deep_researcher_start
Initiate a detailed AI-powered research task to analyze and synthesize complex queries. The tool performs extensive web searches, evaluates information, and generates a comprehensive research report. Use with deep_researcher_check to monitor progress and retrieve findings.
Instructions
Start a comprehensive AI-powered deep research task for complex queries. This tool initiates an intelligent agent that performs extensive web searches, crawls relevant pages, analyzes information, and synthesizes findings into a detailed research report. The agent thinks critically about the research topic and provides thorough, well-sourced answers. Use this for complex research questions that require in-depth analysis rather than simple searches. After starting a research task, IMMEDIATELY use deep_researcher_check with the returned task ID to monitor progress and retrieve results.
Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| instructions | Yes | Complex research question or detailed instructions for the AI researcher. Be specific about what you want to research and any particular aspects you want covered. | |
| model | No | Research model: 'exa-research' (faster, 15-45s, good for most queries) or 'exa-research-pro' (more comprehensive, 45s-2min, for complex topics). Default: exa-research |
Input Schema (JSON Schema)
{
"$schema": "http://json-schema.org/draft-07/schema#",
"additionalProperties": false,
"properties": {
"instructions": {
"description": "Complex research question or detailed instructions for the AI researcher. Be specific about what you want to research and any particular aspects you want covered.",
"type": "string"
},
"model": {
"description": "Research model: 'exa-research' (faster, 15-45s, good for most queries) or 'exa-research-pro' (more comprehensive, 45s-2min, for complex topics). Default: exa-research",
"enum": [
"exa-research",
"exa-research-pro"
],
"type": "string"
}
},
"required": [
"instructions"
],
"type": "object"
}
Implementation Reference
- src/tools/index.ts:250-259 (handler)MCP tool handler for deep_researcher_start: parses arguments using schema and calls ExaClient.startDeepResearchcase 'deep_researcher_start': { const params = deepResearchStartSchema.parse(args); const result = await client.startDeepResearch(params); return { content: [{ type: "text", text: JSON.stringify(result, null, 2) }] }; }
- src/tools/index.ts:63-69 (schema)Zod input schema validation for deep_researcher_start tool parameters// Deep Research Start Tool Schema const deepResearchStartSchema = z.object({ topic: z.string().describe("Research topic or question"), research_type: z.enum(['comprehensive', 'news', 'academic', 'market']).optional().default('comprehensive').describe("Type of research to perform"), max_results: z.number().optional().default(50).describe("Maximum total results to collect"), time_range: z.enum(['week', 'month', 'quarter', 'year', 'all']).optional().describe("Time range for research") });
- src/client.ts:221-248 (handler)Core implementation of deep_researcher_start: creates asynchronous task, stores metadata globally, kicks off background research via performDeepResearchasync startDeepResearch(params: { topic: string; research_type?: 'comprehensive' | 'news' | 'academic' | 'market'; max_results?: number; time_range?: 'week' | 'month' | 'quarter' | 'year' | 'all'; }) { const taskId = `research_${Date.now()}_${Math.random().toString(36).substr(2, 9)}`; // Store task metadata (in production, this would be in a database) global.researchTasks = global.researchTasks || {}; global.researchTasks[taskId] = { status: 'in_progress', started_at: new Date().toISOString(), params: params }; // Simulate async research (in production, this would be a queue job) this.performDeepResearch(taskId, params).catch(error => { global.researchTasks[taskId].status = 'failed'; global.researchTasks[taskId].error = error.message; }); return { task_id: taskId, status: 'started', message: 'Deep research task initiated. Use deep_researcher_check to monitor progress.' }; }
- src/client.ts:274-326 (helper)Background helper function that performs the actual deep research: generates section-specific queries, executes multiple web searches, collects results, and updates task statusprivate async performDeepResearch(taskId: string, params: any) { try { const task = global.researchTasks[taskId]; const results: any = { summary: '', sections: [] }; // Determine time range let startDate: string | undefined; if (params.time_range && params.time_range !== 'all') { const now = new Date(); const ranges = { week: 7, month: 30, quarter: 90, year: 365 }; const days = ranges[params.time_range as keyof typeof ranges]; startDate = new Date(now.getTime() - days * 24 * 60 * 60 * 1000).toISOString().split('T')[0]; } // Perform multiple searches based on research type const searchQueries = this.getResearchQueries(params.topic, params.research_type || 'comprehensive'); for (const [section, query] of Object.entries(searchQueries)) { const searchResults = await this.search({ query: query as string, num_results: Math.floor((params.max_results || 50) / Object.keys(searchQueries).length), type: 'neural', start_published_date: startDate, include_text: true, include_summary: true, include_highlights: true }); results.sections.push({ title: section, query: query, results: searchResults }); } // Generate summary results.summary = `Completed deep research on "${params.topic}" with ${results.sections.length} sections and ${results.sections.reduce((acc: number, s: any) => acc + s.results.length, 0)} total results.`; task.status = 'completed'; task.completed_at = new Date().toISOString(); task.results = results; } catch (error) { throw error; } }
- src/server.ts:30-37 (registration)Registers deep_researcher_start as an enabled tool in the ExaServer constructorthis.enabledTools = enabledTools || [ 'web_search_exa', 'company_research_exa', 'crawling_exa', 'linkedin_search_exa', 'deep_researcher_start', 'deep_researcher_check' ];