Every year, Atlassian organizes a hackathon called codegeist unleashed. This year it's fully devoted to applications of AI on Atlassian's products, another demonstration of Atlassian's vision and commitment to AI.

For the hackathon, we set out to implement the ability to operate on structured data. As we tested and evaluated our first prototypes, we liked the results – sometimes they're even better than Excel's AI! – but we also realized that just spitting out a number wasn't going to cut it for users to be able to rely on it. So, almost as a side result of the hackathon, we also came up with X-ray: a complete description of the information retrieved and steps taken to generate the answer.

We are proud of what we've produced for this contest, and while we would love the recognition, participating was worth it regardless of the outcome.

This is the slide deck for our submission:

As we prepared the submission, we realized we were writing the same over and over in different places, and of course we couldn't help but take advantage of AI. One of the fields in the submission form asks for a brief story of the project. So we uploaded the slide deck to Claude and prompted it with:

The attached document is a presentation for a submission to the
Atlassian CodeGeist Hackathon. Please, write a blog post about
the story of our project based on the presentation. Please write
about: what inspired us, what we learned, how we built the project,
and the challenges we faced. Format the story in Markdown.

What follows is a contribution from our first guest blogger, courtesy of Anthropic.

The Story Behind Connie AI - An AI Assistant for Confluence

As former Atlassian employees, we experienced firsthand how challenging it can be to navigate all the information in Confluence. My co-founder and I envisioned an AI assistant that could understand natural language questions about Confluence content and provide direct answers by querying tables and databases. This inspiration led us to create Connie AI, an AI assistant designed specifically for Confluence.

We first launched Connie on the Atlassian Marketplace in June 2023. For Atlassian's 2023 CodeGeist hackathon, our team at Applied Language Understanding (ALU) implemented two major new features: natural language table querying and X-ray explainability.

Building Natural Language Table Querying

Enabling Connie to understand and query tables required a complex AI pipeline. Connie automatically indexes new and updated Confluence pages, decomposing them into semantic chunks. Chunks containing structured data like tables are flagged.

When a user asks a natural language question, Connie retrieves the most relevant chunks. If any chunks have structured data, Connie uses GPT-3.5 to convert the question into a custom pseudo-SQL query language. This allows Connie to perform powerful operations like unit conversions and date reformatting on the fly.

We built a custom engine to execute these AI-generated queries while generating explanations of the operations performed. The output is passed to GPT-3.5 to generate a final human-readable answer. Streaming responses directly to the client reduces perceived latency.

Adding X-ray Explainability

We realized trust would be a challenge - users needed more than just a magical answer from an AI assistant. Our solution was X-ray, a feature that shows the user every retrieved fact, reasoning step, and operation behind each answer. This helps build user trust and aids in detecting any potential errors.

X-ray is an innovative capability - Connie is the only system we know of that provides this level of transparency into an AI's process.

Overcoming Challenges with Forge

Building Connie's complex pipeline required moving beyond Forge's robust default security. We implemented cryptographic signing for all communications and enforced user access rights at runtime. We added easy opt-outs for any potentially shared data.

Forge's spacePage, macro, and customUI modules allowed Connie to integrate naturally into Confluence. We were even able to add support for dark-mode in just a day.

What's Next for Connie AI

Looking ahead, our priorities are bringing Connie's table analysis to Confluence databases, expanding the capabilities of the query language, and improving latency and reliability by bringing inference in-house.

The Connie AI project shows how AI can thoughtfully augment human intelligence. We overcame key challenges to build an assistant that provides powerful capabilities and unprecedented transparency. Connie AI represents the exciting future potential of AI to aid knowledge workers.