So many things can impact a marketing campaign—from economic shifts to weather events and major news. So, to help marketers understand how the external world is affecting their campaigns, Adobe Research and the Adobe Customer Journey Analytics (CJA) team collaborated to create Project Perfect Context, a prototype showcased as a Sneak at Adobe Summit. The tool allows users to choose from a wide range of real-world insights, and then an AI agent connects those insights with a team’s own data to help make sense of customer behavior in context.
In March, the marketing world got an early glimpse of the experimental Project Perfect Context in a sneak peek on the Summit stage in Las Vegas.
Project Perfect Context grew out of early work on LLM-based AI assistants
Back in August 2023, Senior Research Scientist Sungchul Kim was using LLMs (large language models) to create AI assistants that could help people analyze their data. “The goal was to provide a chat interface so that CJA users, such as data scientists and analysts, could easily generate a chart or report by simply asking a question about their data. They could say something like, ‘How does our revenue compare by country?’ and CJA would produce a chart that could answer the question,” explains Kim. “This was the very beginning of using LLMs for agentic applications at Adobe.”
With that work under way, the CJA product team came to Kim with a new idea—one that would eventually become the Sneak Project Perfect Context. They wanted the CJA AI assistant to be able to leverage external data sources to answer even more complex questions. So Kim’s small team began building a model that brings information like stock market trends, social media sentiment, and even global weather events into the analysis.
To do it, the team needed to create new ways to understand users’ needs. For example, the model helps fill in the gaps when users ask vague questions. “Our advanced technology can help clarify uncertainty—because users often come up with very unclear queries. We needed to be able to clarify those queries by using the context we have. It’s a very challenging research problem, and we did some exciting work on it,” says Kim.
Two interns helped with the challenge of decoding vague queries, and their projects led to novel innovations, published papers, and some of the technology that powers Project Perfect Context. Yeonjun In’s work used AI to interpret short, uncertain queries and give them a clear intention so the agent can produce the best analysis. And Yu Xia’s research focused on using data sources to get enough context to help perform an analysis correctly.
After two years of close collaboration between the Adobe Research team, the CJA product team, and Adobe designers, the team had built Project Perfect Context to demonstrate how marketers can ask their AI assistant to tap into data they already have and to choose what kinds of external data to consider.
Yeonjun and Xia both plan to intern with Adobe Research in the summer of 2025 to continue developing their research.
From Research to Summit Sneak
When Project Perfect Context was selected as a Summit Sneak, the CJA product manager, designers, and engineers got to work building the demo and developing a script for the Summit stage, and Kim collaborated with the team to polish the use case.
“This is the first time my work has been part of a Sneak, and I’ve enjoyed it a lot,” Kim adds. “The whole collaboration has been great and we’re excited about the future of this technology.”
Project Perfect Context reached an even bigger audience through press coverage on ZDNET after Summit.
Key contributors
Presenter: Ben Gaines
Adobe Research collaborators: Yeonjun In (intern), Tu Xia (intern), and Sungchul Kim
Additional collaborators: Prithvi Bhutani, Brandon George, and Jordan Walker
Wondering what else is happening inside Adobe Research? Check out our latest news here.
