Dbhub vs Voyage AI

Voyage AI's capability story is retrieval accuracy. It ships three lanes: general-purpose voyage-3.5 and 3.5 lite, domain models tuned for finance, legal, and code, and company-specific models fine-tuned on your proprietary data. The Voyage 4 series and voyage-multimodal-3.5 extend the same pipeline to images and text. Embeddings run 3x-8x shorter in dimensionality, which directly cuts vector storage and search cost, models are 4x smaller, and 32K-token context lets you embed long documents whole instead of chunking aggressively. rerank-2.5 and rerank-2.5-lite add instruction following so you can steer ranking, voyage-context-3 keeps chunk-level detail while preserving global document context, and the Batch API handles large-scale embedding jobs.

DBHub answers a completely different question: how does an AI coding client reach a database? It loads two tools by default — execute_sql with transactions and safety controls, and search_objects for schemas, tables, columns, indexes, and procedures — at roughly 1.4k tokens. explain_sql and health_check are opt-in so they cost nothing until you need query plans or connection-pool stats. Custom Tools turn recurring SQL into reusable parameterized operations, TOML multi-connection config drives several databases from one process, and read-only mode, row limits, and query timeouts act as guardrails. v1.0.0 adds MCP 2026-07-28 support with a stateless model, cacheable tool lists, and header-based routing.

Voyage makes retrieval smarter. DBHub makes databases reachable. They share no capability surface.

Voyage AI lists pricing_type as contact — there is no public self-serve price in the provided data. That is the signal that matters for budgeting: you are in enterprise procurement territory, with SOC 2 and HIPAA compliance called out as requirements for regulated buyers, and a not_for list that explicitly excludes small teams and hobby projects testing whether RAG works before committing budget. The value argument is cost avoidance rather than sticker price: 3x-8x shorter embeddings cut vector storage and search bills, the models are 2x cheaper to run inference on and 4x smaller, and rerank-2.5/2.5-lite were released as a new price-performance frontier. If you cannot get a quote, you are not the buyer.

DBHub is free and MIT-licensed, installs via Docker or NPM, and has zero dependencies. The cost comparison DBHub itself makes is context, not dollars: ~1.4k tokens for its two default tools versus roughly 19.0k for MCP Toolbox's 28 defaults and 19.3k for Supabase MCP's full tool set, per the vendor's own table. That's headroom returned to the model's reasoning. The real expense of DBHub is configuration and operational care — TOML connections, read-only modes, row limits, timeouts, SSH tunneling, and SSL/TLS — plus the absence of a vendor SLA or support contract, which its not_for list states plainly.