees now get answers in seconds instead of waiting
- Managers and senior staff reclaimed time from repetitive questions
- New hires onboard faster with self-service access to knowledge
- Leadership gained visibility into what employees actually ask
The system is live in production and is considered a successful resolution of the original support and knowledge challenges.
