Web Research

Researches a topic across the internet, gathering information from multiple sources and synthesizing it into a report with citations.

The workflow:

Clarifies what you’re asking (if ambiguous)
Checks existing knowledge we’ve already gathered
Searches the web iteratively, going deeper based on what it finds
Writes a report with cited sources

When to Use This

This workflow is intended for general web-based research—understanding topics, gathering current information, or surveying perspectives on a question.

For academic papers, use Literature Review. For books, use Book Discovery. For processing a document you already have, use Document Processor.

How It Works

flowchart TD
    subgraph understand["1. Understanding Your Question"]
        A[Your Question] --> B{Need Clarification?}
        B -->|Yes| C[Ask Follow-up Questions]
        C --> D[Create Research Brief]
        B -->|No| D
    end

    subgraph prepare["2. Preparation"]
        D --> E[Check What We Already Know]
        E --> F[Make a Research Plan]
    end

    subgraph research["3. Research Loop"]
        F --> G[Coordinator Reviews Progress]
        G --> H{What's Needed?}
        H -->|More Info| I[Send Researchers to Find Answers]
        I --> J[Combine What They Found]
        J --> K[Update Draft Report]
        K --> G
        H -->|Enough Coverage| L[Move to Final Report]
    end

    subgraph finish["4. Finishing Up"]
        L --> M[Write Final Report]
        M --> N[Format Citations]
        N --> O[Save for Future Reference]
    end

    O --> P[Your Report]

    style understand fill:#e8f4f8
    style prepare fill:#f0f8e8
    style research fill:#fff8e8
    style finish fill:#f8e8f4

The Steps Explained

1. Understanding Your Question If your question is ambiguous (e.g., “Tell me about Apple”—the company or the fruit?), the workflow asks clarifying questions before proceeding.

2. Preparation Checks our existing knowledge base for relevant prior research, then creates a structured plan of areas to explore.

3. Research Loop A coordinator manages parallel “researchers” who search the web. Each round:

The coordinator identifies gaps in coverage
Researchers search for information on specific questions
Findings are combined and a draft report is updated
The coordinator assesses coverage—if insufficient, another round begins

The number of rounds depends on thoroughness setting.

4. Finishing Up Writes the final report, formats citations, and saves findings for future reference.

Inputs

Input	Description	Example
Topic or question	What you want to research	”How are companies using AI for customer service in 2025?”
Thoroughness	How deep to go (see below)	“Standard” for most cases
Language (optional)	Research and write in this language	”es” for Spanish, “zh” for Chinese

Outputs

Output	Description
Report	Written report with inline citations
Sources	Web sources consulted, saved for reference
Status	Success, partial, or failed

Thoroughness Settings

Setting	Depth	Intended Use
Quick	Fewer sources, faster	Quick overview of a topic
Standard	Balanced	General research
Comprehensive	Broader coverage	Topics requiring more depth
High Quality	Extensive sourcing	Publication-quality work

Language Support

Supports 29 languages. When you specify a language:

Search queries are conducted in that language
Sources in that language are prioritized
The final report is written in that language

Example

Input:

Topic: “What are the main approaches to reducing plastic waste in oceans?”
Thoroughness: Standard

Typical output: A ~3,000 word report covering areas like problem scale, cleanup initiatives, prevention strategies, policy approaches, and limitations. Cites 15-25 sources (news articles, reports, expert commentary).

Developer Reference

Entry point: workflows/research/web_research/graph/api.py — exposes deep_research() function

Graph construction: workflows/research/web_research/graph/construction.py — defines the LangGraph state machine, node wiring, and routing logic

State definitions: workflows/research/web_research/state/

workflow_state.py — main DeepResearchState
supervisor_state.py / researcher_state.py — agent-specific state
input_types.py — input validation schemas

Core nodes: workflows/research/web_research/nodes/

clarify_intent.py — disambiguation logic
create_brief.py — research brief generation
search_memory.py — queries existing knowledge stores
iterate_plan.py — builds research plan from brief + memory
supervisor/ — coordinator logic (action selection, gap identification)
refine_draft.py — iterative draft improvement
final_report.py — final synthesis
process_citations/ — citation formatting and Zotero integration
save_findings.py — persists results to stores

Subgraphs: workflows/research/web_research/subgraphs/

web_researcher.py — web search agent (uses researcher_base/ for scraping, query generation)

Prompts: workflows/research/web_research/prompts/ — all LLM prompts, organized by function

Config: workflows/research/web_research/config/ — quality tiers, language definitions

about thala

Explorer