Exa Websets MCP Server

/** * Webset Portal - Deep-dive parallel research through webset URLs */ export async function websetPortal( websetId: string, researchQuery: string, maxPortals: number = 5 ): Promise<string> { if (!websetId) { return `Please provide a webset ID to analyze. Use \`list_websets()\` to see available websets.`; } if (!researchQuery) { return `Please provide a research query describing what insights you're looking for.`; } return `# 🌐 Webset Portal: Deep-Dive Parallel Research ## 🎯 Analyzing Webset: ${websetId} ## 🔍 Research Focus: "${researchQuery}" ## 🚀 Max Parallel Portals: ${maxPortals} This workflow will identify the most relevant URLs in your webset and conduct parallel deep-dive research through each "portal" to synthesize comprehensive insights. ## Phase 1: Identify Portal Candidates First, let's find the most relevant URLs based on your research query: \`\`\`javascript // Step 1: Load webset items and analyze relevance async function identifyPortals() { // Get webset metadata to understand original search const websetInfo = await get_webset("${websetId}"); const originalQuery = websetInfo.searchQuery; // Fetch items from the webset const items = await get_webset_items("${websetId}", 100); // Score each item based on relevance to research query const scoredItems = items.map(item => { let relevanceScore = 0; // Check title relevance const titleMatch = calculateSimilarity(item.title, "${researchQuery}"); relevanceScore += titleMatch * 0.3; // Check content relevance const contentMatch = calculateSimilarity(item.text, "${researchQuery}"); relevanceScore += contentMatch * 0.3; // Bonus for original query match (these are your "full matches") const originalMatch = calculateSimilarity(item.title + " " + item.text, originalQuery); relevanceScore += originalMatch * 0.4; return { ...item, relevanceScore, portalUrl: item.url, portalReason: determineRelevanceReason(item, "${researchQuery}") }; }); // Sort by relevance and take top candidates const topPortals = scoredItems .sort((a, b) => b.relevanceScore - a.relevanceScore) .slice(0, ${maxPortals}) .filter(item => item.relevanceScore > 0.7); // Quality threshold console.log(\`Identified \${topPortals.length} high-relevance portals\`); return topPortals; } // Helper to determine why this URL is relevant function determineRelevanceReason(item, query) { const reasons = []; // Check for exact entity matches const entities = extractEntities(item.text); const queryEntities = extractEntities(query); const matchedEntities = entities.filter(e => queryEntities.includes(e)); if (matchedEntities.length > 0) { reasons.push(\`Contains entities: \${matchedEntities.join(', ')}\`); } // Check for domain authority const domain = new URL(item.url).hostname; if (isAuthorativeDomain(domain)) { reasons.push(\`Authoritative source: \${domain}\`); } // Check for recency if (item.publishedDate && isRecent(item.publishedDate)) { reasons.push('Recent publication'); } return reasons.join('; '); } \`\`\` ## Phase 2: Prepare Parallel Research Tasks Create research tasks for Claude Code subagents: \`\`\`javascript // Step 2: Define research tasks for each portal async function prepareResearchTasks(portals) { const researchTasks = portals.map((portal, index) => ({ taskId: \`portal_research_\${index + 1}\`, portalUrl: portal.portalUrl, portalTitle: portal.title, relevanceScore: portal.relevanceScore, relevanceReason: portal.portalReason, // Research instructions for subagent instructions: \` Research Task: Portal Analysis #\${index + 1} URL: \${portal.portalUrl} Context: \${portal.title} Research Query: "${researchQuery}" Your mission: 1. Visit the portal URL and extract key information 2. Follow relevant links (max depth: 2) for additional context 3. Focus on finding information related to: "${researchQuery}" 4. Extract specific data points, quotes, and facts 5. Identify any contradictions or confirmations with other sources 6. Note any new leads or references worth exploring Deliverables: - Key findings related to the research query - Supporting evidence (quotes, data, facts) - Credibility assessment of the source - Related links worth further investigation - Summary of insights gained \`, // Expected output structure outputSchema: { portalUrl: portal.portalUrl, keyFindings: [], supportingEvidence: [], credibilityScore: 0, relatedLinks: [], summary: "", metadata: {} } })); return researchTasks; } \`\`\` ## Phase 3: Execute Parallel Research (Claude Code Subagents) **For Claude Code execution:** \`\`\`javascript // Step 3: Launch parallel subagents async function executeParallelResearch(researchTasks) { console.log(\`Launching \${researchTasks.length} parallel research agents...\`); // Claude Code will create subagents for each task const subagentPromises = researchTasks.map(async (task) => { console.log(\`🔍 Subagent \${task.taskId} researching: \${task.portalUrl}\`); // Each subagent will: // 1. Use web_search_exa or websets_search for deep dive // 2. Create a focused mini-webset if needed // 3. Analyze content specific to research query // 4. Return structured findings return await executeSubagentResearch(task); }); // Wait for all subagents to complete const results = await Promise.all(subagentPromises); console.log('✅ All portal research completed'); return results; } // Subagent research execution async function executeSubagentResearch(task) { try { // Deep dive into the portal const portalContent = await web_search_exa( \`site:\${new URL(task.portalUrl).hostname} ${researchQuery}\`, 10 ); // Extract and analyze findings const findings = await analyzePortalContent(portalContent, task); // Enrich with additional context if needed if (findings.relatedLinks.length > 0) { const additionalContext = await exploreRelatedLinks( findings.relatedLinks.slice(0, 3), task.instructions ); findings.additionalContext = additionalContext; } return { taskId: task.taskId, portal: