Skip to content

~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.

37.7 KB
Raw Blame
import sys
from pathlib import Path
from unittest.mock import Mock

import pytest

# Handle import paths for testing
sys.path.append(str(Path(__file__).parent.parent))
from local_deep_research.report_generator import (
    IntegratedReportGenerator,
)


@pytest.fixture
def mock_search_system():
    """Create a mock search system for testing."""
    mock = Mock()
    mock.analyze_topic.return_value = {
        "findings": [{"content": "Test finding"}],
        "current_knowledge": "Test knowledge",
        "iterations": 1,
        "questions_by_iteration": {1: ["Question 1?", "Question 2?"]},
    }
    mock.all_links_of_system = [
        {"title": "Source 1", "link": "https://example.com/1"},
        {"title": "Source 2", "link": "https://example.com/2"},
    ]
    mock.strategy.settings_snapshot = {"search.iterations": 3}
    mock.strategy.max_iterations = 3
    return mock


@pytest.fixture
def sample_findings():
    """Sample findings for testing."""
    return {
        "findings": [
            {"content": "Finding 1 about AI research"},
            {"content": "Finding 2 about machine learning applications"},
        ],
        "current_knowledge": "AI research has made significant progress in recent years with applications in various fields.",
        "iterations": 2,
        "questions_by_iteration": {
            1: [
                "What are the latest advances in AI?",
                "How is AI applied in healthcare?",
            ],
            2: [
                "What ethical concerns exist in AI development?",
                "What is the future of AI research?",
            ],
        },
    }


@pytest.fixture
def report_generator(mock_llm, mock_search_system, monkeypatch):
    """Create a report generator for testing."""
    monkeypatch.setattr(
        "local_deep_research.report_generator.get_llm", lambda: mock_llm
    )
    generator = IntegratedReportGenerator(search_system=mock_search_system)
    return generator


def test_init(mock_llm, mock_search_system, monkeypatch):
    """Test initialization of report generator."""
    monkeypatch.setattr(
        "local_deep_research.report_generator.get_llm", lambda: mock_llm
    )

    # Test with provided search system
    generator = IntegratedReportGenerator(search_system=mock_search_system)
    # Check that a model was set (might be wrapped)
    assert generator.model is not None
    assert generator.search_system == mock_search_system

    # Test with default search system
    mock_system_class = Mock()
    mock_system_instance = Mock()
    mock_system_class.return_value = mock_system_instance

    monkeypatch.setattr(
        "local_deep_research.report_generator.AdvancedSearchSystem",
        mock_system_class,
    )

    generator = IntegratedReportGenerator()
    # Check that a model was set (might be wrapped)
    assert generator.model is not None
    assert generator.search_system == mock_system_instance


def test_determine_report_structure(report_generator, sample_findings):
    """Test determining report structure from findings."""
    # Mock the LLM response to return a specific structure
    structured_response = """
    STRUCTURE
    1. Introduction
       - Background | Provides historical context of the research topic
       - Significance | Explains why this research matters
    2. Key Findings
       - Recent Advances | Summarizes the latest developments
       - Applications | Describes how the technology is being applied
    3. Discussion
       - Challenges | Identifies current limitations and obstacles
       - Future Directions | Explores potential future developments
    END_STRUCTURE
    """
    report_generator.model.invoke.return_value = Mock(
        content=structured_response
    )

    # Call the method
    structure = report_generator._determine_report_structure(
        sample_findings, "AI research advances"
    )

    # Verify structure was parsed correctly
    assert len(structure) == 3
    assert structure[0]["name"] == "Introduction"
    assert len(structure[0]["subsections"]) == 2
    assert structure[0]["subsections"][0]["name"] == "Background"
    assert (
        structure[0]["subsections"][0]["purpose"]
        == "Provides historical context of the research topic"
    )

    assert structure[1]["name"] == "Key Findings"
    assert structure[2]["name"] == "Discussion"

    # Verify LLM was called (can't check exact args if wrapped)
    assert report_generator.model.invoke.called or hasattr(
        report_generator.model, "invoke"
    )


def test_research_and_generate_sections(report_generator):
    """Test researching and generating sections."""
    # Define sample structure
    structure = [
        {
            "name": "Introduction",
            "subsections": [
                {"name": "Background", "purpose": "Historical context"}
            ],
        },
        {
            "name": "Findings",
            "subsections": [
                {"name": "Key Results", "purpose": "Main research outcomes"}
            ],
        },
    ]

