Today we are very excited to share a new open source project with you called Panoramica. Panoramica makes it easy for developers to work with conversational data, one of the most prevalent kinds on the internet. In particular, Panoramica helps leverage AI and Large Language Models (LLMs) in the context of single conversations, groups, or large panoramas like online communities or social networks.
Why did we build Panoramica?
Josh: At Orbit, most of the data we work with is conversational. Our most popular integrations are to community platforms and social networks. Both varieties of community platform—chat and forum—are conversational in nature. Even GitHub is conversational - conversations happen on issues, pull requests, pull request reviews, and of course GitHub Discussions. Not all community data is conversational, but most of it is.
As community builders, we know how important conversations are. They are the conduit through which connections are made, information is shared, and work is done. Back in April, I started to explore ways that Orbit could bring more insights to users by looking at the conversation level vs. single activities. At the same time, I started looking at how Orbit could help our users leverage AI and the rapid advancement of LLMs in the context of community building. Since many LLMs are optimized for conversations with humans, I thought they might also be very good at analyzing conversations between humans.
Together, these two investigations led to Panoramica. As I got underway with "conversationalizing" Orbit's data, I saw first-hand how tricky it could be. Orbit has integrations to dozens of conversational platforms and each one works a different way. Some platforms have threads, others have multiple levels of replies. Some platforms have channels, others have categories, others have more unique conversational boundaries, such as single issues or PRs on GitHub. Email is its own animal. There is no standard to work off of.
Creating a conversation graph
Making it harder, when conversationalizing the data I wanted to account for a variety of things that I felt were important to understanding the whole conversation. These include:
- The full Parent/Reply Tree
- Mentions of people in messages
- Entities used in messages (hashtags, keywords, extractable topics, etc.)
- Reactions to messages in the thread
- Actors who wrote the messages
One of the first choices I needed to make was about how to store all of this information. I reached for a graph database because of how relationship-oriented the data is, and also how common certain kinds of traversals are for analyzing the data. For example, to find all descendants of an original post or all members or entities anywhere in a thread. Or who is connected to who, which is very useful in the community context. It turns out that conversations get much easier to work with when you have a graph data model. I chose to use Memgraph as the graph database engine since it is open-source, fast, and developer-friendly.
The graph query language I learned, called CYPHER, is very short and expressive. Here's a portion of a query from Panoramica that finds all descendants of each activity in a project and yields a list of nodes that then can be looped over.
MATCH (p:Project)-[:OWNS]->(a:Activity)
WITH DISTINCT a
MATCH path = (a)<-[:REPLIES_TO*0..]-()
WITH a, path
UNWIND nodes(path) as node
This makes it really easy to get all members, entities, mentions, replies, etc. from across an entire thread. And the performance is very fast, even over large datasets.
Adding AI to the mix
Of course, staring at a screenful of nodes and edges isn't for everyone, so I started to look at how AI could allow end users to get these insights without actually knowing or caring that there's a graph involved. At this point, one of Orbit's most senior engineers, Steeve Bete, got involved in the project.
Steeve had been doing some AI explorations of his own and had ideas about how to engineer prompts that would allow end users to ask simple questions and still get the right answers. Questions like:
- Who in our community knows the most about Python?
- What are the most popular websites that our members share?
- What was the most common problem reported in our support channel this month?
- How many members have had a conversation with Paula? What about?
- Who is our most influential community member?
For the last month, Steeve and I have been working to bring the first version of Panoramica to you. It is certainly not complete, and not yet capable of answering every question you might have regarding your universe of conversations. But we think it's a decent start and that it could already help developers start to build AI experiences with conversational data.
