Deep Research on Collections: A Comprehensive Guide

What is Deep Research?

Deep Research is a powerful feature that automatically transforms your saved collection of research materials into a comprehensive, academic-quality report. Instead of manually reading through all your sources and writing summaries yourself, the system orchestrates an intelligent process that:

Analyzes what you've collected to understand your research topic
Plans a structured approach with focused questions
Reuses research you've already completed to save time
Generates detailed sections with proper citations
Synthesizes everything into a polished final report
Evaluates the quality and suggests improvements

Think of it as having a research assistant who can read through all your materials, understand the connections between them, and write a well-organized report with proper academic citations.

Understanding Collections

Before diving into Deep Research, it's important to understand what collections are in this system.

What is a Collection?

A collection is like a digital research folder where you gather and organize materials related to a specific topic. You might create a collection for "Climate Change Policy," "Machine Learning Applications," or "Renaissance Art History."

For detailed information about collections, see the Collections Guide.

What Can You Put in a Collection?

Collections can contain three types of items:

Excerpts - Highlighted passages or quotes from academic papers, books, articles, or other texts you've read. These are the raw materials of your research.
Research Results - Previous research answers you've generated on related topics. These are complete responses to past research questions.
Research Queries - Specific research questions you've asked or want to explore. These help organize your thinking.

Example Collection

Imagine you're researching "Artificial Intelligence Ethics." Your collection might contain: - 10 excerpts from academic papers about algorithmic bias - 5 excerpts from articles about privacy concerns - 3 previous research results about fairness frameworks - 2 research queries about regulation approaches

When you have at least 5 items in a collection, Deep Research becomes available.

How Deep Research Works

Deep Research follows a systematic process that mirrors how a human researcher would approach synthesizing information from multiple sources.

The Big Picture

Understand Your Topic - The system analyzes your collection to grasp what you're researching
Create a Plan - It develops focused sub-questions that break down your topic
Find What You Already Know - It checks if you've already researched similar questions
Gather Relevant Information - For each question, it selects the most relevant materials
Generate Detailed Answers - It writes comprehensive sections with citations
Check Quality - It evaluates whether the research meets academic standards
Combine Everything - It synthesizes all sections into one cohesive report

Why This Approach?

Breaking research into focused sub-questions makes it: - More manageable - Each question is tackled separately - More thorough - Every important aspect gets attention - Better organized - The final report has clear structure - Higher quality - Focused questions lead to detailed answers

Two Ways to Use Deep Research

Deep Research offers two modes depending on how much control you want:

Mode 1: Manual Research (Step-by-Step Control)

Best for: When you want to review and customize at each stage, or when you're working on something important that needs careful oversight.

How it works: - You progress through each step one at a time - You can review, edit, and approve at every stage - You have full visibility into what the system is doing - You can adjust the research plan, choose which results to reuse, and refine outputs

The steps you'll follow:

Review the Research Plan - See the suggested goal and sub-questions; edit if needed
Choose Result Reuse - For each question, decide whether to reuse existing research or generate new content
Verify Context Selection - See what materials were selected for each question
Generate Each Section - Create detailed answers for each sub-question
Review the Synthesis - See how everything comes together
Check Quality - Review metrics and decide if refinement is needed

Time investment: More hands-on, but gives you complete control over the research direction.

Mode 2: Automated Research (Fully Automatic)

Best for: Quick insights, well-curated collections, or when you trust the system to make good decisions on your behalf.

How it works: - You click one button: "Start Automated Research" - The system handles every step in the background - A progress indicator shows you which phase it's currently working on - You get notified when the report is ready - All the same quality checks happen automatically

What happens automatically: - Research planning (generates sub-questions) - Result matching (finds reusable research) - Context assembly (selects relevant materials) - Section generation (writes detailed answers) - Quality evaluation (checks and refines if needed) - Final synthesis (combines into report)

Time investment: Minimal - just initial setup and final review.