    # Mock the search system to return specific results for each subsection
    report_generator.search_system.analyze_topic.side_effect = [
        {
            "current_knowledge": "Background section content about historical context."
        },
        {
            "current_knowledge": "Key results section content with main findings."
        },
    ]

    # Call the method
    sections = report_generator._research_and_generate_sections(
        {"current_knowledge": "Initial findings"}, structure, "Research query"
    )

    # Verify sections were generated correctly
    assert "Introduction" in sections
    assert "Findings" in sections
    assert "# 1. Introduction" in sections["Introduction"]
    assert "Background section content" in sections["Introduction"]

    assert "# 2. Findings" in sections["Findings"]
    assert "Key results section content" in sections["Findings"]

    # Verify search system was called the correct number of times (once per subsection)
    assert report_generator.search_system.analyze_topic.call_count == 2


def test_format_final_report(report_generator, monkeypatch):
    """Test formatting the final report."""
    # Define sample structure and sections
    structure = [
        {
            "name": "Introduction",
            "subsections": [
                {"name": "Background", "purpose": "Historical context"}
            ],
        },
        {
            "name": "Findings",
            "subsections": [
                {"name": "Key Results", "purpose": "Main research outcomes"}
            ],
        },
    ]

    sections = {
        "Introduction": "# Introduction\n\n## Background\n\nBackground content here.",
        "Findings": "# Findings\n\n## Key Results\n\nKey results content here.",
    }

    # Mock format_links_to_markdown after the visible source set is resolved.
    seen = {}

    def mock_format_links(all_links):
        seen["all_links"] = all_links
        return "1. [Source 1](https://example.com/1)\n2. [Source 2](https://example.com/2)"

    monkeypatch.setattr(
        "local_deep_research.utilities.search_utilities.format_links_to_markdown",
        mock_format_links,
    )

    # Call the method
    report = report_generator._format_final_report(
        sections, structure, "Test query"
    )

    # Verify report structure
    assert "content" in report
    assert "metadata" in report
    assert "# Table of Contents" in report["content"]
    assert "Introduction" in report["content"]
    assert "Findings" in report["content"]
    assert "Background content here" in report["content"]
    assert "Key results content here" in report["content"]
    assert "## Sources" in report["content"]

    # Verify metadata
    assert report["metadata"]["query"] == "Test query"
    assert "generated_at" in report["metadata"]
    assert "sections_researched" in report["metadata"]
    assert report["metadata"]["sections_researched"] == 2

    assert (
        seen["all_links"] == report_generator.search_system.all_links_of_system
    )


def test_format_final_report_forwards_body_to_filter(
    report_generator, monkeypatch
):
    """Regression: dropping prose= would silently disable filtering."""
    import local_deep_research.utilities.search_utilities as su

    captured = {}

    def mock_filter(all_links, prose, mode):
        captured["all_links"] = all_links
        captured["prose"] = prose
        captured["mode"] = mode
        return all_links

    monkeypatch.setattr(su, "filter_cited_links", mock_filter)
    monkeypatch.setattr(
        su,
        "format_links_to_markdown",
        lambda all_links: "formatted",
    )
    structure = [
        {"name": "Intro", "subsections": [{"name": "B", "purpose": "p"}]}
    ]
    sections = {"Intro": "Body text with citation [1]."}

    report_generator._format_final_report(sections, structure, "q")

    assert (
        captured["all_links"]
        == report_generator.search_system.all_links_of_system
    )
    assert captured["prose"] == (
        "# Table of Contents\n\n1. **Intro**\n\n\n\n"
        "# Research Summary\n\n"
        "This report was researched using an advanced search system.\n\n"
        "Research included targeted searches for each section and subsection.\n\n"
        "\n---\n\n\nBody text with citation [1].\n\n"
    )
    assert captured["mode"] == "fallback"


def test_format_final_report_filters_uncited_sources(
    report_generator, monkeypatch
):
    """End to end through the real filter: uncited hits leave Sources."""
    monkeypatch.setattr(
        "local_deep_research.config.thread_settings.get_setting_from_snapshot",
        lambda key, default=None, **kwargs: default,
    )
    report_generator.search_system.all_links_of_system = [
        {"title": "Cited", "link": "https://cited.test/a", "index": "1"},
        {"title": "Uncited", "link": "https://uncited.test/b", "index": "2"},
    ]
    structure = [
        {"name": "Intro", "subsections": [{"name": "B", "purpose": "p"}]}
    ]
    sections = {"Intro": "Findings show progress [1]."}

