Performance Tuning
Reduce latency and cost through model selection, caching strategy, database optimization, and workflow runtime tuning
Reduce latency and operating costs without sacrificing response quality through systematic tuning.
Model Selection
Choose from models authorized for your team based on actual needs — don't assume fixed speed or quality rankings. Measure latency, answer quality, context capacity, and cost on representative prompts before deciding on production configuration.
Caching Strategy
- Only cache data with a clear invalidation policy
- JWT is the primary session mechanism; Redis supports token blacklist, optional single-session state, plus Celery brokering, rate limits, and temporary caches
- Do not add a global LLM response cache unless prompt, model, team permissions, tools, and freshness requirements make reuse safe
Database Optimization
- Add indexes on fields used by frequent filters and joins
- Configure connection pooling appropriate to the deployment
- Measure query plans and retrieval latency before changing schema or cache settings
Workflow Runtime
- Bound every loop. The editor defaults to 10 iterations and the executor caps loops at 1000
- Use an Iteration node for array/batch processing with explicit, verifiable termination conditions
Key Limits
| Parameter | Value | Description |
|---|---|---|
| Editor default loop count | 10 | Default iteration limit in the visual editor |
| Executor max loop count | 1000 | Runtime hard cap — exceeded loops are force-terminated |
How is this guide?