Democratizing institutional knowledge: Building an AI-powered knowledge management system with AWS
Amazon Web Services Nneoma Okoroafor
AWS built a knowledge system that talks and listens. It keeps company know-how from walking out the door with retiring staff.
Based on reporting by Amazon Web Services, Nneoma Okoroafor — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AWS is pitching a new way to keep institutional knowledge from evaporating when experienced people leave. The idea is simple enough: instead of burying procedures in stale docs and scattered folders, build a cloud-based system that lets workers ask questions by voice or text and get answers grounded in the company’s own files.
The setup leans on a browser interface, an avatar layer, and a stack of AWS services underneath. Amazon Cognito handles sign-in, API Gateway controls access, and Amazon Bedrock Knowledge Bases pulls answers from documents stored in Amazon S3. Those documents are chunked, embedded with Amazon Titan Text Embeddings, and indexed in an Amazon OpenSearch Serverless vector store. DynamoDB caches responses, and Lambda ties the workflow together.
The voice part is not an afterthought. Amazon Transcribe turns speech into text, Amazon Polly speaks the reply, and the avatar system provides the visual wrapper. AWS says the same system can be used in manufacturing, healthcare, financial services, energy, and government, with particular appeal for people in control rooms, training rooms, quality labs, and maintenance-planning offices who need hands-free access.
What makes this accelerator different, AWS argues, is that it’s built for non-technical users and knowledge owners alike. People uploading content do not need to redesign it or build a retrieval pipeline by hand; they can drop in Word, PDF, plain text, Markdown, or JSON files, and the system ingests them after a sync. New documents are queryable shortly after upload, though not instantly. An optional AWS Glue ETL job can convert source material into Markdown or structured JSON first.
The deployment story is meant to be quick, too. AWS says the prototype can go out through CloudFormation in hours. It also includes a caching layer that can reuse answers for repeated questions, and in testing the company says cache hit rates of 50% to 70% were possible when the workload was full of repeat queries. There is, however, a standing cost floor: the OpenSearch Serverless vector store bills on an always-on minimum and is described as the largest fixed component of the setup.
AWS frames this as the middle path between building a custom Bedrock system from scratch and using a plain text chatbot. The pitch is not subtle: get voice, avatars, retrieval, and caching without stitching the whole thing together yourself. For organizations worried about institutional memory walking out the door, that’s a pragmatic answer, not a flashy one. Which is probably why it works.
My take — AI-written commentary, not fact-checked reporting
This is the kind of AI project that actually makes sense: boring, useful, and aimed at people who do real work instead of demo theater. The avatar is the part that sounds most gimmicky, but the real story is the old one—companies keep losing knowledge because they never bothered to capture it properly. Funny how the future keeps showing up as better filing cabinets.
Read more about this at: Amazon Web Services