    report = report_generator._format_final_report(sections, structure, "q")

    assert "https://cited.test/a" in report["content"]
    assert "https://uncited.test/b" not in report["content"]
    assert report["metadata"]["initial_sources"] == 1


def test_format_final_report_disabled_mode_keeps_all_sources(
    report_generator, monkeypatch
):
    """report.uncited_sources_mode=disabled restores the legacy view."""
    monkeypatch.setattr(
        "local_deep_research.config.thread_settings.get_setting_from_snapshot",
        lambda key, default=None, **kwargs: "disabled",
    )
    report_generator.search_system.all_links_of_system = [
        {"title": "Cited", "link": "https://cited.test/a", "index": "1"},
        {"title": "Uncited", "link": "https://uncited.test/b", "index": "2"},
    ]
    structure = [
        {"name": "Intro", "subsections": [{"name": "B", "purpose": "p"}]}
    ]
    sections = {"Intro": "Findings show progress [1]."}

    report = report_generator._format_final_report(sections, structure, "q")

    assert "https://cited.test/a" in report["content"]
    assert "https://uncited.test/b" in report["content"]
    assert report["metadata"]["initial_sources"] == 2


def test_format_final_report_strict_mode_drops_uncited_sources(
    report_generator, monkeypatch
):
    monkeypatch.setattr(
        "local_deep_research.config.thread_settings.get_setting_from_snapshot",
        lambda key, default=None, **kwargs: "strict",
    )
    report_generator.search_system.all_links_of_system = [
        {"title": "Cited", "link": "https://cited.test/a", "index": "1"},
        {"title": "Uncited", "link": "https://uncited.test/b", "index": "2"},
    ]
    structure = [
        {"name": "Intro", "subsections": [{"name": "B", "purpose": "p"}]}
    ]
    sections = {"Intro": "No citations here."}

    report = report_generator._format_final_report(sections, structure, "q")

    assert "https://cited.test/a" not in report["content"]
    assert "https://uncited.test/b" not in report["content"]
    assert report["metadata"]["initial_sources"] == 0


def test_generate_report(report_generator, sample_findings, monkeypatch):
    """Test the full report generation process."""

    # Mock the component methods with Mock objects
    mock_determine_structure = Mock(
        return_value=[
            {
                "name": "Section",
                "subsections": [{"name": "Subsection", "purpose": "Purpose"}],
            }
        ]
    )

    mock_research = Mock(return_value={"Section": "Section content"})

    mock_format = Mock(
        return_value={
            "content": "Report content",
            "metadata": {"query": "Test query"},
        }
    )

    monkeypatch.setattr(
        report_generator,
        "_determine_report_structure",
        mock_determine_structure,
    )
    monkeypatch.setattr(
        report_generator, "_research_and_generate_sections", mock_research
    )
    monkeypatch.setattr(report_generator, "_format_final_report", mock_format)

    # Call generate_report
    result = report_generator.generate_report(sample_findings, "Test query")

    # Verify component methods were called with correct arguments
    mock_determine_structure.assert_called_once_with(
        sample_findings, "Test query"
    )

    # Get the expected structure result
    structure_result = mock_determine_structure.return_value
    mock_research.assert_called_once_with(
        sample_findings, structure_result, "Test query", progress_callback=None
    )

    # Get the expected sections result
    sections_result = mock_research.return_value
    mock_format.assert_called_once_with(
        sections_result, structure_result, "Test query"
    )

    # Verify result is the formatted report
    assert result == mock_format.return_value


def test_generate_error_report(report_generator):
    """Test generating an error report."""
    error_report = report_generator._generate_error_report(
        "Test query", "Error message"
    )

    assert "Test query" in error_report
    assert "Error message" in error_report


def test_context_accumulation_across_sections(report_generator):
    """Test that previous section content is passed to subsequent sections."""
    # Define a structure with multiple sections
    structure = [
        {
            "name": "Section A",
            "subsections": [{"name": "Part 1", "purpose": "First part"}],
        },
        {
            "name": "Section B",
            "subsections": [{"name": "Part 2", "purpose": "Second part"}],
        },
        {
            "name": "Section C",
            "subsections": [{"name": "Part 3", "purpose": "Third part"}],
        },
    ]

