NLP Cloud (direct)

Direct NLP Cloud API calls (nlpcloud SDK or raw HTTP, no framework). NLP Cloud is a multi-provider aggregator: one API surfaces a catalog of open-source and proprietary text-generation/embedding models behind a single key, at https://api.nlpcloud.io/v1/<model>/<endpoint> (well- documented publicly; NLP Cloud's own docs pages returned 403 to programmatic fetch during this session, so the base URL and per-model path shape could not be independently re-verified live here — treat the JSONPaths below as the standard gen_ai.* convention, not a vendor-confirmed field layout). Unlike Voyage/Jina/Mixedbread, NLP Cloud fronts chat/completion-style models as well as embeddings, so an output_tokens metric is meaningful here and is kept. IMPORTANT for chargeback: because NLP Cloud is an aggregator, the OTLP spans this connector ingests only carry NLP Cloud's own request/response — whichever underlying model actually served the call, cost and usage land on NLP Cloud as the billing entity, not on the underlying model vendor. Per- underlying-model attribution is only possible if the customer's own instrumentation additionally captures the model id NLP Cloud reports back (gen_ai.response.model), which this connector does try via model_fallback_paths. Auth is NOT bearer-prefixed: NLP Cloud expects `Authorization: Token <api_key>` (not `Bearer <api_key>`), reflected below via header_prefix. NLP Cloud has no admin/billing API of its own beyond the dashboard, so spend is only capturable via this customer-side OTLP push. No telara-observability provider wrapper exists for NLP Cloud today (the package's optional-dependencies only cover langchain/langgraph/ crewai/autogen/llamaindex/openai-agents/openai/anthropic/azure-openai) — spans must come from the customer's own OTel instrumentation around the NLP Cloud SDK/HTTP call. Spans should be tagged telara.span.kind=provider so the cost rollup excludes them when an active framework adapter already covers the same call (double-count guard).