stop sequences

**Stop sequences** is the **configured text patterns that cause generation to stop when matched in the decoded output stream** - they are widely used to enforce response boundaries at application level. **What Is Stop sequences?** - **Definition**: String-level termination triggers checked during incremental decoding. - **Matching Behavior**: Generation halts when output suffix matches any configured sequence. - **Use Cases**: Template completion, tool protocol boundaries, and multi-message formatting. - **Difference**: Operates on decoded text rather than raw token IDs. **Why Stop sequences Matters** - **Protocol Control**: Prevents model from writing beyond expected sections. - **Integration Safety**: Essential when model output is consumed by parsers or downstream tools. - **UX Consistency**: Keeps response endings aligned with interface constraints. - **Cost Savings**: Stops output as soon as required content is complete. - **Operational Flexibility**: Easy to update without retraining or model changes. **How It Is Used in Practice** - **Sequence Design**: Choose unambiguous markers unlikely to appear in normal content. - **Tokenizer Testing**: Validate boundary detection across tokenization edge cases. - **Escaping Strategy**: Handle quoted and escaped delimiters in structured outputs. Stop sequences is **a practical high-level termination mechanism for production apps** - careful sequence design prevents accidental truncation and parsing failures.