    # Track the queries passed to analyze_topic
    captured_queries = []

    def capture_query(query):
        captured_queries.append(query)
        return {"current_knowledge": f"Content for query about {query[:50]}"}

    report_generator.search_system.analyze_topic.side_effect = capture_query

    # Generate sections
    report_generator._research_and_generate_sections(
        {"current_knowledge": "Initial findings"}, structure, "Test query"
    )

    # Verify that analyze_topic was called 3 times
    assert len(captured_queries) == 3

    # First section should NOT have previous context
    assert "CONTENT ALREADY WRITTEN" not in captured_queries[0]

    # Second section SHOULD have previous context from first section
    assert "CONTENT ALREADY WRITTEN" in captured_queries[1]
    assert "Section A" in captured_queries[1]

    # Third section SHOULD have previous context from first and second sections
    assert "CONTENT ALREADY WRITTEN" in captured_queries[2]
    assert "Section A" in captured_queries[2]
    assert "Section B" in captured_queries[2]


def test_context_accumulation_limits_to_last_3_sections(report_generator):
    """Test that only the last 3 sections are included in context."""
    # Define a structure with 5 sections
    structure = [
        {
            "name": f"Section {i}",
            "subsections": [{"name": f"Part {i}", "purpose": f"Purpose {i}"}],
        }
        for i in range(1, 6)
    ]

    # Track the queries
    captured_queries = []

    def capture_query(query):
        captured_queries.append(query)
        return {
            "current_knowledge": f"Content for section {len(captured_queries)}"
        }

    report_generator.search_system.analyze_topic.side_effect = capture_query

    # Generate sections
    report_generator._research_and_generate_sections(
        {"current_knowledge": "Initial findings"}, structure, "Test query"
    )

    # The 5th section should only have context from sections 2, 3, 4 (last 3)
    # Section 1 content should NOT be in the 5th query
    fifth_query = captured_queries[4]

    # Count how many section references are in the context
    # Sections 2, 3, 4 should be present, Section 1 should not
    assert "Section 2" in fifth_query
    assert "Section 3" in fifth_query
    assert "Section 4" in fifth_query
    # Section 1 should have been dropped (only last 3 kept)
    # Check that the context section exists but Section 1 is not in it
    assert "CONTENT ALREADY WRITTEN" in fifth_query


def test_context_truncation_for_large_content(report_generator):
    """Test that context is truncated when it exceeds 4000 characters."""
    # Define structure with sections that will generate large content
    structure = [
        {
            "name": "Section A",
            "subsections": [
                {"name": "Large Part", "purpose": "Generate large content"}
            ],
        },
        {
            "name": "Section B",
            "subsections": [
                {"name": "Next Part", "purpose": "Should see truncated context"}
            ],
        },
    ]

    # Track queries
    captured_queries = []

    def capture_query(query):
        captured_queries.append(query)
        # Return large content for first section (over 4000 chars)
        if len(captured_queries) == 1:
            return {"current_knowledge": "X" * 5000}  # 5000 chars
        return {"current_knowledge": "Normal content"}

    report_generator.search_system.analyze_topic.side_effect = capture_query

    # Generate sections
    report_generator._research_and_generate_sections(
        {"current_knowledge": "Initial findings"}, structure, "Test query"
    )

    # Second query should have truncated context
    second_query = captured_queries[1]
    assert "CONTENT ALREADY WRITTEN" in second_query
    assert "[...truncated]" in second_query


def test_context_includes_section_labels(report_generator):
    """Test that accumulated content includes section/subsection labels."""
    structure = [
        {
            "name": "Introduction",
            "subsections": [
                {"name": "Overview", "purpose": "Provide overview"}
            ],
        },
        {
            "name": "Details",
            "subsections": [
                {"name": "Specifics", "purpose": "Provide details"}
            ],
        },
    ]

    captured_queries = []

    def capture_query(query):
        captured_queries.append(query)
        return {"current_knowledge": "Some content here"}

    report_generator.search_system.analyze_topic.side_effect = capture_query

    report_generator._research_and_generate_sections(
        {"current_knowledge": "Initial"}, structure, "Query"
    )

