Aeion AI + Voyage — Best-Benchmark Embeddings
Configure Voyage as the embedding provider in Aeion AI. voyage-4-large leads Voyage's current lineup for retrieval quality; voyage-4 is the cost / quality sweet spot; domain-specific variants for code, finance, and law. Pairs with Aeion's built-in vector store for production RAG.
Setup
- Get an API key at
https://dash.voyageai.com/. Free tier: 50M tokens/month (very generous for development).
Model Family — When to Use Which
Voyage's current lineup (Voyage 4 generation, plus the still-active Voyage 3 and domain-specific v2 models).
Why Voyage Over OpenAI / Cohere
Embedding quality directly impacts RAG retrieval precision. The MTEB (Massive Text Embedding Benchmark) leaderboard tracks ~50 retrieval tasks across domains.
RAG Wiring — Voyage + Aeion's Vector Store
End-to-end RAG flow using Aeion's built-in pgvector-backed store.
Domain-Specific Variants
Voyage's domain models outperform general models on their respective corpora.
Common Gotchas
- Token limit per request — Voyage caps at 1000 input texts × 16,000 tokens each per request. Aeion's adapter batches automatically.
FAQ
Mixedbread if you can self-host the embedding model (lowest cost, ~equivalent quality). Voyage if you want hosted + domain-specific variants. Both beat OpenAI on benchmarks.
All are viable. Voyage's `voyage-multilingual-2` benchmarks well on cross-lingual retrieval; Cohere has a broader integration ecosystem. Test on your specific corpus.
For RAG quality, yes — two-stage retrieval (initial top-N by embedding similarity, then re-rank to a smaller top-K) materially improves answer quality. Use Cohere Rerank or Mixedbread for that step, since Voyage's rerank models aren't wired up in Aeion yet.
Aeion's built-in vector store is pgvector-backed — no separate service to run, data stays in the same Postgres database as everything else.
Voyage doesn't ship multimodal embeddings. Use Jina (`jina-embeddings-v3` is multimodal) for image-text retrieval instead.