Imp Ports DSPy for Declarative LLM Programming to Elixir's BEAM

Introduction to Imp

Imp is a new framework that brings the functionalities of DSPy to the BEAM ecosystem, specifically for Elixir developers. It allows for the creation of declarative and self-improving language model programs. This port aims to combine DSPy's approach to LLM programming with the reliability and concurrency features of Elixir's OTP.

Declarative LLM Programming

With Imp, developers describe the inputs and outputs of each language model step and define how the model should process information. The framework then uses an optimizer to improve the model's performance based on provided examples. This approach eliminates the need for manual prompt writing or parser development, as Imp constructs prompts from signatures and validates replies against them, returning typed fields.

Integration with BEAM and OTP

A key aspect of Imp is its integration with the BEAM runtime. In this setup, an agent functions as a process, maintaining its own state, receiving messages, and operating under a supervisor alongside other application components. This allows for building applications ranging from single typed calls to multiple long-running agents, with each part measurable and optimizable.

Optimization and Evaluation

Imp includes mechanisms for optimizing and evaluating language model programs. Developers provide labeled examples across training, validation, and test sets. An optimizer, such as GEPA, learns from failures and rewrites instructions to improve performance. Metrics are used to score the program's effectiveness before and after optimization.