    # Second query should have labeled content from first section
    second_query = captured_queries[1]
    assert "[Introduction > Overview]" in second_query


def test_no_context_for_first_section(report_generator):
    """Test that the first section has no previous context."""
    structure = [
        {
            "name": "First Section",
            "subsections": [{"name": "First Part", "purpose": "First purpose"}],
        },
    ]

    captured_queries = []

    def capture_query(query):
        captured_queries.append(query)
        return {"current_knowledge": "Content"}

    report_generator.search_system.analyze_topic.side_effect = capture_query

    report_generator._research_and_generate_sections(
        {"current_knowledge": "Initial"}, structure, "Query"
    )

    # First (and only) query should not have previous context markers
    assert "CONTENT ALREADY WRITTEN" not in captured_queries[0]
    assert "DO NOT REPEAT" not in captured_queries[0]


def test_context_accumulation_with_multiple_subsections(report_generator):
    """Test that context accumulates correctly across multiple subsections within a section."""
    structure = [
        {
            "name": "Main Section",
            "subsections": [
                {"name": "Sub A", "purpose": "First subsection"},
                {"name": "Sub B", "purpose": "Second subsection"},
                {"name": "Sub C", "purpose": "Third subsection"},
            ],
        },
    ]

    captured_queries = []

    def capture_query(query):
        captured_queries.append(query)
        return {
            "current_knowledge": f"Content from subsection {len(captured_queries)}"
        }

    report_generator.search_system.analyze_topic.side_effect = capture_query

    report_generator._research_and_generate_sections(
        {"current_knowledge": "Initial"}, structure, "Query"
    )

    # Should have 3 queries (one per subsection)
    assert len(captured_queries) == 3

    # First subsection: no context
    assert "CONTENT ALREADY WRITTEN" not in captured_queries[0]

    # Second subsection: should have Sub A content
    assert "CONTENT ALREADY WRITTEN" in captured_queries[1]
    assert "Sub A" in captured_queries[1]

    # Third subsection: should have Sub A and Sub B content
    assert "CONTENT ALREADY WRITTEN" in captured_queries[2]
    assert "Sub A" in captured_queries[2]
    assert "Sub B" in captured_queries[2]


def test_context_includes_critical_instruction(report_generator):
    """Test that the CRITICAL instruction is included in prompts with context."""
    structure = [
        {
            "name": "Section 1",
            "subsections": [{"name": "Part 1", "purpose": "Purpose 1"}],
        },
        {
            "name": "Section 2",
            "subsections": [{"name": "Part 2", "purpose": "Purpose 2"}],
        },
    ]

    captured_queries = []

    def capture_query(query):
        captured_queries.append(query)
        return {"current_knowledge": "Some content"}

    report_generator.search_system.analyze_topic.side_effect = capture_query

    report_generator._research_and_generate_sections(
        {"current_knowledge": "Initial"}, structure, "Query"
    )

    # Second query should have the CRITICAL instruction
    second_query = captured_queries[1]
    assert "CRITICAL" in second_query
    assert "Do NOT repeat" in second_query
    assert "Focus on NEW information" in second_query


def test_context_with_empty_subsection_content(report_generator):
    """Test that empty subsection results don't break context accumulation."""
    structure = [
        {
            "name": "Section 1",
            "subsections": [
                {"name": "Empty Part", "purpose": "Returns nothing"}
            ],
        },
        {
            "name": "Section 2",
            "subsections": [
                {"name": "Normal Part", "purpose": "Returns content"}
            ],
        },
        {
            "name": "Section 3",
            "subsections": [{"name": "Final Part", "purpose": "Should work"}],
        },
    ]

    call_count = [0]

    def capture_query(query):
        call_count[0] += 1
        if call_count[0] == 1:
            # First subsection returns no content
            return {"current_knowledge": None}
        return {"current_knowledge": f"Content {call_count[0]}"}

    report_generator.search_system.analyze_topic.side_effect = capture_query

    # Should not raise an error
    sections = report_generator._research_and_generate_sections(
        {"current_knowledge": "Initial"}, structure, "Query"
    )