The Research Process Explained

Let's dive deeper into each phase of the research process:

Phase 1: Research Planning

Template Used: research_planning (customizable in Settings > Templates)

What happens: The system analyzes your collection to understand: - What topic you're researching (based on collection name and description) - What materials you have available (excerpts, results, queries) - What key themes emerge from your materials - What questions would be most valuable to explore

What it produces: A research plan containing: - A clear research goal statement - 3-7 focused sub-questions that break down your topic - An explanation of why each question matters - An estimate of how much information is needed for each question - A synthesis approach describing how sections will be combined

Example: For a collection on "AI Ethics," the plan might include sub-questions like: - "What are the main types of algorithmic bias in AI systems?" - "How can fairness be measured in machine learning models?" - "What regulatory frameworks exist for AI governance?"

Phase 2: Result Matching and Reuse

What happens: Before generating new research, the system checks if you've already answered similar questions. This is both cost-effective and efficient.

How it works: 1. Semantic Comparison - Compares each new sub-question to all your past research results using advanced similarity matching (not just keywords, but meaning) 2. Quality Scoring - Evaluates past results based on: - Number of citations included - Recency of the research - Comprehensiveness of the answer - Diversity of sources used 3. Recommendation - Suggests one of four strategies: - Exact Reuse - The existing result perfectly answers the question - Partial Reuse - Use the existing result as a foundation and expand - Ensemble - Combine multiple related results - Generate New - Create fresh research for this question

Why this matters: - Saves money - Reduces AI processing costs by reusing work - Maintains consistency - Builds on your previous research - Speeds up process - No need to regenerate what you already know - Improves quality - Leverages your best previous work

Thresholds: - Results must match at least 85% similarity to be considered - Results must score at least 0.75 out of 1.0 on quality metrics

Phase 3: Context Assembly

What happens: For each sub-question, the system gathers the most relevant information from your collection.

The process: 1. Selection - Identifies which collection items are relevant to this specific question 2. Retrieval - Fetches the full text of selected items 3. Deduplication - Removes repeated passages (if the same quote appears in multiple items) 4. Prioritization - Orders items by relevance and type: - Research results (most relevant) - Excerpts (original sources) - Research queries (for context) 5. Token Management - Ensures the context fits within limits (typically 20,000-40,000 tokens, roughly 5,000-10,000 words)

Why careful context selection matters: - Too little context → Superficial answers - Too much context → Information overload and wasted processing - Wrong context → Answers that miss the point

The system aims for the "Goldilocks zone" - just the right amount of relevant information.

Phase 4: Section Generation

Template Used: section_synthesis (customizable in Settings > Templates)

What happens: Using the assembled context, the system generates a detailed, well-cited answer to each sub-question.

What goes into each section: - Comprehensive answer - Thorough exploration of the question - Inline citations - Proper attribution in format like [Author Year] - Source integration - Synthesizes information from multiple sources - Academic tone - Professional, objective writing style - Structured content - Logical flow with clear arguments

Metadata captured: - Word count - Number of citations included - Number of unique sources cited - Which collection items were used - Generation timestamp

Quality focus: The system aims for: - High citation density (frequent references to sources) - Source diversity (drawing from multiple works, not just one or two) - Coherent narrative (reads smoothly, not choppy) - Comprehensive coverage (addresses all aspects of the question)

Phase 5: Quality Evaluation

Template Used: quality_evaluation (customizable in Settings > Templates)

What happens: After generating content, the system automatically evaluates whether it meets academic standards.

What gets measured:

Citation Coverage (0-100%)
What percentage of your collection sources are cited?
Higher is better - shows comprehensive use of materials
Target: At least 60-70% coverage
Source Diversity (count)
How many unique sources are cited?
More diversity = stronger, more balanced research
Target: At least 5-7 different sources for major topics
Coherence (qualitative)
Does the text flow logically?
Are arguments clear and well-structured?
Is the writing style consistent?
Evaluated by analyzing text structure and transitions

What happens if quality is low: - The system can automatically trigger refinement - Refinement regenerates the section with adjusted parameters - Can attempt up to 2 refinements per section - Each refinement aims to improve specific weak points (more citations, better diversity, clearer structure)

Manual override: In manual mode, you can review quality metrics and decide whether to: - Accept the section as-is - Trigger refinement - Regenerate with different context - Edit manually

Phase 6: Final Synthesis

Template Used: synthesis (customizable in Settings > Templates)

What happens: All individual sections are combined into a polished, cohesive final report.

