@arivie/embeddings

Provider adapters for RAG embedding: thin factories over AI SDK v6 (@ai-sdk/openai, @ai-sdk/cohere, @ai-sdk/voyage) that return an EmbeddingProvider record (model, modelName, dimensions, costPerMillionTokens). Downstream code calls embedMany({ model: provider.model, values }) using the re-exported embed / embedMany from this package — no Arivie-owned embed wrapper.

import { embedMany, openAIEmbeddings } from "@arivie/embeddings";

const provider = openAIEmbeddings({ apiKey: process.env.OPENAI_API_KEY! });
const { embeddings, usage } = await embedMany({
  model: provider.model,
  values: ["chunk one", "chunk two"],
});
// cost rollup (e.g. buildIndex): usage.tokens * provider.costPerMillionTokens / 1_000_000

All three providers use createProvider({ apiKey }).embedding(modelId) (AI SDK v6 idiom). Cohere and Voyage accept arbitrary model strings at runtime; metadata dimensions defaults are caller overrides when the model variant differs from the package default.

buildIndex

Chunk a semantic layer with ParagraphChunker, embed via embedMany, and upsert into a Mastra vector store (typically PgVector):

import { PgVector } from "@mastra/pg";
import { buildIndex, openAIEmbeddings } from "@arivie/embeddings";

const provider = openAIEmbeddings({ apiKey: process.env.OPENAI_API_KEY! });
const vector = new PgVector({
  id: "arivie-rag",
  connectionString: process.env.DATABASE_URL!,
});

const { chunkCount, totalEmbeddingCost } = await buildIndex({
  layer,
  provider,
  vector,
  indexName: "sem",
});

Returns { chunkCount, totalEmbeddingCost } where cost is usage.tokens × provider.costPerMillionTokens / 1_000_000 summed across embedding batches.

retrieve

Embed a single query string, run similarity search against a populated index, and return ranked Chunk[] (same shape as ParagraphChunker):

import { PgVector } from "@mastra/pg";
import { buildIndex, openAIEmbeddings, retrieve } from "@arivie/embeddings";

const provider = openAIEmbeddings({ apiKey: process.env.OPENAI_API_KEY! });
const vector = new PgVector({
  id: "arivie-rag",
  connectionString: process.env.DATABASE_URL!,
});

await buildIndex({ layer, provider, vector, indexName: "sem" });

const chunks = await retrieve({
  query: "monthly revenue by region",
  vector,
  indexName: "sem",
  provider,
  topK: 5,
  entityHint: "orders", // optional metadata filter: { entity: "orders" }
});

Uses AI SDK embed (single value) for the query vector, then vector.query with optional entityHint filter. Results are sorted by descending similarity score.

Contract and RAG pipeline: RFC-002 §4.4 (amended in Sprint 2 — EmbeddingProvider holds EmbeddingModel + cost metadata, not a custom embed() method).