the new-information contribution of an utterance; the part that moves a discourse forward
A Python ecosystem for generating inspectable multi-party conversations from profiles, scenarios, domain constraints, and discourse-state operations.
structured inputs recombined into new conversations; generated, not scripted
Rhemix is a Python ecosystem for building synthetic dialogue datasets from domain-specific ingredients. A domain pack defines the world: participant roles, prompts, tools, constraints, and quality checks. Profile generation creates concrete participants for that world. The engine then runs multi-party conversations through explicit discourse-state operations, producing transcripts, state snapshots, tool traces, and generation metadata.
Rhemix is built for evaluation engineers, data scientists, and researchers who need controlled dialogue data rather than scraped, hand-authored, or single-pass conversations: conversational AI evaluation, redaction and compliance testing, dialogue-system regression tests, and research workflows that need structured conversational state and mental state modeling alongside plain text.
The open framework separates conversation generation from domain content: rhemix-engine provides the discourse runtime, dashboard, and batch generator; rhemix-profiles provides profile-generation infrastructure; domain packs provide the concrete roles, prompts, data, tools, and evaluation rules for a vertical. The first domain pack (private license) targets customer-service dialogue with synthetic profiles, PII, transfers, channel artifacts, and quality checks.
The name is a portmanteau of rheme and remix. A rheme is the new-information contribution an utterance makes to a discourse. Remix points to the system’s method: participant profiles, scenarios, domain constraints, and discourse operations are combined to generate new conversations with inspectable structure, rather than replaying templates or reusing source dialogues.