genai-services_skill

Overview

This skill helps implement and operate OCI Generative AI services with practical guidance on model selection, cost optimization, token management, rate-limit handling, and healthcare/compliance for PHI/PII. It focuses on safe production usage: redaction, validation, monitoring, and efficient cost trade-offs. Use it to reduce surprise costs, prevent data leaks, and stabilize GenAI integrations.

How this skill works

The skill inspects common failure modes and operational patterns for OCI GenAI: rate limit 429s, 400 token errors, model token budgets, and cost-per-token trade-offs. It provides concrete retry/backoff patterns, truncation and prompt-optimization techniques, PHI redaction strategies, response validation checks, and a decision approach for when to use embeddings versus generation. It does not guess live CLI syntax or pricing; use referenced commands and up-to-date API docs when performing CLI/API calls.

When to use it

• Designing GenAI workloads or choosing models for cost vs performance
• Troubleshooting rate limits (429) or token errors (400) in production
• Implementing PHI/PII redaction, audit logging, and HIPAA controls
• Optimizing recurring costs via caching, embeddings, and prompt trimming
• Validating outputs for hallucinations before updating critical systems

Best practices

• Never include direct PHI/PII in prompts; redact and store mappings securely
• Prefer embeddings+RAG or embeddings-only searches when possible to save generation costs
• Implement exponential backoff on 429s and explicit handling for 400 token errors
• Use command-r for high-volume short tasks and command-r-plus only for long-context needs
• Validate responses with structure and PII checks; flag for human review before committing

Example use cases

• Medical transcription pipeline: redact identifiers, call GenAI, validate structure, store AI-suggested text for clinician review
• Cost modeling: estimate tokens per note, apply cache-hit assumptions, and compare command-r vs command-r-plus costs
• Rate-limit resilience: implement retry with jitter and circuit-breaker for high-throughput endpoints
• RAG for long documents: embed chunks, retrieve top-K, and query a smaller model to reduce generation tokens
• Compliance checklist: BAA verification, audit logs, access controls, and retention policies