LLM Application Patterns

LLM applications need normal software engineering controls plus specific controls for prompts, retrieval, evaluation, cost, privacy, and model behavior. Treat the model as a dependency with probabilistic output, not as deterministic business logic.

Common patterns

• Chat or assistant interfaces.
• Retrieval-augmented generation.
• Document summarization and extraction.
• Classification and routing.
• Code, query, or configuration generation.
• Agentic workflows with tool access.

Architecture baseline

Required controls

Every production LLM application should define:

• Data allowed in prompts and responses.
• Prompt and model versioning.
• Retrieval source ownership and freshness.
• Evaluation suite for important behaviors.
• Human review requirements for high-impact actions.
• Cost limits and usage monitoring.
• Abuse, prompt injection, and data exfiltration controls.
• Logging policy that avoids sensitive data leakage.

Tool use

When models can call tools, define:

• Which tools are available.
• Authorization model for each tool.
• Input validation and output filtering.
• Dry-run support for risky actions.
• Approval steps for irreversible changes.
• Kill switch or disablement path.

Watchouts

• Prompt changes are production changes.
• Retrieval quality determines answer quality.
• Model confidence is not evidence.
• Sensitive data can leak through logs, traces, prompts, and eval sets.
• Agentic workflows need strict permissions and observability.