Aeion Search + Voyage AI Embeddings
Voyage AI's embedding models consistently top retrieval benchmarks. `voyage-4-large` (2048 dim) is Voyage's flagship MoE embedding model, ranked #1 on the RTEB leaderboard. `voyage-code-3` is the best-in-class for code-aware retrieval (engineering agents, codebase Q&A). Aeion Search integrates Voyage natively via the AI gateway — per-tenant BYOK, code-aware embeddings, and side-by-side A/B testing with other embedding providers.
Voyage Model Lineup
`
Why Voyage Over OpenAI?
`
Contextual Quantization
Voyage's embeddings API natively supports int8/binary output dtypes, giving up to 4x (int8) or 32x (binary) storage reduction. Aeion's Voyage adapter currently requests the model's default full-precision output and does not yet forward a quantized dtype — so today, storage reduction with Voyage-backed collections is achieved at the Postgres layer (halfvec columns) rather than by requesting quantized vectors from Voyage's API.
Code-Aware Retrieval with voyage-code-3
Code-search is qualitatively different from natural-language search — function signatures, type annotations, AST structure all matter.
Domain-Specialized Models
`
FAQ
Generic English text: OpenAI 3-small is cheaper at similar accuracy. Code: voyage-code-3 wins. Finance/Law: voyage-finance-2/law-2 wins. Long documents (>8K tokens): Voyage's 32K context avoids chunking.
Add a new column, backfill async, shadow-test for 1-2 weeks, cut over. Aeion Search supports multi-column embeddings natively.
Voyage's API supports int8/binary output dtypes natively, but Aeion's Voyage adapter doesn't yet forward a quantized dtype request — it always requests the model's default full-precision output. Storage reduction today comes from Postgres-side `halfvec` columns, not Voyage-side quantization.
Voyage's default tier is 1M TPM. Aeion Search's embedding worker respects per-tenant TPM/RPM with backoff.
Comparable — 30-100ms per batch. Voyage's longer context window can reduce total cost if it eliminates chunking.
1M documents × 1K tokens avg × $0.06/1M = $60 one-time for voyage-3. Daily refresh on 10K updated docs: $0.60/day. Yearly: ~$280 for a 1M-doc corpus.
Voyage's rerank-2 model is priced at $0.05/1M tokens on Voyage's own API. Aeion's AI module currently wires reranking through the Cohere rerank provider (`ctx.ai.rerank()`) — a direct Voyage rerank adapter isn't implemented yet, so pick Cohere as the rerank provider for post-retrieval re-ranking today.
No retention by default. Enterprise tier offers SOC 2 + HIPAA BAA available.
Yes. Multi-column schema, route per-experiment, attribute results, decide based on conversion / MRR. Same pattern as the OpenAI integration page.