    # Sections should still be generated
    assert "Section 1" in sections
    assert "Section 2" in sections
    assert "Section 3" in sections


def test_context_format_has_clear_delimiters(report_generator):
    """Test that context block has clear start and end delimiters."""
    structure = [
        {
            "name": "First",
            "subsections": [{"name": "A", "purpose": "First"}],
        },
        {
            "name": "Second",
            "subsections": [{"name": "B", "purpose": "Second"}],
        },
    ]

    captured_queries = []

    def capture_query(query):
        captured_queries.append(query)
        return {"current_knowledge": "Content here"}

    report_generator.search_system.analyze_topic.side_effect = capture_query

    report_generator._research_and_generate_sections(
        {"current_knowledge": "Initial"}, structure, "Query"
    )

    second_query = captured_queries[1]

    # Check for clear delimiters
    assert "=== CONTENT ALREADY WRITTEN (DO NOT REPEAT) ===" in second_query
    assert "=== END OF PREVIOUS CONTENT ===" in second_query

    # Verify delimiters appear in correct order
    start_pos = second_query.find("=== CONTENT ALREADY WRITTEN")
    end_pos = second_query.find("=== END OF PREVIOUS CONTENT")
    assert start_pos < end_pos


def test_section_level_vs_subsection_level_prompts(report_generator):
    """Test that both section-level (single subsection) and subsection-level prompts get context."""
    structure = [
        {
            "name": "Standalone Section",
            "subsections": [
                {"name": "Only Part", "purpose": "Single subsection"}
            ],
        },
        {
            "name": "Multi Section",
            "subsections": [
                {"name": "Part A", "purpose": "First of multiple"},
                {"name": "Part B", "purpose": "Second of multiple"},
            ],
        },
    ]

    captured_queries = []

    def capture_query(query):
        captured_queries.append(query)
        return {"current_knowledge": f"Content {len(captured_queries)}"}

    report_generator.search_system.analyze_topic.side_effect = capture_query

    report_generator._research_and_generate_sections(
        {"current_knowledge": "Initial"}, structure, "Query"
    )

    # Query 0: Standalone section (section-level, no context)
    assert "CONTENT ALREADY WRITTEN" not in captured_queries[0]

    # Query 1: First part of multi-section (subsection-level, has context)
    assert "CONTENT ALREADY WRITTEN" in captured_queries[1]
    assert "Standalone Section" in captured_queries[1]

    # Query 2: Second part of multi-section (subsection-level, has context)
    assert "CONTENT ALREADY WRITTEN" in captured_queries[2]


def test_context_preserves_content_integrity(report_generator):
    """Test that the actual content is preserved in context, not just labels."""
    structure = [
        {
            "name": "Section A",
            "subsections": [{"name": "Part 1", "purpose": "First"}],
        },
        {
            "name": "Section B",
            "subsections": [{"name": "Part 2", "purpose": "Second"}],
        },
    ]

    unique_content = "UNIQUE_MARKER_12345_THIS_SHOULD_APPEAR_IN_CONTEXT"
    captured_queries = []

    def capture_query(query):
        captured_queries.append(query)
        if len(captured_queries) == 1:
            return {"current_knowledge": unique_content}
        return {"current_knowledge": "Other content"}

    report_generator.search_system.analyze_topic.side_effect = capture_query

    report_generator._research_and_generate_sections(
        {"current_knowledge": "Initial"}, structure, "Query"
    )

    # The unique content from section 1 should appear in section 2's query
    assert unique_content in captured_queries[1]


def test_context_with_special_characters(report_generator):
    """Test that content with special characters doesn't break context."""
    structure = [
        {
            "name": "Section 1",
            "subsections": [{"name": "Part 1", "purpose": "First"}],
        },
        {
            "name": "Section 2",
            "subsections": [{"name": "Part 2", "purpose": "Second"}],
        },
    ]

    special_content = "Content with 'quotes', \"double quotes\", {braces}, [brackets], and $pecial ch@rs!"
    captured_queries = []

    def capture_query(query):
        captured_queries.append(query)
        if len(captured_queries) == 1:
            return {"current_knowledge": special_content}
        return {"current_knowledge": "Normal"}

    report_generator.search_system.analyze_topic.side_effect = capture_query

    # Should not raise any errors
    report_generator._research_and_generate_sections(
        {"current_knowledge": "Initial"}, structure, "Query"
    )