task.portalUrl, findings: findings, completedAt: new Date().toISOString() }; } catch (error) { console.error(\`Error in subagent \${task.taskId}:\`, error); return { taskId: task.taskId, portal: task.portalUrl, error: error.message, completedAt: new Date().toISOString() }; } } \`\`\` ## Phase 4: Synthesize Research Findings Combine all parallel research into cohesive insights: \`\`\`javascript // Step 4: Synthesize findings from all portals async function synthesizeFindings(portalResults, originalQuery) { const synthesis = { researchQuery: "${researchQuery}", websetId: "${websetId}", timestamp: new Date().toISOString(), portalsAnalyzed: portalResults.length, // Core insights organized by theme insights: {}, // Evidence supporting each insight evidence: {}, // Contradictions found across sources contradictions: [], // New questions raised emergingQuestions: [], // Actionable recommendations recommendations: [] }; // Group findings by theme const themeGroups = groupFindingsByTheme(portalResults); // For each theme, synthesize insights for (const [theme, findings] of Object.entries(themeGroups)) { synthesis.insights[theme] = { summary: generateThemeSummary(findings), confidence: calculateConfidenceLevel(findings), sources: findings.map(f => ({ url: f.portal, credibility: f.findings.credibilityScore })) }; // Collect supporting evidence synthesis.evidence[theme] = findings .flatMap(f => f.findings.supportingEvidence) .filter(e => e.relevance > 0.8); } // Identify contradictions synthesis.contradictions = findContradictions(portalResults); // Extract emerging questions synthesis.emergingQuestions = extractEmergingQuestions(portalResults); // Generate recommendations synthesis.recommendations = generateRecommendations(synthesis); return synthesis; } \`\`\` ## Phase 5: Generate Analytical Report Create a comprehensive report with insights: \`\`\`javascript // Step 5: Generate the final analytical report async function generateAnalyticalReport(synthesis) { const report = \` # 📊 Webset Portal Analysis Report **Research Query:** "${researchQuery}" **Webset ID:** ${websetId} **Analysis Date:** \${new Date().toLocaleDateString()} **Portals Analyzed:** \${synthesis.portalsAnalyzed} ## 🎯 Executive Summary \${generateExecutiveSummary(synthesis)} ## 🔍 Key Insights \${Object.entries(synthesis.insights).map(([theme, insight]) => \` ### \${theme} **Summary:** \${insight.summary} **Confidence Level:** \${insight.confidence}/10 **Supporting Sources:** \${insight.sources.length} portals \${insight.sources.map(s => \`- \${s.url} (Credibility: \${s.credibility}/10)\`).join('\\n')} \`).join('\\n')} ## 📋 Supporting Evidence \${formatEvidence(synthesis.evidence)} ## ⚠️ Contradictions & Discrepancies \${formatContradictions(synthesis.contradictions)} ## 🤔 Emerging Questions Based on this research, the following questions warrant further investigation: \${synthesis.emergingQuestions.map((q, i) => \`\${i + 1}. \${q}\`).join('\\n')} ## 💡 Recommendations \${synthesis.recommendations.map((r, i) => \` ### Recommendation \${i + 1}: \${r.title} \${r.description} **Next Steps:** \${r.nextSteps.map(step => \`- \${step}\`).join('\\n')} \`).join('\\n')} ## 📊 Research Methodology - **Portal Selection:** Top \${synthesis.portalsAnalyzed} URLs by relevance score - **Research Depth:** 2-level deep dive from each portal - **Analysis Method:** Parallel subagent research with synthesis - **Quality Control:** Cross-validation across multiple sources ## 🔗 Further Research Paths \${generateFurtherResearchPaths(synthesis)} \`; return report; } \`\`\` ## 📝 Complete Workflow Example \`\`\`javascript // Full portal research workflow async function runWebsetPortalAnalysis() { try { console.log("🌐 Starting Webset Portal Analysis..."); // Phase 1: Identify portals const portals = await identifyPortals(); console.log(\`Found \${portals.length} relevant portals\`); // Phase 2: Prepare tasks const tasks = await prepareResearchTasks(portals); // Phase 3: Execute parallel research console.log("🚀 Launching parallel research agents..."); const results = await executeParallelResearch(tasks); // Phase 4: Synthesize findings console.log("🔄 Synthesizing research findings..."); const synthesis = await synthesizeFindings(results); // Phase 5: Generate report console.log("📄 Generating analytical report..."); const report = await generateAnalyticalReport(synthesis); // Save report const reportPath = \`./reports/portal-analysis-\${Date.now()}.md\`; await saveToFile(reportPath, report); console.log(\` ✅ Portal Analysis Complete! - Portals analyzed: \${portals.length} - Insights discovered: \${Object.keys(synthesis.insights).length} - Report saved to: \${reportPath} \`); return { synthesis, report, reportPath }; } catch (error) { console.error("Portal analysis failed:", error); throw error; } } \`\`\` ## 💡 Usage Tips 1. **Quality over Quantity**: Better to deeply analyze 3-5 highly relevant portals than skim 20 2. **Leverage Parallelism**: Claude Code's subagents can research simultaneously 3. **Follow the Thread**: Let subagents explore 2-3 levels deep from each portal 4. **Cross-Validate**: Look for confirmations and contradictions across sources 5. **Time Management**: Set time limits for each subagent (e.g., 5 minutes per portal) ## 🎯 Best Use Cases - **Deep Company Research**: Analyze multiple sources about a specific company - **Technology Assessment**: Understand a technology through various expert sources - **Market Intelligence**: Gather insights from industry publications and analyses - **Academic Research**: Cross-reference findings across scholarly sources - **Competitive Analysis**: Deep dive into competitor information from multiple angles Ready to start portal analysis? The system will: 1. Identify the most relevant URLs from your webset 2. Launch parallel research through each portal 3. Synthesize findings into actionable insights 4. Generate a comprehensive analytical report`; }