Embeddings & Reranking Models
ZeroEntropy zembed-1 vs Voyage embeddings: which performs better on domain-heavy corpora like legal or support tickets?
7 min read
Most teams only discover their embedding model is failing once it hits real traffic: critical clauses go missing in legal search, or the “obvious” past ticket never surfaces in customer support. That’s exactly where the difference between ZeroEntropy’s zembed-1 and Voyage’s embeddings shows up most clearly.
Quick Answer: For domain-heavy corpora like legal documents and support tickets, ZeroEntropy’s zembed-1 typically delivers higher retrieval quality and better production economics than Voyage embeddings, especially when you care about nuance, multilingual content, and large-scale indexing costs.
Frequently Asked Questions
Which performs better for legal and support-ticket search: zembed-1 or Voyage?
Short Answer: In most domain-heavy workloads, zembed-1 will give you more relevant top-k results and better recall on nuanced queries than Voyage embeddings, with strong multilingual support and lower token cost.
Expanded Explanation:
Voyage ships solid general-purpose embedding models. They do well on webby benchmarks and straightforward semantic similarity tasks. But when you move into legal corpora or noisy, multi-turn support logs, the failure modes change: domain vocabulary, cross-jurisdiction nuance, and “buried” relevant evidence become the bottleneck. That’s the scenario we built zembed-1 for.
zembed-1 is trained and tuned specifically for retrieval in high-stakes, domain-heavy environments. On internal and public benchmarks, we see consistent NDCG@10 lifts over general-purpose baselines (including models in Voyage’s class), particularly on long-context, jargon-heavy documents. You also get sub-200ms latency and a token price point of $0.05 per million tokens, which matters when you’re embedding millions of pages or reindexing continuously.
Key Takeaways:
- zembed-1 is optimized for domain-heavy retrieval with better top-k relevance and calibrated behavior on nuanced queries.
- For large legal or support-ticket corpora, zembed-1 typically beats Voyage on both retrieval quality and total indexing cost.
How should I evaluate zembed-1 vs Voyage on my own corpus?
Short Answer: Run a small, labeled evaluation on your actual documents, comparing NDCG@10 and recall@k for zembed-1 and Voyage under the same conditions.
Expanded Explanation:
Benchmarks are useful, but the only evaluation that matters is on your data: your contracts, your clinical notes, your support tickets. The right way to compare zembed-1 and Voyage isn’t “vibes” (the results look good) or generic leaderboards – it’s a controlled experiment with the same queries, same candidate set, and a simple relevance labeling pass.
For legal, that might mean queries around specific clauses or precedents and marking which retrieved documents truly answer the question. For support, you can use historical tickets where you already know the “right” similar past conversations or resolutions. Then you compute NDCG@10 or recall@10 and compare.
Steps:
- Sample your data: Take 50–200 realistic queries from your legal or support workflows, plus the associated relevant documents.
- Embed with both models: Generate embeddings for the same corpus using zembed-1 and Voyage, keeping index and search parameters identical.
- Measure retrieval quality: For each query, compute NDCG@10 and recall@k for both models, then compare aggregate metrics and qualitative failure modes (missed nuance, language issues, lost-in-the-middle).
How do zembed-1 and Voyage embeddings differ in capabilities for domain-heavy use cases?
Short Answer: Voyage focuses on general-purpose semantic similarity, while zembed-1 is tuned for high-precision retrieval in domain-heavy, multilingual corpora with strong cost and latency characteristics.
Expanded Explanation:
Voyage models are solid “one-size-fits-most” embedding models aimed at common semantic search tasks. They’re a good upgrade over naive keyword search but don’t explicitly target the failure zones of legal, medical, or complex support corpora.
zembed-1 was built around those failure modes:
- Domain vocabulary and nuance: It stays robust when the query uses different surface forms than the document (e.g., “change-of-control carve-out” vs the way a clause is phrased in a contract).
- Multilingual by design: zembed-1 supports 100+ languages with strong recall on non-English text, including legal and enterprise jargon. You don’t need separate pipelines per language.
- Economics and scale: At $0.05 per million tokens with sub-200ms latency, you can afford to embed entire knowledge bases, archive ticket logs, and reindex routinely without the bill exploding.
Comparison Snapshot:
- Option A: Voyage embeddings: General-purpose semantic embeddings; good for standard search but less explicitly tuned for domain-heavy nuance and multilingual legal/support workloads.
- Option B: zembed-1: Retrieval-first embeddings optimized for domain-heavy, multilingual corpora with state-of-the-art accuracy, sub-200ms latency, and very low token cost.
- Best for: Teams running serious legal, compliance, or support-ticket search and RAG that care about top-k precision, predictability, and cost at scale.
How do I actually implement zembed-1 if I’m already using Voyage?
Short Answer: Swapping Voyage for zembed-1 is typically an API-level change: re-embed your corpus with zembed-1, rebuild the index, and then point your search or RAG pipeline at the new embeddings.
Expanded Explanation:
From an engineering standpoint, both Voyage and ZeroEntropy expose developer-friendly embedding APIs. Replacing Voyage embeddings with zembed-1 usually means: obtain a ZeroEntropy API key, re-embed your documents, and plug those vectors into your existing vector DB or search backend. If you’re already on a vector DB (like Pinecone, Weaviate, pgvector, etc.), it’s a straight index swap. If you want to go further, you can also adopt ZeroEntropy’s Search API and hybrid retrieval stack (dense + sparse + rerank) for an end-to-end upgrade.
Since zembed-1 is priced at $0.05 per million tokens, you can re-embed large corpora (millions of documents) without blowing up your migration budget. Most teams do it as a background job, cut over read traffic once the new index is warm, and then deprecate the old pipeline.
What You Need:
- A ZeroEntropy API key and access to zembed-1 via the SDK or HTTP API.
- A plan to re-embed and reindex your corpus (vector DB or ZeroEntropy’s Search API) plus a simple cutover strategy for your search/RAG stack.
Strategically, when does it make sense to choose zembed-1 over Voyage?
Short Answer: Choose zembed-1 when search quality directly impacts risk, revenue, or support efficiency—especially in legal, compliance, medical, or complex customer support where nuance and multilingual coverage matter.
Expanded Explanation:
If your search stack is “nice to have,” a generic embedding model might be enough. But when retrieval quality ties directly to real-world outcomes—finding the right indemnity clause before signing, retrieving the correct clinical guideline, or surfacing the most relevant past ticket that unlocks a 30-second support resolution—the bar is higher.
zembed-1 is designed for that bar. It’s part of a larger retrieval stack (with rerankers like zerank-2 and a full hybrid Search API) that treats retrieval as a measurable system: you can track NDCG@10, understand p50/p99 latency, and control token spend in a predictable way. Voyage is a capable embedding vendor, but it isn’t built around the same “human-level retrieval at machine speed” mandate we optimize for.
Why It Matters:
- Business impact: Better top-k retrieval in legal or support search reduces risk (missed clauses, wrong precedents) and improves operational metrics (first-contact resolution, handle time).
- Total cost of ownership: With zembed-1’s low token price and tighter retrieval, you can send fewer, higher-quality chunks to your LLM, cutting RAG spend while increasing answer quality.
Quick Recap
For domain-heavy corpora like legal documents and support tickets, the embedding model you choose is not cosmetic—it determines whether your RAG and search systems actually find what matters. Voyage embeddings are solid general-purpose models, but zembed-1 is purpose-built for this problem space: domain nuance, multilingual coverage, high top-k relevance, and aggressive cost efficiency at scale. If you’re running legal, compliance, or support workloads where retrieval quality isn’t negotiable, zembed-1 is usually the better fit.