nacrecontextlayer/nacre-embedding-adapter repository overview

⁠Nacre — embedding adapter

One protocol in, several vendors out. POST /embeddings takes exactly what the Nacre⁠ worker and search path already send — {"model": …, "input": […]} — and answers exactly what they already read; POST /rerank speaks Text Embeddings Inference's shape for the same reason. It exists so the core does not learn three protocols: anything OpenAI-shaped already works by pointing an embedding_providers row's endpoint at it, and this adds the vendors that do not, with a credential, without either reaching Postgres.

Part of the open core — github.com/nacre-work/nacre⁠, Apache 2.0. Standard library only.

⁠Nothing has a default

• No route exists unless NACRE_EMBED_ROUTES names it.
• An unrouted model is refused by name. There is no fallback vendor, because a fallback is a decision made on the operator's behalf about where their documents go.
• With no routes at all the process refuses to start.

Routing a model here means the text of your documents leaves your installation. That is the trade, and it is stated in those words in docs/config.md⁠. The adapter is absent from the airgapped Compose profile rather than disabled in it.

⁠Configuration

_API_KEY and _API_KEY_FILE together is refused rather than resolved by precedence. The rerank credential is separate from the embedding one even where the vendor is the same. The full table is in the adapter's README⁠.

Tags: {version} and latest, amd64 and arm64, always the same version as nacre. ghcr.io/nacre-work/nacre-embedding-adapter is the canonical address; this repository is a mirror pushed by the same build at the same tags.

⁠Links

nacre.work⁠ · Configuration⁠ · Releases⁠