AI LLM Integration

Workflow

1. Identify the user goal, model provider, latency budget, cost budget, privacy constraints, and failure tolerance.
2. Identify whether the work is prompting, retrieval, tool calling, evaluation, or integration plumbing.
3. Route to a narrower skill when one applies.
4. Prefer existing SDKs, clients, tracing, retry, and secret-management patterns.
5. Design deterministic boundaries around nondeterministic model calls.

Implementation Guidance

• Prompts: use prompt-engineering.
• Retrieval and embeddings: use rag-pipeline.
• Agents and tool calls: use agent-tool-calling.
• Evals and model comparison: use llm-evaluation.

Verification