Enterprise Knowledge AI Assistant
AWS Bedrock · Claude Sonnet 4.6 · Amazon Titan · RDS PostgreSQL + pgvector · S3
An AI-powered assistant that lets employees search through hundreds of operational documents by simply asking questions in plain English — and get accurate, sourced answers in seconds.
The Problem
Large organizations store critical knowledge across hundreds of PDFs, procedures, manuals, and engineering documents. When employees need answers, they spend hours manually searching through files — often finding nothing, or finding outdated information.
At Duke Energy, operational teams work with 1,000+ documents covering transformer health, asset management, safety protocols, and wildfire risk. Finding the right information quickly is a real operational challenge.
The Solution
Instead of searching through documents manually, employees can now just ask a question — like "What is the health status of transformer TF-554867?" — and get an accurate answer with the exact source documents cited.
The system reads and indexes all documents, understands the meaning of questions (not just keywords), finds the most relevant content, and uses Claude AI to generate a clear, grounded answer.
How It Works
Two simple pipelines — one that indexes documents, one that answers questions.
Step 1 — Indexing Documents
Runs once when new documents are added
Documents uploaded to AWS S3
Secure cloud storage
System reads & splits documents
Into small meaningful sections
Amazon Titan AI converts each section
Into a mathematical embedding
Embeddings stored in RDS PostgreSQL
Searchable by meaning, not just keywords
Step 2 — Answering Questions
Runs every time a user asks a question
User types a question in plain English
Natural language input
Question converted to the same format
Mathematical embedding via Amazon Titan
System finds 5 most relevant sections
Semantic similarity search via pgvector
Claude Sonnet reads those sections
Via AWS Bedrock — writes a clear answer
Answer returned with source citations
Users can verify the information
Key Design Decisions
Every architectural choice was made for a reason.
Why pgvector instead of a dedicated vector database?
Keeping the vector search inside PostgreSQL means one database for everything — document content, metadata, and embeddings together. Simpler to manage, faster to query, and no extra service to maintain.
Why HNSW index instead of IVFFlat?
HNSW (a type of search index) works accurately at any data size. The alternative (IVFFlat) requires a minimum number of records to work correctly. HNSW was the right choice for a system that starts small and grows.
Why Amazon Titan instead of other embedding models?
Titan is built into AWS Bedrock — same login, same security, same billing. In enterprise environments, reducing external dependencies is a governance and security requirement.
Why chunk documents at 500-700 tokens?
Too large and the search becomes imprecise. Too small and the chunks lose context. 500-700 tokens hits the sweet spot — enough context for accurate answers, small enough for precise retrieval.
What Powers It
Enterprise-grade AWS infrastructure — the same stack used in production environments.
AWS Bedrock
Secure, enterprise AI platform that runs Claude Sonnet — no data leaves your AWS environment
Claude Sonnet 4.6
Anthropic's enterprise AI model — reads the retrieved documents and writes accurate, clear answers
Amazon Titan Embeddings
Converts text into mathematical vectors that capture meaning — enabling search by concept, not just keywords
RDS PostgreSQL + pgvector
Enterprise database that stores both document content and AI embeddings — enables fast semantic search
AWS S3
Secure cloud storage for all source documents — encrypted, access-controlled, enterprise-grade
Node.js Pipeline
The ingestion engine that reads documents from S3, chunks them, generates embeddings, and loads them into the database
Try It Live
Ask any question about transformer health, wildfire risk, asset management, or operational procedures. Every answer comes directly from indexed enterprise documents — with source citations.
Hello! I am the Enterprise Knowledge AI Assistant. Ask me anything about the indexed operational documents — transformer health, wildfire risk, procedures, or asset management.
Why This Matters
Saves Hours of Search Time
What used to take 30-60 minutes of manual document searching now takes seconds. Employees get accurate answers instantly.
Answers You Can Trust
Every response cites the exact source document it came from. No guessing, no hallucinations — just grounded, verifiable answers.
Works on Any Document Set
The same system works for any organization's documents — maintenance manuals, safety procedures, HR policies, legal contracts, or financial reports.
Upload Your Own Document
Upload any PDF or text file and ask questions about it. Prototype mode — document processed for session only.
Drag & drop your document
or click to browse
PDF, TXT, DOC accepted
If This Was at Enterprise Scale
How this system would evolve for a Fortune 500 deployment.
Partition the database by document category. Move to Aurora Serverless for automatic scaling. Add hybrid search combining keyword and semantic matching for even better accuracy.
Add row-level security so each organization only sees their own documents. Separate storage per client. Individual user authentication per organization.
Add an automatic screening step before indexing that detects and removes personal information using Amazon Comprehend. Store a full audit log of every query for compliance.
Want to discuss this architecture?
I'm available to talk through RAG systems, AWS Bedrock, or enterprise AI adoption.