Intelligent Prescriptive Guidance

Solution Architecture

Inspiration

Use Case Building a conversational AI application for a custom domain, public (or) private data which could be a knowledge hub to answer questions and also provide prescriptive guidance for the product or service adoption.

Market Opportunities

In the areas where it requires a prescriptive guidance for the products or service adoptions, such as user manuals, blog posts or even immersion days.
Providing interactive and guided coaching – Ex: eLearning courses with Video tutorials
Interactive service manuals – such as product installation instructions provided by Ikea

What it does

“Intelligent Prescriptive Guidance”: This application provides guidance to educate, build, deploy and operate your custom product or the service. Instead of traditional one way content like documentation or blog posts, this application will provide an interactive method of learning/building, powered by intelligent AI chatbot which provides the prescriptive guidance based on the changing users situations. It would be like a virtual buddy overlooking your shoulder and helping you through the entire learning or build journey. Each step in the guidance flow is dynamically chosen using traversal graph in GraphML based on the users navigation.

How we built it

This was a fantastic use case and had an opportunity to leverage the best of AI/ML services: #amazonbedrock, #huggingface (Large Language Models - #llm), #neptune (GraphML), #pinecone (Vector Database), #langchain (Prompt Engineering), #dynamodb (State Management and Chat persistence)

Retrieval Augmented Generation (RAG) based implementation with LangChain and SageMaker LLM models.
Dynamic adaption of the solution guidance workflow using GraphML depending on the users current state.
Fine-tuned Prompt Engineering with LangChain to provide: Use Case discovery (reasoning) Prescriptive Guidance for building solutions, Q&A and Error Troubleshooting
State management during the conversation
Multi Lingual support at the run time
Custom ontology analytics to check what is queried based on the geo location and focus our scaling efforts respectively

Challenges we ran into

Accomplishments that we're proud of

Understanding based applications which get adjusted to the questions asked, instead of Circuit based which have explicit and rigid structures.
Adaptability of the prescriptive guidance flows based on the feedback received.
Use Custom Ontology for Prompt Engineering and analytics.
State management of each user and provide a summary of the conversation.
Highly secured, the entire solution could be deployed in a private subnet and your custom private data does not leave your VPC.
Decoupled architecture which supports scaling independently at each layer and provides pluggability. Like you can chose the LLM models or the AWS services with out impacting the pipeline.
Extensibility, this solution could be easily extended to provide voice interface.
Multi Lingual, now we can have single knowledge base which can support multi lingual at the run time – instead of creating the documentation for each language and training individual models.