Onyx AI: RAG for Research
Onyx AI released an open-source RAG platform that the authors claim outperforms Claude on deep research tasks and now supports 40+ data sources with local self-hosting. The project is getting traction on GitHub and sparked debate about enterprise RAG versus live runtime querying. (x.com)
Onyx published a detailed benchmarking post on Feb. 19, 2026 that reports a 68.1% win rate against Claude Enterprise on a 99-question workplace research set and an average “time to last token” (response completion) of 34.7 seconds versus Claude’s 36.2 seconds. (onyx.app) The project’s repos show heavy engineering activity with thousands of commits and an active contributor base, and the wider project has drawn community attention on GitHub (roughly 18K stars reported by third‑party trackers). (github.com) (openapps.pro) Onyx’s “Deep Research” feature is built as a long‑running research mode that composes multi‑document reports; their benchmark repo states these jobs are capped at 30 minutes and typically produce outputs around 10,000 tokens (with some runs reaching ~20,000 tokens), which explains why comparisons against short‑response chat assistants focus on different user needs. (github.com) The platform ingests content through a connector system (software plugins that pull and sync data from other tools) and then builds an index for retrieval; Onyx documents list over 40 supported connectors including Slack, Confluence, Google Drive, GitHub, Salesforce, and others, and also detail permission‑syncing so access control from the source can be preserved. (docs.onyx.app) (onyx.app) For live runtime querying, Onyx exposes a Model Context Protocol server (MCP is a standard that lets an external model client request structured context and tools from a server at inference time), and Onyx’s MCP server documentation shows configuration examples for connecting MCP‑compatible clients such as Claude Desktop to an Onyx knowledge base. (docs.onyx.app) (modelcontextprotocol.io) Community discussion has centered on the production tradeoffs this design surfaces — whether to rely on periodically indexed document stores versus providing a live context runtime — with active threads on Hacker News and technical writeups comparing “batteries‑included” RAG platforms like Onyx to build‑your‑own frameworks and arguing about where live querying (MCP) becomes necessary for enterprise workflows. (news.ycombinator.com) (dzone.com)
