Research & Web Search
Research & summarize searches the selected topic or current note title, asks your configured LLM to summarize the available context, and appends the result to the active Markdown note. Select Tavily or DuckDuckGo explicitly. Search results and model-generated citations require review.
Overview
Configure the search provider separately from the LLM that writes the summary. Tavily requires a key. DuckDuckGo does not require one, but still makes network requests and can be throttled. A local summarization model does not make web research offline.
How It Works
Search-then-Summarize Pipeline
- Open the target note in a Markdown editor. Select a short topic, or leave the selection empty to use the note title.
- Run Research & summarize from the sidebar or command palette.
- Notemd queries the configured search provider. Tavily snippets are used as context; DuckDuckGo results can trigger page-content fetches.
- If local knowledge retrieval is enabled for this task, scoped local excerpts can also enter the prompt.
- The research provider/model writes a summary. Notemd appends it under a
Research Summaryheading identifying the topic and context source. - Review the appended text and open relevant source URLs. Search retrieval does not establish that every generated statement or citation is supported.
Tavily vs. DuckDuckGo
| Choice | Setup | Operational boundary |
|---|---|---|
| Tavily | Select Tavily and enter its API key | Uses the account's search quota and returned snippets |
| DuckDuckGo | Select DuckDuckGo explicitly | No key; subject to search throttling and result-page availability |
An empty or invalid Tavily key does not automatically select DuckDuckGo. If search produces no useful context, inspect the report. When enabled and available, local knowledge context can still support a summary; its source label distinguishes that result from web-backed research.
Research During Title Generation
Enable Enable research in "Generate from title" to include research in title generation, including its batch variant. It is off by default. This is the supported folder-generation route; there is no standalone generic Research Folder action.
Configuration
| Setting | Default | Effect |
|---|---|---|
searchProvider | tavily | Select the search backend |
tavilyApiKey | '' | Supply a key before using Tavily |
tavilyMaxResults | 5 | Requested Tavily result count |
tavilySearchDepth | basic | Tavily search depth |
ddgMaxResults | 5 | Requested DuckDuckGo result count |
ddgFetchTimeout | 15 | Result-page fetch timeout in seconds |
maxResearchContentTokens | 3000 | Approximate web-context budget |
researchProvider | DeepSeek | Summarization provider when task-specific routing is enabled |
researchModel | '' | Use that provider's configured model when empty |
researchSummarizeLanguage | en | Output choice when task-specific languages are enabled |
enableResearchInGenerateContent | false | Opt in to research during title generation |
enableLocalKnowledgeRetrieval | false | Enable scoped local retrieval separately |
enableLocalKnowledgeForResearchSummarize | false | Use local retrieval for this task |
Context And Model Selection
Context limits use an estimate rather than provider-exact tokenization. Truncation can omit evidence. Choose a model and output budget appropriate for the source material, and keep the research question focused. Local indexing is lexical; it does not imply embedding-based semantic search or complete recall.
Example
Create a note titled Relative positional encoding. Run research, then verify the explanation against the linked original papers or official documentation. Keep the question, actual source URLs and any corrections in the note so another reader can audit the conclusion. Example text in a guide is not a research result produced for your Vault.
Cancellation And Privacy
Cancel stops further task work where supported; completed appends remain. A provider or page fetch already in progress may finish. The search service receives the query, fetched websites receive requests, and the selected LLM receives the collected prompt context, including local excerpts when enabled. Review diagnostic reports before sharing them.
Tips
- Distinguish a web source, a local note and a model-generated inference.
- Verify quotations, dates and high-impact claims in the original material.
- Use a small test query to diagnose search and LLM failures separately.
- Inspect existing summaries before rerunning to avoid redundant appended material.
Next Steps
- Configuration: search and local-context settings.
- Custom prompts: supported research placeholders.
- Troubleshooting: credentials, quotas and missing context.