    # Special content should be in second query's context
    assert special_content in captured_queries[1]


def test_accumulated_findings_count_matches_subsections(report_generator):
    """Test that the number of accumulated findings matches successful subsections."""
    structure = [
        {
            "name": "Section 1",
            "subsections": [
                {"name": "Sub 1", "purpose": "P1"},
                {"name": "Sub 2", "purpose": "P2"},
            ],
        },
        {
            "name": "Section 2",
            "subsections": [
                {"name": "Sub 3", "purpose": "P3"},
            ],
        },
    ]

    call_count = [0]

    def capture_query(query):
        call_count[0] += 1
        return {"current_knowledge": f"Content {call_count[0]}"}

    report_generator.search_system.analyze_topic.side_effect = capture_query

    report_generator._research_and_generate_sections(
        {"current_knowledge": "Initial"}, structure, "Query"
    )

    # Should have called analyze_topic 3 times (once per subsection)
    assert call_count[0] == 3


def test_context_separator_between_sections(report_generator):
    """Test that sections in context are separated by the correct delimiter."""
    structure = [
        {
            "name": f"Section {i}",
            "subsections": [{"name": f"Part {i}", "purpose": f"P{i}"}],
        }
        for i in range(1, 4)
    ]

    captured_queries = []

    def capture_query(query):
        captured_queries.append(query)
        return {"current_knowledge": f"Content {len(captured_queries)}"}

    report_generator.search_system.analyze_topic.side_effect = capture_query

    report_generator._research_and_generate_sections(
        {"current_knowledge": "Initial"}, structure, "Query"
    )

    # Third query should have separator between Section 1 and Section 2 content
    third_query = captured_queries[2]
    assert "---" in third_query  # The separator used between sections


def test_truncate_at_sentence_boundary_no_truncation(report_generator):
    """Test that short text is not truncated."""
    text = "This is short. It should not be truncated."
    result = report_generator._truncate_at_sentence_boundary(text, 1000)
    assert result == text
    assert "[...truncated]" not in result


def test_truncate_at_sentence_boundary_at_sentence(report_generator):
    """Test that truncation happens at sentence boundary when possible."""
    # Use a longer limit so sentence boundary falls within 80% threshold
    text = "First sentence. Second sentence. Third sentence that is very long and goes beyond the limit here."
    result = report_generator._truncate_at_sentence_boundary(text, 40)
    # "First sentence. Second sentence." = 33 chars, which is > 80% of 40 (32)
    # So it should truncate at the sentence boundary
    assert result.startswith("First sentence. Second sentence.")
    assert "[...truncated]" in result
    assert "Third sentence" not in result


def test_truncate_at_sentence_boundary_fallback(report_generator):
    """Test that truncation falls back to hard cut when no good boundary exists."""
    # No sentence boundaries in the first 80% of the limit
    text = "A" * 100  # No sentence boundaries
    result = report_generator._truncate_at_sentence_boundary(text, 50)
    # Should hard truncate at 50 chars
    assert len(result) == 50 + len("\n[...truncated]")
    assert result.startswith("A" * 50)
    assert "[...truncated]" in result


def test_truncate_at_sentence_boundary_with_question_mark(report_generator):
    """Test that question marks are recognized as sentence boundaries."""
    text = "Is this a question? Yes it is. More content that exceeds the limit."
    result = report_generator._truncate_at_sentence_boundary(text, 35)
    assert "Is this a question?" in result
    assert "[...truncated]" in result


def test_truncate_at_sentence_boundary_with_exclamation(report_generator):
    """Test that exclamation marks are recognized as sentence boundaries."""
    text = "Wow! Amazing! This is great content that exceeds the limit significantly."
    result = report_generator._truncate_at_sentence_boundary(text, 20)
    assert "Amazing!" in result or "Wow!" in result
    assert "[...truncated]" in result


def test_build_previous_context_empty(report_generator):
    """Test that empty accumulated findings returns empty string."""
    result = report_generator._build_previous_context([])
    assert result == ""


def test_build_previous_context_single_section(report_generator):
    """Test context building with a single section."""
    accumulated = ["[Section A > Part 1]\nContent for section A"]
    result = report_generator._build_previous_context(accumulated)
    assert "CONTENT ALREADY WRITTEN" in result
    assert "Section A" in result
    assert "CRITICAL" in result
    assert "END OF PREVIOUS CONTENT" in result


def test_build_previous_context_respects_max_sections(report_generator):
    """Test that only max_context_sections sections are included."""
    # Use the instance's max_context_sections (defaults to DEFAULT_MAX_CONTEXT_SECTIONS)
    max_sections = report_generator.max_context_sections

