Production AI Architecture

Key findings

• 01Finding 1 — retrieval. Naive RAG fails in production for five predictable, documented reasons. Hybrid retrieval (BM25 + dense, fused with reciprocal rank fusion) plus cross-encoder reranking is the single highest-ROI architectural change available to most enterprise pipelines.
• 02Finding 2 — agentic patterns. Agentic systems sit on a five-tier maturity ladder. Tiers 1–2 are broadly production-ready; Tiers 3–5 are workload-specific and frequently overclaimed in current practitioner discourse.
• 03Finding 3 — anti-patterns. Six anti-patterns recur across documented production retrospectives — the evaluation cliff, debugging opacity, latency drift, integration impedance, the POC-to-production gap, and the verification vacuum — each with a diagnostic signature and a mitigation.
• 04Finding 4 — decision framework. Pattern selection can be made almost deterministically given five workload characteristics: corpus characteristics, query complexity, latency budget, error tolerance, and verifiability surface.
• 05Series link. These findings give the operating model (Pillar 3) its technical object, the cost analysis (Pillar 4) its cost drivers, and the governance practice (Pillar 5) its evaluation surface. The architecture is necessary but not sufficient — it reaches production only through the organisational conditions the later pillars examine.