@kybernesis/brain-embed-openai

The default EmbeddingProvider for Cortex — a portable OpenAI embedding provider for @kybernesis/brain-core's EmbeddingProvider seam. It produces vectors (text-embedding-3-small, 1536-dim); the vector store lives separately in @kybernesis/brain-storage-vec. See the 0.8.0 kickoff §3.5 for why the embedder sits in the inference plane, not the storage plane.

Usage

import { createOpenAIEmbedder } from '@kybernesis/brain-embed-openai';
import { setEmbeddingProvider } from '@kybernesis/brain-core';

setEmbeddingProvider(createOpenAIEmbedder());

Once wired, indexChunk populates vectors.db and semanticSearch returns real results. Without it, brain-core's vector path no-ops (the seam's default).

Key resolution (tenant-aware)

The provider resolves OPENAI_API_KEY per tenant:

1. <tenant.paths.home>/.env — where the agent's per-brain key is written;
2. process.env.OPENAI_API_KEY — daemon-global fallback.

One OpenAI client is cached per tenant.slug. A tenant with no key degrades to null (no embedding) rather than throwing — so a mixed fleet (some agents keyed, some not) behaves per-agent.

openai is an optional peer dependency

The host application provides the openai SDK (>=4). Tests inject a fake client via createOpenAIEmbedder({ clientFactory }), so the package builds and tests without it.

Constraint: 1536 dimensions

The vector store schema is built for EMBEDDING_DIM = 1536. A different embedding model must emit 1536-dim vectors, or EMBEDDING_DIM + the vectors.db schema move together and you re-index. Don't mix models within one populated vectors.db — distances across models aren't comparable.