    # Create more sections than the limit
    accumulated = [f"[Section {i}]\nContent {i}" for i in range(10)]
    result = report_generator._build_previous_context(accumulated)

    # Should only include the last max_sections
    for i in range(10 - max_sections):
        # Earlier sections should NOT be in the context
        if f"Section {i}" in result:
            # Check it's not just a coincidental match
            assert f"[Section {i}]" not in result or i >= 10 - max_sections

    # Last sections SHOULD be present
    assert f"Section {10 - 1}" in result  # Last section


def test_build_previous_context_truncates_long_content(report_generator):
    """Test that very long content is truncated."""

    # Create content that exceeds max_context_chars (default 4000)
    long_content = "This is a sentence. " * 500  # ~10000 chars
    accumulated = [f"[Section A]\n{long_content}"]
    result = report_generator._build_previous_context(accumulated)

    # Result should contain truncation marker
    assert "[...truncated]" in result
    # Total context (excluding delimiters) should be around max_context_chars
    # The result includes delimiters so it will be larger


def test_configurable_max_context_sections(
    mock_llm, mock_search_system, monkeypatch
):
    """Test that max_context_sections can be configured via settings_snapshot."""
    monkeypatch.setattr(
        "local_deep_research.report_generator.get_llm", lambda: mock_llm
    )

    # Create generator with custom settings
    settings_snapshot = {
        "report.max_context_sections": 5,
        "report.max_context_chars": 8000,
    }
    generator = IntegratedReportGenerator(
        search_system=mock_search_system,
        settings_snapshot=settings_snapshot,
    )

    # Verify settings were applied
    assert generator.max_context_sections == 5
    assert generator.max_context_chars == 8000


def test_configurable_max_context_sections_affects_context_building(
    mock_llm, mock_search_system, monkeypatch
):
    """Test that custom max_context_sections affects _build_previous_context."""
    monkeypatch.setattr(
        "local_deep_research.report_generator.get_llm", lambda: mock_llm
    )

    # Create generator with only 2 sections in context
    settings_snapshot = {
        "report.max_context_sections": 2,
    }
    generator = IntegratedReportGenerator(
        search_system=mock_search_system,
        settings_snapshot=settings_snapshot,
    )

    # Create 5 sections
    accumulated = [f"[Section {i}]\nContent {i}" for i in range(5)]
    result = generator._build_previous_context(accumulated)

    # Should only include Section 3 and Section 4 (last 2)
    assert "[Section 3]" in result
    assert "[Section 4]" in result
    # Section 0, 1, 2 should NOT be in context (only last 2 kept)
    assert "[Section 0]" not in result
    assert "[Section 1]" not in result
    assert "[Section 2]" not in result


def test_configurable_max_context_chars_affects_truncation(
    mock_llm, mock_search_system, monkeypatch
):
    """Test that custom max_context_chars affects truncation behavior."""
    monkeypatch.setattr(
        "local_deep_research.report_generator.get_llm", lambda: mock_llm
    )

    # Create generator with small context limit
    settings_snapshot = {
        "report.max_context_chars": 100,
    }
    generator = IntegratedReportGenerator(
        search_system=mock_search_system,
        settings_snapshot=settings_snapshot,
    )

    # Create content that exceeds 100 chars but is under default 4000
    content = "This is a sentence. " * 20  # ~400 chars
    accumulated = [f"[Section A]\n{content}"]
    result = generator._build_previous_context(accumulated)

    # Should be truncated due to small limit
    assert "[...truncated]" in result


def test_default_context_settings_when_no_snapshot(
    mock_llm, mock_search_system, monkeypatch
):
    """Test that default values are used when no settings_snapshot is provided."""
    from local_deep_research.report_generator import (
        DEFAULT_MAX_CONTEXT_SECTIONS,
        DEFAULT_MAX_CONTEXT_CHARS,
    )

    monkeypatch.setattr(
        "local_deep_research.report_generator.get_llm", lambda: mock_llm
    )

    generator = IntegratedReportGenerator(search_system=mock_search_system)

    # Should use defaults
    assert generator.max_context_sections == DEFAULT_MAX_CONTEXT_SECTIONS
    assert generator.max_context_chars == DEFAULT_MAX_CONTEXT_CHARS