Laurian/context-compression-experiments-2508

Implements a practical prompt optimization pipeline for retrieval-augmented generation, using DSPy's GEPA (genetic algorithm) to evolve context-compression prompts that improve gpt-4o-mini performance on real production failures. The system extracts query-relevant document sections while preserving original formatting and structure, trained on 296 document-query pairs where gpt-4o succeeded but gpt-4o-mini initially failed, achieving measurable improvements in extraction accuracy across iterations.

No commits in the last 6 months.