{"type":"analysis","id":"analysis:research-dropbox-used-dspy-and-gepa-to-cut-incomplete-dash-chat-answers-by-26-percent","slug":"research-dropbox-used-dspy-and-gepa-to-cut-incomplete-dash-chat-answers-by-26-percent","title":"Dropbox used DSPy and GEPA to cut incomplete Dash chat answers by 26 percent","description":"A 25 June 2026 Dropbox engineering post describes a two-stage DSPy loop: calibrate LLM judges on human-labeled internal chats, then run GEPA on offline replay to rewrite the Dash chat system prompt. First-party launch metrics show fewer incomplete answers and lower token use. Independent outlets have not audited those figures.","published_at":"2026-11-08T06:00:00.000Z","updated_at":"2026-11-08T06:00:00.000Z","record_date":"2026-11-08","topics":["prompts","reasoning","research"],"source_urls":["https://dropbox.tech/machine-learning/how-we-turned-ai-evaluations-into-better-responses-in-dash-chat","https://dropbox.tech/machine-learning","https://dropbox.tech/machine-learning/llm-human-labeling-improving-search-relevance-dropbox-dash","https://dropbox.tech/machine-learning/vp-josh-clemm-knowledge-graphs-mcp-and-dspy-dash","https://dropbox.tech/machine-learning/optimizing-dropbox-dash-relevance-judge-with-dspy","https://dspy.ai/current/api/optimizers/MIPROv2/","https://dspy.ai/current/api/optimizers/GEPA/overview/","https://arxiv.org/abs/2507.19457","https://dspy.ai/community/use-cases/","https://www.infoq.com/news/2026/02/dropbox-context-engine/","https://dropbox.tech/machine-learning/practical-blueprint-evaluating-conversational-ai-at-scale-dash"],"language":"en","routes":{"html":"https://newruntime.com/analysis/research-dropbox-used-dspy-and-gepa-to-cut-incomplete-dash-chat-answers-by-26-percent/","markdown":"https://newruntime.com/analysis/research-dropbox-used-dspy-and-gepa-to-cut-incomplete-dash-chat-answers-by-26-percent.md","json":"https://newruntime.com/analysis/research-dropbox-used-dspy-and-gepa-to-cut-incomplete-dash-chat-answers-by-26-percent.json"},"author":{"role":"author","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]},"editorial_provenance":{"desk":"authorial","target_ref":"deep_research:7c207692-b11a-4a80-9819-fb4affc872d9","authorial_article_id":"7c207692-b11a-4a80-9819-fb4affc872d9"},"next_reads":[{"type":"topic","path":"/topics/prompts/","url":"https://newruntime.com/topics/prompts/","title":"Prompts","reason":"Explore the prompts topic hub.","media_type":"text/html","markdown_url":"https://newruntime.com/topics/prompts.md","json_url":"https://newruntime.com/topics/prompts.json"},{"type":"collection","path":"/analysis/","url":"https://newruntime.com/analysis/","title":"Analysis","reason":"Open the latest authorial analysis.","media_type":"text/html","json_url":"https://newruntime.com/analysis.json"},{"type":"collection","path":"/sources/","url":"https://newruntime.com/sources/","title":"Sources","reason":"See the original sources this archive cites.","media_type":"text/html"}]}