@tekmemo/voyageai

Production-ready Voyage AI embedding adapter for TekMemo.

This package is intentionally provider-specific and BYOK-ready. It accepts a Voyage API key from the host application and does not store secrets.

Install

pnpm add @tekmemo/voyageai

Quickstart

import { createVoyageEmbedder } from "@tekmemo/voyageai";

const embedder = createVoyageEmbedder({
  apiKey: process.env.VOYAGE_API_KEY!,
  model: "voyage-4-lite",
  outputDimension: 1024
});

const result = await embedder.embedTexts({
  texts: ["TekMemo is file-first memory infrastructure."],
  inputType: "document"
});

console.log(result.embeddings[0]?.embedding);

Query vs document embeddings

Use inputType: "document" when embedding memory chunks.

Use inputType: "query" when embedding a user query for recall.

await embedder.embedTexts({
  texts: ["What did we decide about memory?"],
  inputType: "query"
});

BYOK

createVoyageEmbedder({
  apiKey: userProvidedVoyageKey,
  model: "voyage-4-lite"
});

The package does not persist or log keys.

Production features

• REST client for POST /v1/embeddings
• BYOK-ready config
• model and output dimension validation
• batch splitting
• retry with exponential backoff and optional jitter
• timeout handling
• response shape validation
• embedding count validation
• finite-number vector validation
• fake Voyage client for tests

What this package does not own

• vector storage
• recall store contracts
• Upstash/Turso/Qdrant/Pinecone integrations
• reranking
• billing
• cloud BYOK encryption