Reference
API reference
OpenAI-compatible. Endpoints, schemas, error codes, rate limits.
§ 01
Base URL and authentication
All requests go to https://api.grunden.ai. Send your key in the Authorization header:
§ 02
Endpoints
Alla endpoints kräver Authorization: Bearer sk-grunden-… och accepterar/returnerar JSON. Klicka på en kapacitet för fullständig anropssyntax och kodexempel.
| Endpoint | Beskrivning | Kapacitet |
|---|---|---|
| POST /v1/chat/completions | Text-in / text-ut, OpenAI-kompatibel. Streaming + tool-calling stöds. | Chat (GLM 5.2) → |
| POST /v1/embeddings | 1024-dim multilingual embeddings (BGE-M3). OpenAI-kompatibel. | Embeddings → |
| POST /v1/rerank | Cross-encoder. Skicka query + lista texter, få relevans-scores i retur. | Rerank → |
| POST /v1/pii/redact | Detekterar namn, personnummer, e-post, telefon, adress. Returnerar maskerad version. | PII-detektering → |
| POST /v1/messages | Anthropic-kompatibel (wire-format). Servar glm-5.2 i Anthropics request/response-form — kör INTE Claude. Låter Claude Code & Anthropic-SDK peka hit via ANTHROPIC_BASE_URL. Auth även via x-api-key. Streaming + tool-use stöds. | OpenAPI-spec → |
| GET /v1/models | Returnerar alla aktiva modeller med typ + endpoint. Kräver auth. | Modeller & priser → |
| POST /v1/knowledge-bases | Kunskapsbas-API (RAG): skapa/lista/radera KB, ladda upp dokument (PDF/DOCX/TXT m.fl.) via /{id}/documents, text via /{id}/snippets, länka nyckel via /{id}/keys. | OpenAPI-spec → |
| POST /v1/chat/completions (knowledge_base_id) | Fråga en kunskapsbas: lägg knowledge_base_id i chat-anropet → relevanta utdrag injiceras automatiskt som kontext (OpenAI-kompatibel RAG). | OpenAPI-spec → |
§ 03
Error codes
- 401 Unauthorized — missing or invalid Authorization header.
- 402 Payment Required — wallet balance is 0 or negative. Top up at /t/<orgId>/billing.
- 413 Payload Too Large — uploaded file exceeds the limit.
- 429 Too Many Requests — rate limit at api-key level. Default 60 req/min, contact us for increases.
- 400 Bad Request — missing required header or invalid query param.
§ 04
Vision (multimodal input)
Vision/multimodal input (image → model) is not active in the current deployment. GLM 5.2-FP8 (MoE 700 GB VRAM) uses all 8 H200 cards, so we cannot run a parallel vision model yet. When we scale up the hardware it will be published via GET /v1/models and documented here.