Pattern library
Eight reusable decisions for model access, retrieval, agent coordination, operations, and governance. Start with the problem your system needs to survive.
01 / Platform control
Place one governed access layer between applications and model providers for routing, authentication, budgets, policy, and auditability.
Architecture decision
Use it when several teams, models, or providers need one enforceable operating boundary.
02 / Grounded generation
Separate ingestion, indexing, retrieval, answer generation, and citation checks so each stage can be evaluated and repaired.
Architecture decision
Use it when answers must be grounded in private or frequently changing source material.
03 / Agent systems
Assign narrow roles, explicit tool contracts, shared state, and recovery paths instead of relying on one unconstrained agent loop.
Architecture decision
Use it when work has genuinely different responsibilities, permissions, or verification steps.
04 / Tool connectivity
Expose tools and context through small, typed interfaces with least-privilege access, validation, and observable execution.
Architecture decision
Use it when several assistants or agents need safe access to the same systems and data.
05 / Model operations
Version prompts and datasets, evaluate behavior, control releases, and monitor quality, latency, and cost as one operating loop.
Architecture decision
Use it when an LLM feature has moved beyond a prototype and changes need controlled promotion.
06 / Operations
Trace model calls, tool use, retrieval, latency, cost, and evaluation results across the complete request path.
Architecture decision
Use it before production incidents force you to reconstruct agent behavior from partial logs.
07 / Risk control
Combine identity, data boundaries, approval gates, policy checks, and evidence retention around model and tool access.
Architecture decision
Use it whenever AI touches sensitive data, consequential actions, or regulated processes.
08 / Retrieval infrastructure
Choose storage by retrieval quality, filtering, tenancy, operations, and total system cost—not benchmark speed alone.
Architecture decision
Use it before committing a retrieval architecture to a database that cannot meet production constraints.
Implementation source