Build · RAG · Model fit

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.