Deep Research

Agentic multi-step research for complex questions

Overview

Deep Research is an agentic feature that enables Onyx to perform multi-step research for complex questions. Unlike standard chat interactions, Deep Research runs multiple cycles of thinking, research, and actions to produce comprehensive, well-sourced answers.

Use Deep Research when you need to:

• Synthesize information from many sources
• Answer questions requiring significant reasoning
• Research topics that span multiple knowledge domains
• Generate comprehensive reports on complex subjects

Deep Research may take up to several minutes and could cost many times (>10x) the token cost of a normal inference.

How it works

Deep Research follows a four-phase process to deliver thorough, accurate results:

1. Clarification

The system analyzes your question to understand the full scope of what you’re asking. If needed, it may ask clarifying questions to ensure the research covers exactly what you need.

2. Planning

Based on your question, the system creates a research plan that breaks down the query into sub-questions and identifies:

• Which knowledge sources to search
• What actions or tools to use
• The order of research steps

3. Orchestrated research

The system executes the research plan, running multiple cycles of:

• Searching internal knowledge via Connectors
• Web searching for external information (if enabled)
• Reasoning through intermediate findings
• Iterating based on what it discovers

Each cycle builds on previous findings, allowing the system to follow leads, fill gaps, and cross-reference sources automatically.

4. Report generation

After completing research, the system synthesizes all findings into a comprehensive response with:

• Clear conclusions and insights
• Citations from all sources used
• Structured formatting for readability

Usage

Enabling Deep Research

Toggle Deep Research using the hourglass icon in the input bar before sending your message.

When to use Deep Research

Use Case	Example
Multi-source synthesis	”Summarize our Q3 performance across all department reports”
Complex analysis	”What are the competitive advantages and risks of our new product line?”
Research reports	”Create a market analysis for expanding into the healthcare sector”
Cross-domain questions	”How do our engineering practices compare to industry standards?”

When standard chat is sufficient

• Quick factual lookups
• Simple questions with single-source answers
• Conversational follow-ups
• Tasks that don’t require extensive research

Admin configuration

Deep Research is available by default when enabled by your Onyx administrator.

Requirements

• An LLM provider configured in AI Models
• Sufficient token limits for extended research sessions
• Optionally, Web Search for external research

Token considerations

Deep Research uses significantly more tokens than standard chat. Administrators can manage costs by:

• Setting appropriate rate limits
• Configuring token limits per user or group
• Monitoring usage via Analytics

Best practices

Frame your question with context about what you’re trying to accomplish. The more specific your question, the more targeted the research.

Writing effective queries

Good query:

“Analyze our customer support ticket trends over the past 6 months and identify the top 3 areas where we could improve response times, based on our internal documentation and industry benchmarks.”

Less effective:

“How can we improve customer support?”

Tips for best results

1. Be specific - Include relevant timeframes, departments, or constraints
2. State your goal - Explain what you’ll do with the information
3. Mention key sources - Reference specific documents or systems if relevant
4. Set scope boundaries - Clarify what should or shouldn’t be included

Working with results

• Review the citations in the right sidebar to verify sources
• Use follow-up questions to drill deeper into specific findings
• Share research sessions with teammates using the share button
• Export or copy findings for use in other documents