The Best Answers in Enterprise AI

Your team asks questions all day. The wrong answer wastes hours, the right one closes deals.

Onyx is answering thousands of questions a week at Ramp. We tried a variety of other AI tools but none had the same answer reliability as Onyx. It's been a huge productivity boost as we continue to scale.

Tony Rios, Director of Product Ops at Ramp
Read the Ramp case study →

30x

ROI measured by Ramp

1000+

questions answered / week

~30min

saved per user per day

Trusted by top teams

Benchmarks

Answer Quality

• Enterprise workplace questions
• Deep Research
• Multi-step reasoning tasks

Onyx Wins Head-to-Head Against Every Major Competitor

99 real workplace questions. 220K internal documents. Onyx beat ChatGPT Enterprise, Claude Enterprise, and Notion AI in every matchup.

Head-to-head win rates

• vs ChatGPT 64% win rate
  Onyx 64%
  ChatGPT 36%
• vs Claude 68.1% win rate
  Onyx 68.1%
  Claude 31.9%
• vs Notion AI 76% win rate
  Onyx 76%
  Notion AI 24%

About this benchmark

99 real workplace questions. 220K documents from Slack, Google Drive, GitHub, Gmail, and more. Onyx vs. ChatGPT Enterprise, Claude Enterprise, and Notion AI, scored blind by two independent LLM judges.

Methodology

• 99 questions spanning fact lookup, synthesis, and multi-hop reasoning
• Blind evaluation by GPT-5.2 and Claude Opus 4.5
• 220K documents indexed across 6 enterprise tools

Time to answer

• 34.7s Onyx
• 36.2s Claude
• 45.4s ChatGPT
• 46.7s Notion

Under the hood

Why Onyx finds what others miss

Most enterprise AI tools run a single search and hope for the best. Onyx runs a 6-stage retrieval pipeline that filters noise before the LLM ever sees it.

1. LLM Query Generation: LLM generates multiple parallel queries: a semantic rephrasing, keyword-heavy variants, and broad searches. Multi-part questions are split automatically.
2. Search & Recombination: Each query hits the hybrid search index (vector + BM-25). Results are combined via weighted Reciprocal Rank Fusion and adjacent chunks are merged for continuous context.
3. LLM Selection: The LLM reviews all retrieved chunks across documents and selects the most promising results. Reduces noise and downstream hallucination risk.
4. Context Expansion: For each selected document, the LLM reads surrounding chunks to decide how much context it needs. Runs in parallel per document for reliability.
5. Prompt Building: Selected and expanded document sections are assembled into a structured prompt with citations, chat history, and keyword-matched references.
6. Answer Synthesis: The LLM generates a grounded answer with inline citations linking back to source documents.

What this means in practice

• 343ms median retrieval across 6.8M chunks. The full pipeline adds less than a second, even on CPU.
• 23% recall improvement from adaptive query classification. Onyx detects whether your question needs keyword or semantic search and adjusts automatically.
• Steps 3-4 are the biggest drivers of accuracy. The LLM selects the best chunks, then expands context per document in parallel. Most RAG systems skip both.

It gets smarter over time. User upvotes, admin boosts, and time decay continuously refine ranking. The more your team uses Onyx, the better it gets.