Enhancing Developer Support at Adobe: The Unified Support Initiative
Overview of Adobe’s Creative Ecosystem
Introduction to Unified Support
Solution Architecture and Objectives
Improving Document Retrieval and Deployment
Multi-Tenancy Through Metadata Filtering
Utilizing the Retrieve API
Experimental Insights and Metrics
Conclusion: Building a Scalable Future for Developer Support
About the Authors
Enhancing Developer Productivity at Adobe with Unified Support
Adobe Inc. has long been a beacon of creativity and innovation, offering a comprehensive suite of tools that empower artists, designers, and developers in various digital disciplines. From web design and photo editing to vector graphics and video production, Adobe’s product landscape forms the backbone of countless creative projects globally. However, behind its powerful tools lies an equally potent support engine—Adobe’s commitment to enhancing developer productivity.
The Challenge
As internal developers at Adobe utilize extensive documentation, including wiki pages, software guidelines, and troubleshooting guides, the challenge of swiftly locating pertinent information became evident. The need for an efficient system that consolidates crucial resources was paramount. This realization spurred Adobe’s Developer Platform team to innovate with a centralized support system known as Unified Support.
Unified Support was designed to streamline the process for thousands of internal developers, providing immediate answers to a plethora of questions—from troubleshooting software upgrades to setting up continuous integration and delivery (CI/CD) pipelines in new AWS Regions. By centrally organizing information, Adobe aimed to minimize both time and costs associated with developer support.
Building the Solution
The endeavor began with a prototype that unearthed significant insights about the necessary features and capabilities required for a system operating at Adobe’s scale. Key focus areas included scalability, simplifying resource onboarding, content synchronization, and optimizing infrastructure efficiency. The next logical step was enhancing retrieval precision to ensure developers could quickly find the information they needed.
To tackle these challenges, Adobe partnered with the AWS Generative AI Innovation Center, leveraging technologies such as Amazon Bedrock Knowledge Bases and Amazon OpenSearch Serverless. This collaboration resulted in a notable 20% increase in retrieval accuracy, significantly enriching the developer experience. By implementing metadata filtering, developers could fine-tune their searches, helping them surface relevant answers quickly across complex, multi-domain knowledge sources.
Technical Overview
The project encompassed two primary objectives:
-
Document Retrieval Engine Enhancement:
- Creation of a robust system to improve search result accuracy.
- Development of a data ingestion pipeline for preprocessing and indexing in a vector database, with performance evaluations against Adobe’s ground truth data.
- Scalable, Automated Deployment:
- Designing a reusable blueprint to accommodate large-scale data ingestion and flexible configurations, including embedding model selection and chunk size adjustment.
The architecture’s backbone, Amazon Bedrock Knowledge Bases, operates through several stages:
- Data Ingestion: Pulling data from sources like Amazon S3 buckets.
- Chunking: Breaking data into manageable pieces.
- Vectorization: Transforming chunks into numerical vectors using embedding models.
- Storage: Storing vectors in Amazon OpenSearch Serverless for efficient searching.
When a developer submits a query, the system:
- Performs query vectorization.
- Conducts a similarity search and retrieval based on similarity scores.
- Ranks and presents the most relevant results.
Multi-Tenancy through Metadata Filtering
On the developer’s journey of problem-solving, challenges often span multiple domains. Metadata filtering was introduced to empower developers to retrieve well-defined subsets of information based on specific criteria, enhancing the relevancy of search outputs. Each source data file is associated with a metadata file, facilitating multi-tenancy and precise filtering across Adobe’s expansive knowledge landscape.
Experimentation and Results
To ensure the highest accuracy and efficiency, rigorous testing was conducted to fine-tune the system. Metrics for evaluating success included document relevance—assessing how effectively the retrieved information addressed developer questions—and Mean Reciprocal Rank (MRR), which evaluated the ranking of the first relevant item for queries.
Through a variety of data chunking strategies—including fixed-size chunking and semantic chunking—the team was able to determine that the 400-token fixed-size chunking method yielded the highest accuracy. This iterative approach enabled optimization tailored to Adobe’s specific needs, striking the perfect balance between effectiveness and efficiency.
Conclusion
Adobe’s journey to develop the Unified Support search and retrieval system has successfully harnessed the capabilities of Amazon Bedrock Knowledge Bases to establish a scalable and efficient developer support framework. With a 20% increase in retrieval accuracy and the implementation of metadata filtering, Adobe can now offer a seamless navigation experience amidst the complexities of its information landscape.
As technology continues to evolve, this fully managed solution lays a robust foundation for ongoing enhancements in developer support at Adobe, showcasing the company’s commitment to empowering its teams with the tools they need to excel.
For those inspired to collaborate on similar projects with AWS, the Generative AI Innovation Center provides a wealth of resources. To explore Amazon Bedrock Knowledge Bases further, visit the AWS documentation to learn how to retrieve data and generate AI responses effectively.
About the Authors
The collaborative efforts of a diverse team of experts, including data scientists, engineering managers, and software developers, underscore the innovation driving the Unified Support initiative at Adobe. With extensive experience in AI, cloud technologies, and developer support, the authors share a common passion for enhancing productivity and user experience in the ever-evolving digital landscape.