The synthesis process: 1. Integration - Sections are arranged in logical order 2. Executive Summary - A high-level overview is generated 3. Introduction - Context and research goals are stated 4. Body - Individual sections for each sub-question 5. Discussion - Connections and implications are drawn 6. Limitations - Any gaps or constraints are acknowledged 7. References - Complete bibliography of all cited sources

Output format: - Professional markdown formatting - Clear headings and subheadings - Properly formatted citations - Clean, readable structure - Ready to export to PDF, Word, or other formats

What Results You Get

Research Plan Document

A structured plan showing:

Research Goal: [Your overarching research objective]

Key Themes:
- Theme 1
- Theme 2
- Theme 3

Sub-Questions:
1. [Focused question exploring one aspect]
   Rationale: [Why this question matters]
   Estimated context needed: [Token count]

2. [Next question]
   ...

Synthesis Approach: [How sections will be combined]

Individual Research Sections

For each sub-question, you get: - Detailed answer (typically 500-2000 words) - Inline citations linking to your source materials - Quality metrics (citation count, source diversity) - Metadata (generation date, context used, reuse info)

Final Research Report

A comprehensive document containing:

Front Matter: - Title - Executive summary (2-3 paragraphs) - Research questions listed

Body: - Findings for each sub-question - Well-structured paragraphs with proper citations - Logical flow between sections - Synthesized narrative connecting all parts

Back Matter: - Discussion of limitations and caveats - Complete reference list with all sources - Metadata about the research process

Quality Assessment

A report card showing: - Citation coverage: "85% of sources cited" - Source diversity: "12 unique sources used" - Coherence score: "High - well-structured with clear flow" - Completeness: "All sub-questions thoroughly addressed" - Recommendations: Suggestions for improvement if needed

Key Benefits

1. Time Efficiency

Instead of spending hours reading through materials and manually synthesizing, Deep Research automates the heavy lifting. What might take a full day of work can be completed in 30 minutes to an hour.

2. Cost Efficiency

By reusing existing research results, the system minimizes expensive AI processing. If you've already researched related topics, those insights are leveraged rather than regenerated.

3. Academic Rigor

Every claim is backed by citations. Source attribution is maintained throughout, ensuring you never lose track of where information came from. This meets academic standards for proper research.

4. Quality Control

Automatic evaluation catches common research pitfalls: - Too few sources cited - Over-reliance on single sources - Poor organization or flow - Incomplete coverage of topics

5. Flexibility

Choose between hands-on control (manual mode) or full automation based on your needs, timeline, and trust level with the system.

6. Resumability

Life happens. If you need to stop mid-research, your progress is automatically saved. Return anytime and pick up exactly where you left off with a simple "Resume Session" button.

7. Comprehensive Output

You don't just get raw information - you get a polished, professional report ready to share, present, or use as a foundation for further work.

8. Transparency

You can always see: - What sources were used for each section - Why certain materials were selected - What quality scores were achieved - Whether content was reused or generated fresh

Real-World Example

Let's walk through a complete example to see how this works in practice.

Your Starting Point

Collection Name: "Climate Change and Agriculture"

Collection Contents: - 8 excerpts from academic papers on temperature impacts - 6 excerpts from reports on drought effects - 4 excerpts about adaptation strategies - 3 previous research results: - "How does rising temperature affect crop yields?" - "What are the main greenhouse gases from agriculture?" - "Case studies of successful agricultural adaptation" - 2 research queries about future projections

Total: 23 items in your collection

What Deep Research Does

Step 1: Planning

The system analyzes your collection and proposes:

Research Goal: "Understand the multifaceted impacts of climate change on agricultural systems and evaluate potential adaptation strategies"

Sub-Questions Generated: 1. How do rising temperatures and changing precipitation patterns affect major crop yields? 2. What are the mechanisms by which drought impacts agricultural productivity? 3. What adaptation strategies have proven successful in climate-vulnerable agricultural regions? 4. How do different climate scenarios project future agricultural viability? 5. What policy interventions can support agricultural resilience to climate change?

