proscenium
Proscenium
Proscenium is a small, experimental library of composable glue that allows for succinct construction of enterprise AI applications. It was started in February 2025 and is still in early development.
Currently, proscenium development prioritizes support for domains where the creation and use of structured data is critical.
See the repository on GitHub.
Demos / Quickstart
The Proscenium repository includes a set of simple applications that demonstrate key features of the library.
There are two ways to get started quickly:
-
Command Line Interface to demos, ither from a local repo clone or in a new GitHub Codespace
Resources
For more background and future plans, see the wiki
To find the Proscenium community, see the discussions
Motivation
Enterprise AI applications share many challenges with other Generative AI applications, including the speed of evolution of model capabilities and the software ecosystem.
However, they also have qualities that warrant dedicated focus. Proscenium is a library of patterns that emerge when building enterprise AI applications and products for which accuracy and accountability are critical.
The use cases are “big picture” systems, not just product features. For example they include steps such as:
- Produce high-confidence, auditable results
- Securely process enterprise data
- Evaluate and report system quality and performance
- Handle multi-modal user queries and data
- Efficiently consume relevant data and reference material
- Integrate with appropriate tools
- Guide owners and subject-matter experts through system evolution options
The applications themselves are complete enough to provide insight into specific user interactions, in a range of settings such as integrations with chat systems or collaborative document editors.
Although the initial development of Proscenium targets python notebooks (running locally or on Colab) and command-line tools, Proscenium ultimately must be easily deployable to enterprise-friendly platforms. The first two target platforms are AWS and Llama Stack.
Finally, Proscenium is being developed with an understanding that details from specific enterprise application domains should be prioritized in the early phases of its design.
Library Design
The AI model, software, hardware, and cloud ecosystem is evolving rapidly. Proscenium is designed to co-evolve with the ecosystem by keeping interfaces simple and loosely-coupled.
It’s simplest offerings are low-level “verbs” such as: chunk, complete (inference), extract, and invoke (tools). In many cases these are thin wrappers around well-known libraries.
Those verbs are composed to form higher-level generic systems including:
- Document Enrichment
- Knowledge Graph Construction
- Entity Resolution
- RAG
- Graph RAG
- Tool Use
Demonstrations in the Legal Domain
Legal use cases rely heavily on deriving structured data from unstructured data, so this is an ideal domain to demonstrate several related subsystems.
Federal Caselaw and Statute Research
The first use case in the legal domain is question-answering related to large, public-domain legal datasets including U.S. federal case law and statutes.
The questions chosen highlight categories of questions where the ability to traverse a Knowledge Graph provides advantages over a naive RAG approach.
Email Generation with Context from Matter Management System
The second use case in development is a email authoring capability that integrates with a Matter Management system.
Common Features and Requirements
Demo applications will be demonstrated as Slack integrations as well as Python notebooks.
They will use Llama 4 models hosted by Together.AI, and make use of a Neo4j graph database.
These applications will also demonstrate the creation and tuning of custom benchmarks. Existing legal benchmarks today are very narrow and/or do not map well to customer value.