Build / Web development

Confluence-Powered AI Chatbot

Streamlining HR support with a Confluence-powered AI chatbot.

Our internal HR department was spending significant time answering recurring employee questions, despite all information being available on our company's Confluence wiki. To solve this, we developed a proof-of-concept AI chatbot that securely connects to our Confluence data. This tool provides employees with instant, summarized answers and direct links to source pages, significantly reducing the HR team's workload.

Services

Web development

Tech

React

Challenge

The primary challenge was twofold: first, to relieve our HR department from the burden of repeatedly answering simple questions about leave requests, company car policies, or sick leave procedures. Second, we wanted to build an AI chatbot that could securely access and understand our private, internal Confluence knowledge base without requiring a complete, custom-trained LLM model. The goal was to create an intelligent, self-service tool that could provide accurate information on demand.

The solution

We engineered an intelligent agent using a Retrieval-Augmented Generation (RAG) architecture. This approach allows the chatbot to fetch information from our existing Confluence pages and use it to generate relevant, natural-language answers. The solution is built entirely on AWS, with a cost-effective data pipeline.

AI Agent on AWS Bedrock: The core of the chatbot is an AI agent built in AWS Bedrock, utilizing Amazon's Nova Pro model for understanding natural language.
RAG with Pinecone Vector Store: Instead of a costly direct integration, we implemented a system where a simple script scrapes our Confluence space, stores the content in an AWS S3 bucket, and syncs it to a Pinecone vector store. This provides a fast and affordable knowledge base.
Intelligent Orchestration: The agent uses pre-processing to sanitize user queries and post-processing to format answers neatly in Markdown, always including a link to the source Confluence page.
Modern Tech Stack: The user-facing application consists of a React frontend and a Node.js backend that communicates securely with the AWS Bedrock agent.

Our approach

Our process began with research to identify the best technical architecture, quickly landing on the now commonly used RAG system as the ideal solution. We prioritized using AWS Bedrock to explore its capabilities for building enterprise-grade AI agents.

Initial Proof of Concept: We first built a version using AWS OpenSearch for the vector store. However, we discovered this option was prohibitively expensive for our use case.

Iteration & Optimization: We pivoted to a more cost-effective architecture using Pinecone as the vector database. This required creating a custom data pipeline: a script scrapes Confluence and pushes JSON files to an S3 bucket. The content was then converted into vector embeddings using Amazon's Titan Text Embeddings V2 model before being synced with Pinecone.

Versioning & Testing: The agent is managed with alias IDs in AWS Bedrock, allowing us to develop and test new versions of the agent internally without affecting the live chatbot.

The outcome

The project successfully demonstrated our ability to create secure, knowledgeable AI chatbots for internal use. The proof-of-concept evolved into a fully functional internal tool that has been rolled out for our team, validating the effectiveness of the RAG approach and the AWS Bedrock platform. As a result, the chatbot now fields common employee questions, significantly reducing the HR workload and freeing up the team to focus on more complex tasks.

Furthermore, we have established a reusable and cost-effective architectural pattern for connecting AI agents to any internal data source, such as PDFs, text or HTML files, by leveraging an S3 bucket as the central data source.

By optimizing our architecture for both capability and cost, we built an intelligent tool that was truly sustainable for company-wide deployment.

Yorrick Schoonheydt AI Lead at icapps

Key features

Natural Language Interface: Employees can ask questions in plain English and receive conversational, accurate answers.

Automated Knowledge Sync: A daily script ensures the chatbot's knowledge base is always up-to-date with the latest information from Confluence.

Source-Cited Answers: Every answer includes a direct link to the detailed Confluence page, ensuring transparency and providing a path to more comprehensive information.

Let's create a digital product together!

Contact