ADR-026: OpenRouter for Embed, Chat, and Rerank
Status: Accepted Date: 2026-06-02 Authors: paulhumanity (with grazianognoll)
Context
humanwork already routes chat completions through OpenRouter. RAGflow had its own bundled embedder and reranker; migrating off it forces upstream-AI vendor decisions.
Options
- Region-tied managed model platform, rejected due to region and access-management lock-in.
- Direct OpenAI plus direct Cohere, rejected because it requires two vendors, two keys, and two billing lines.
- OpenRouter.
Decision
Use OpenRouter for all three traffic classes: embed (/api/v1/embeddings), chat (/api/v1/chat/completions), and rerank (/api/v1/rerank). This gives humanwork a single key, single billing line, and vendor swap by config string.
Verified 2026-06-02
OpenRouter exposes 26 embedding models, 3 rerankers (Cohere Rerank v3.5, Rerank 4 Pro, Rerank 4 Fast), and 342 chat models.