GenAIScript

script({ title: "Deep Research", description: "Performs comprehensive research on a topic using web searches and multiple iterations", parameters: { topic: { description: "The topic to research", type: "string", default: "quantum computing advancements in the last year", }, iterations: { description: "The number of research iterations to perform", type: "number", default: 3, }, }, }) const { output, vars, generator } = env const { runPrompt } = generator // Define schemas for structured outputs const researchQuestionsSchema = { type: "object", properties: { primaryQuestion: { type: "string", description: "The main research question", }, searchQueries: { type: "array", items: { type: "string" }, description: "Specific search queries to use for web searches", }, subtopics: { type: "array", items: { type: "string" }, description: "Subtopics to explore within the main topic", }, }, required: ["primaryQuestion", "searchQueries", "subtopics"], } const searchAnalysisSchema = { type: "object", properties: { summary: { type: "string", description: "Brief summary of the search results", }, keyFindings: { type: "array", items: { type: "string" }, description: "Key facts or findings from the search results", }, newQuestions: { type: "array", items: { type: "string" }, description: "New questions that arise from these findings", }, sourcesQuality: { type: "number", minimum: 1, maximum: 10, description: "Rating of the quality of the sources (1-10)", }, }, required: ["summary", "keyFindings", "newQuestions", "sourcesQuality"], } const researchReportSchema = { type: "object", properties: { title: { type: "string", description: "Descriptive title for the research report", }, executiveSummary: { type: "string", description: "Concise summary of the entire research", }, keyFindings: { type: "array", items: { type: "object", properties: { finding: { type: "string" }, confidence: { type: "string", enum: ["high", "medium", "low"], }, sources: { type: "array", items: { type: "string" } }, }, required: ["finding", "confidence"], }, }, gaps: { type: "array", items: { type: "string" }, description: "Knowledge gaps that require further research", }, furtherResearch: { type: "array", items: { type: "string" }, description: "Suggested areas for further investigation", }, }, required: ["title", "executiveSummary", "keyFindings"], } /** * Main function to conduct deep research on a topic */ async function conductDeepResearch( topic: string, iterations: number = 3 ): Promise<any> { // Initialize research cache let researchMemory = { topic, startTime: new Date().toISOString(), iterations: 0, searches: [], findings: [], questions: { asked: [], pending: [], }, sources: [], primaryQuestion: "", subtopics: [], completionTime: "", report: {}, } // Step 1: Generate initial research questions if we don't have any if ( !researchMemory.questions.asked?.length && !researchMemory.questions.pending?.length ) { const questionsResult = await runPrompt( (ctx) => { ctx.$`You are an expert researcher tasked with planning a comprehensive investigation on the topic: "${topic}". Break down this topic into: 1. A clear primary research question 2. Specific search queries that would yield valuable information 3. Important subtopics that should be explored Generate search queries that are specific, varied in perspective, and would provide comprehensive information.` }, { label: "generate_research_questions", model: "large", cache: "deep-research", responseSchema: researchQuestionsSchema, } ) if (!questionsResult.json) throw new Error("Failed to generate research questions") researchMemory.primaryQuestion = questionsResult.json.primaryQuestion researchMemory.questions.pending = questionsResult.json.searchQueries || [] researchMemory.subtopics = questionsResult.json.subtopics || [] } // Step 2: Iterative research loop while ( researchMemory.iterations < iterations && researchMemory.questions.pending.length > 0 ) { // Get the next question to research const query = researchMemory.questions.pending.shift() researchMemory.questions.asked.push(query) researchMemory.iterations++ console.log( `Research iteration ${researchMemory.iterations}/${iterations}: "${query}"` ) // Perform web search let searchResults try { searchResults = await retrieval.webSearch(query, { count: 5 }) console.log(`Found ${searchResults.length} search results`) } catch (error) { console.error("Search error:", error) searchResults = [] } if (searchResults.length === 0) { console.log("No search results found for query:", query) continue } // Store source URLs searchResults.forEach((result) => { if ( result.filename && !researchMemory.sources.includes(result.filename) ) { researchMemory.sources.push(result.filename) } }) // Format search results for analysis const searchContent = searchResults .map((result, index) => { const content = result.content || "No content available" return `[Source ${index + 1}: ${result.filename}]\n${content.substring(0, 2000)}...\n` }) .join("\n\n") // Analyze search results const analysisResult = await runPrompt( (ctx) => { ctx.def("SEARCH_RESULTS", searchContent) ctx.$`You are an expert research analyst. Review these search results for the query: "${query}" in <SEARCH_RESULTS>. Analyze these results to identify: 1. Key findings and facts 2. New questions that arise from this information 3. Assess the quality and reliability of these sources (1-10 scale)` }, { label: `analyze_search_results_${researchMemory.iterations}`, model: "small", // Using the smaller model for search analysis cache: "deep-research", responseSchema: searchAnalysisSchema, } ) // Store results in research memory researchMemory.searches.push({ query, timestamp: new Date().toISOString(), sourcesCount: searchResults.length, analysis: analysisResult.json, }) // Add findings and new questions researchMemory.findings = [ ...researchMemory.findings, ...analysisResult.json.keyFindings.map((finding) => ({ text: finding, query: query, sourcesQuality: analysisResult.json.sourcesQuality, })), ] // Add new questions to our research queue const newQuestions = analysisResult.json.newQuestions.filter( (q) => !researchMemory.questions.asked.includes(q) && !researchMemory.questions.pending.includes(q) ) researchMemory.questions.pending.push(...newQuestions) } // Step 3: Generate comprehensive research report console.log("Generating final research report...") // Prepare findings for the report const allFindings = researchMemory.findings .map( (f) => `- ${f.text} (Source: ${f.query}, Quality: ${f.sourcesQuality}/10)` ) .join("\n") const report = await runPrompt( (ctx) => { ctx.def("RESEARCH_TOPIC", topic) ctx.def("PRIMARY_QUESTION", researchMemory.primaryQuestion) ctx.def("FINDINGS", allFindings) ctx.def("SEARCHES", researchMemory.questions.asked.join("\n- ")) ctx.def("SOURCES", researchMemory.sources.join("\n- ")) ctx.$`You are an expert researcher creating a comprehensive report on <RESEARCH_TOPIC>. Create a structured research report with these elements: 1. A descriptive title 2. An executive summary of findings 3. Key findings with confidence levels 4. Knowledge gaps requiring further research 5. Recommendations for further research` }, { label: "generate_research_report", model: "large", // Using the larger model for the final synthesis cache: "deep-research", responseSchema: researchReportSchema, } ) // Final update to research memory with completion info researchMemory.completionTime = new Date().toISOString() researchMemory.report = report.json return { report: report.json, stats: { iterations: researchMemory.iterations, searchesConducted: researchMemory.questions.asked.length, findingsDiscovered: researchMemory.findings.length, sourcesConsulted: researchMemory.sources.length, }, } } // Main execution const { topic, iterations } = vars const result = await conductDeepResearch(topic, iterations) output.appendContent("\n" + MD.stringify(result.report, { headings: 1 }))