Build · RAG

Best for — On-prem or edge deployment, cost-critical volume, and well-scoped tasks you can fine-tune for.

Watch out — Limited raw reasoning; usually needs fine-tuning or tight retrieval to hit quality.

Best for — A narrow, high-volume task you can train for — often the best $/quality once it's dialed in.

Watch out — Upfront labeled data and a training/eval loop; brittle outside its trained domain.

Best for — Regulated or sovereign data, deep customization, and avoiding a per-token vendor bill at scale.

Watch out — Real GPU + MLOps burden — you own capacity, uptime, patching, and evals.

Best for — Mixed workloads — route easy questions to a cheap model and hard ones to a flagship, optimizing cost and quality together.

Watch out — You own the routing logic, evals, and fallbacks; a misroute quietly costs quality or money.

Best for — High-volume, latency-sensitive workloads that still need solid reasoning — the workhorse default.

Watch out — Lower ceiling on the genuinely hard queries; still hosted, so residency and lock-in remain.

Best for — Meeting data-residency rules without running infrastructure yourself.

Watch out — Smaller model menu, still a managed dependency, and partial lock-in.

Best for — The hardest reasoning, long-document synthesis, and agentic chains where quality is non-negotiable.

Watch out — Priciest per call, data leaves your boundary, and you inherit vendor lock-in and rate limits.

Best for — Inputs beyond text — scanned documents, screenshots, diagrams, or audio that retrieval needs to read.

Watch out — Pays a premium for vision/audio; text-only quality can trail a text-specialized peer at the same price.

Best for — Genuinely hard analysis, planning, math, or code where a slower, deeper answer is worth it.

Watch out — Slowest and priciest per answer — overkill and frustrating for simple lookups.