Step 2: Result Matching

The system finds that: - Question 1 has 85% similarity to your existing result on temperature impacts → Recommends: Partial Reuse - Question 2 is new → Recommends: Generate New - Question 3 matches your adaptation case studies at 92% similarity → Recommends: Exact Reuse - Question 4 is new → Recommends: Generate New - Question 5 is new → Recommends: Generate New

Result: 1 exact reuse, 1 partial reuse, 3 new generations needed

Step 3: Context Assembly

For each question, relevant excerpts are selected:

Question 1: 8 temperature-related excerpts + existing result (12,000 tokens total)
Question 2: 6 drought-related excerpts (15,000 tokens)
Question 3: Existing case study result (already complete)
Question 4: 2 future projection queries + 4 relevant excerpts (10,000 tokens)
Question 5: Mixed excerpts on policy and adaptation (18,000 tokens)

Step 4: Section Generation

Five detailed sections are generated:

Section 1 (partial reuse): Builds on existing temperature research, adds new drought insights, 1,200 words, 18 citations

Section 2 (new): Comprehensive drought impact analysis, 900 words, 12 citations

Section 3 (reused): Your existing case study work, 1,500 words, 22 citations

Section 4 (new): Future projections analysis, 1,100 words, 15 citations

Section 5 (new): Policy recommendations, 800 words, 10 citations

Step 5: Quality Check

Evaluation results: - Citation coverage: 78% (18 of 23 sources cited) - Source diversity: 18 unique sources - Coherence: High - Assessment: Meets quality standards ✓

Step 6: Final Report

A 5,500-word comprehensive report is generated with: - Executive summary highlighting key findings - Five detailed sections answering each sub-question - 77 total citations properly attributed - Discussion of implications and limitations - Complete reference list of all 18 cited sources - Professional formatting ready for export

Time and Cost Saved

Traditional approach: - Reading all 23 items: ~3 hours - Manual synthesis: ~4 hours - Writing and citations: ~3 hours - Total: ~10 hours

Deep Research approach: - Automated mode: ~30-45 minutes - Manual review: ~15 minutes - Total: ~1 hour

Research reused: 1 exact + 1 partial = ~40% less AI processing cost

Quality achieved: Academic-standard report with comprehensive citations

Getting Started

Prerequisites

Have a collection with at least 5 items
Items should be related to a common research topic
For best results, include a mix of excerpts and past research results

Choosing Your Mode

Choose Manual Mode if: - You're working on high-stakes research (thesis, publication, important project) - You want to learn how Deep Research works by seeing each step - You need to customize the research plan or questions - You prefer having full control and oversight

Choose Automated Mode if: - You need quick insights from your collection - Your collection is well-curated and focused - You trust the system to make good decisions - You want a starting draft that you'll refine later - Time is a priority

Tips for Best Results

Curate your collection carefully - More focused collections produce better research
Include diverse sources - Don't rely on just one or two papers
Add previous research - The reuse feature is incredibly valuable
Write clear collection descriptions - Helps the planning phase understand your goals
Review quality metrics - Use them to assess whether refinement is needed
Iterate if needed - The first pass is often good enough, but refinement can make it great

Conclusion

Deep Research on Collections transforms the laborious process of synthesizing research materials into an automated, quality-controlled workflow. Whether you choose hands-on control or full automation, the result is the same: a comprehensive, well-cited research report that would have taken hours or days to create manually.

The key to success is building good collections, choosing the right mode for your needs, and trusting the quality control mechanisms to ensure academic rigor. With Deep Research, you can focus your time on higher-level thinking - interpreting results, developing new ideas, and applying insights - rather than the mechanical work of reading, organizing, and citing sources.

Start with a well-curated collection, click "Deep Research," and let the system do the heavy lifting while you maintain the level of control you're comfortable with.

Collections Guide - Learn how to create and manage collections
Settings - Customize templates and research parameters
RAG - Understanding the underlying retrieval and generation system