Architecture
Every default is a decision, not a shrug
Six components, six explicit trade-offs. This is what we tell clients when they ask "why this and not that" — no "let the team decide later."
01
Ingestion Pipeline
Trade-off
Step Functions over plain Lambda chains buys per-step retry and observability at the cost of added infra complexity. It breaks down at Glue-scale document volume — that's a batch-ETL job, not a state machine.
02
Vector Store
Trade-off
OpenSearch Serverless carries a real ~$200/mo OCU floor regardless of query volume. For light workloads, pgvector is the honest recommendation — not the default.
03
Agent Runtime
Trade-off
This is decided per engagement up front. Deferring it means re-architecting the memory and guardrail layer later — there is no neutral middle ground.
04
Multi-Agent Topology
Trade-off
Supervisor costs +1 hop of latency but stays debuggable and auditable. Swarm is faster to prototype but can't answer "why did the system do that" — a dealbreaker for an auditor.
05
Evaluation Pipeline
Trade-off
Failures are triaged by category before anyone touches a prompt: retrieval miss → fix chunking/embedding, correct retrieval + wrong answer → fix prompt/model, stale answer → fix the sync pipeline.
06
Security & Access Control
Trade-off
Filtering happens at retrieval time, never after generation — retrieval-time filtering is the only defensible design in a regulated environment.