Production: Batch, Optimization, and Observability
Move to production: batch inference up to 1M requests, predicted outputs, citations with sources, and observability (traces, judges, offline evaluations).
16 lessons About 1 h 15 min Advanced
What you'll learn
- Prepare a batch inference job in JSONL, track its status, and retrieve the results
- Reduce response latency with predicted outputs on rewriting tasks
- Implement citations with reference object to ground responses in their sources
- Trace an application with OpenTelemetry, explore production traces, and automate evaluation with LLM judges
- Build datasets from traces and measure each version with offline evaluations
Course details
- Who it's for
- Developers, engineers, and integrators who want to use Mistral AI via API in their applications.
- Prerequisites
- Proven experience developing applications and deploying them to production.
- Length
- 1 h 15 min · 16 lessons
- Format
- 100% online, self-paced course, accessible 24/7 from a browser. Progress is saved, and you can resume at any time. Estimated duration: 1 h 15 min.
- Assessment
- Self-assessment based on the Key takeaways that conclude each lesson, and direct practice of the examples on your own examples. A final quiz of 5 graded questions at the end of the last lesson validates what you have learned.
- Accessibility
- Course accessible to people with disabilities: text content compatible with screen readers, keyboard navigation, and compliant color contrast. For any adaptation needs, email [email protected].
- Price
- Free
- Recognition
- Certificate of completion issued by Corsen AI, verifiable online.
- Syllabus
- Download the syllabus (PDF)
Syllabus
Batch Inference
Predicted Outputs
Citations and References
Observability
- 11 Introduction to Observability: Traces, Judges, and Evaluations
- 12 Explore: Search and inspect production traces
- 13 Judges: Automated LLM Evaluation
- 14 Offline evaluations: measure before you deploy
- 15 Datasets: Build evaluation sets and iterate
- 16 End-to-end pipeline: traces, judges, datasets, evaluations