rauta-market

List Gpus

List GPU types currently available across providers, with per-provider price, region, and availability. Use when the user names a GPU (e.g. 'A100', 'H100', 'L40S') or wants to browse the market. Data is real benchmarker-sourced market listings, not estimates.

Lookup Model

Look up a Hugging Face model: parameter count, context window, VRAM at fp16/int8/int4, which market GPUs fit it, and (when indexed) its architecture facts incl. serve_needs_gpu / train_needs_gpu. Curated catalog first, then the HuggingFace Hub LIVE; `source` says which. A live answer distinguishes not_found from gated and upstream-unavailable, so a rate limit never reads as a missing model.

Recommend Workload

Best provider+machine for a workload, GPU or CPU: the canonical 'what should I use for X' tool. GPU: model_id derives the VRAM floor, then live listings ranked by price. CPU (compute_class='cpu' or a CPU workload): sizes RAM+vCPU, ranks the CPU market with monthly cost incl. egress, arch warnings, per-row confidence and the API break-even. Encoder/embedding/tabular models take the CPU path, never a GPU.

Leaderboard

Inference leaderboard for one category: 'cloud' (GPU rental providers), 'hardware' (GPU types) or 'inference' (serverless APIs). Use for 'who's fastest/cheapest/best' with no model named. Returns {category, count, rows, notes}; count 0 means not populated yet and notes point at the live tools that have data: report that, never invent numbers.

Cold Start Stats

Real provisioning (cold-start) latency AND supply health per provider+GPU: p50/p95/min provision time, success/error counts, failure rate, and the actionable cause per failing bucket. Use for cold start, spin-up, reliability, 'who can we actually get GPUs from now'. Benchmarker probes only. `dominant_cause`/`blame` per bucket: owner_billing (a human must pay), owner_account, rauta_credential (rotate our key), rauta_adapter (our bug), provider_capacity (no hardware). A 0% provider is signal: surf

Compare Providers

Cross-provider time series of TTFT p50/p95, latency, tokens/sec and availability, head-to-head over time. Pass provider ids (e.g. ['runpod','lambda_labs']); default: providers in the live market. Empty series mean no real samples yet.

Ask Advisor

Ask the Rauta Advisor for one opinionated prose recommendation grounded in the same live data. Only when the user wants Rauta's own summarized answer rather than reasoning over the raw tools. Routes to a self-hosted model that may cold-start (~2 min); prefer the deterministic tools for speed.

Hardware Catalog

Hardware lookup: purchase prices for AI accelerators (default), or compute='cpu' for the rentable CPU host catalogue (vCPU, RAM, arch, monthly/hourly price, egress terms, confidence + source URL per row). Always cite confidence; never present partial/inferred prices as fact.

Procurement Channels

List buy-side procurement entities (OEMs, distributors, VARs, brokers, colo, etc.) with how_to_buy and buy_url. Never invent checkout links.

Buy Vs Reserve

Recommend buy (OEM) vs reserve (neocloud) given utilization, horizon, GPU class, and datacenter readiness. Returns next_step buy_url links.