Skip to content

How to conduct deep research

The /research command performs comprehensive research on a topic by querying multiple sources and synthesizing them into a detailed report.

Prerequisites

With Exa API key (recommended):

  • Higher quality, curated sources
  • Faster research (30-90 seconds)
  • Costs $0.01-0.05 per research

Without Exa API key (free fallback):

  • Uses DuckDuckGo search + your LLM
  • Slower (2-5 minutes)
  • Broader but potentially noisier sources
  • Free (only LLM token costs)

To configure Exa:

  1. Click the ⚙️ Settings button
  2. Get an API key from Exa
  3. Paste it into the "Exa API Key" field
  4. Click Save

If you don't have an Exa key, /research will automatically use the DuckDuckGo fallback.

Basic research

Type /research followed by your research topic:

/research recent advances in CRISPR gene editing for treating sickle cell disease

Press Enter. Canvas Chat will:

  1. Create a RESEARCH node
  2. Stream a comprehensive report as it's generated
  3. Cite sources automatically

The research process typically takes 30-90 seconds depending on the topic's complexity.

How research works

With Exa API key: Exa's research API handles the entire process:

  1. Plans the research - Breaks down your topic into sub-questions
  2. Searches multiple sources - Queries the web for relevant information
  3. Synthesizes findings - Combines information from all sources into a coherent report
  4. Cites sources - Includes links to the pages used

With DuckDuckGo fallback (no Exa key): The system performs iterative research using your LLM:

  1. Generates search queries - Converts your instructions into DuckDuckGo search terms
  2. Searches and fetches - Finds pages and extracts content (3-5 iterations)
  3. Summarizes sources - Your LLM creates tailored summaries for each page
  4. Expands queries - Generates new queries to explore adjacent topics
  5. Synthesizes report - Combines all summaries into a final report

You'll see status updates as it progresses. Example status messages include:

  • "Generating initial search queries..."
  • "Iteration 1/4: searching DuckDuckGo..."
  • "Iteration 1: fetching 12 pages in parallel..."
  • "Synthesizing final report..."

The exact messages depend on which provider is used.

Context-aware research

When you select text or nodes before running /research, the AI refines your instructions based on that context.

Example: Building on a conversation

  1. Have a discussion about business ideas in several nodes
  2. Select a node mentioning "sustainable fashion marketplace"
  3. Type /research market size and competitors
  4. The AI refines this to: "Research market size and competitors for sustainable fashion marketplace, including industry trends and key players"

The research node shows both your original instruction and the refined version.

Research quality: Standard vs Pro

Exa offers two research models:

  • exa-research (default) - Fast, good for most topics
  • exa-research-pro - Slower, more comprehensive, better for complex topics

Future enhancement: Pro mode (exa-research-pro) is not yet exposed in the UI. To use it, you would need to modify the API call in the code. This feature may be added in a future release.

Working with research results

Read the report

Research nodes are wider than normal nodes (500px) to accommodate formatted markdown reports. The report includes:

  • An introduction to the topic
  • Key findings organized by theme
  • Supporting details with citations
  • A conclusion or summary

Citations and sources

Click any cited link to open the source in a new browser tab. Sources appear as markdown links inline: [Source Name](url).

Branch from findings

Select specific text in the research report and click 🌿 Branch to create a highlight node. This lets you:

  • Ask follow-up questions about a specific finding
  • Compare findings from multiple research nodes
  • Build a knowledge graph around key insights

Continue the conversation

Reply to the research node to ask clarifying questions:

Can you explain the third-generation CRISPR tools mentioned in the report?

The AI has access to the full research content and can elaborate on any section.

Research positioning

Research nodes appear automatically:

  • If you have nodes selected: research appears to the right
  • If nothing is selected: research appears to the right of the most recent node

Research nodes are 500px wide (vs 360px for normal nodes) to better display formatted reports.

Tips for effective research

Be specific but not narrow

✅ Good:

/research quantum error correction techniques used in superconducting qubits, including surface codes and recent improvements

❌ Too vague:

/research quantum computing

❌ Too narrow (use /search instead):

/research exact page count of Nature paper 10.1038/12345

Use context for follow-up research

After getting initial results, select the research node and run a follow-up:

/research how could these techniques be applied to topological qubits?

The AI uses your first research as context for the second.

Combine with other features

Research → Matrix evaluation:

  1. Research multiple competing approaches
  2. Select the research node
  3. Run /matrix compare these approaches against ease of implementation, scalability, and cost

Research → Committee:

  1. Get research findings
  2. Run /committee what are the biggest risks with this approach?
  3. Multiple AI models debate the risks based on your research

Use /search when you want to:

  • Browse multiple sources yourself
  • Quickly find a specific page or fact
  • See what information is available
  • Search without an Exa API key (uses DuckDuckGo as fallback)

Use /research when you want:

  • A synthesized report combining multiple sources
  • Comprehensive coverage of a topic
  • Citation-backed analysis

Note: /research works with or without an Exa API key. Without Exa, it uses a free DuckDuckGo-based fallback that performs iterative research using your LLM. The fallback is slower but produces comprehensive reports.

Limits

With Exa:

  • Costs $0.01-0.05 per research
  • Takes 30-90 seconds to complete
  • Can be stopped using the stop button in the node header

With DuckDuckGo fallback:

  • Free (only LLM token costs)
  • Takes 2-5 minutes (more API calls)
  • May be rate-limited by DuckDuckGo (retries automatically)
  • Quality depends on your LLM model
  • Can be stopped using the stop button in the node header

Both:

  • Wide nodes may overflow on small screens
  • Can be stopped and resumed using the stop/continue buttons

Troubleshooting

"Research failed: 402 Payment Required"

  • Your Exa account has run out of credits
  • Add credits at Exa

Research returns very brief results

  • Topic may be too narrow or too obscure
  • Try rephrasing with more context
  • Consider using /search for niche topics

Sources are not clickable

  • Check that the research completed successfully
  • Sources should appear as markdown links [text](url)
  • If plain URLs appear, the research may have been interrupted

DuckDuckGo fallback returns irrelevant sources

  • This may indicate rate limiting
  • The system automatically retries with exponential backoff
  • If you see "Warning: Only found X relevant sources", wait a few minutes and try again
  • Consider adding an Exa API key for more reliable results

Research report is cut off mid-sentence

  • This can happen if the LLM hits token limits during synthesis
  • The system detects truncation and adds a warning note
  • Try rephrasing your research query to be more specific
  • With Exa, this is less common as Exa handles synthesis internally