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.

4.5 KB
Raw Blame
"""Exercise stored source URLs through the route and real document template."""

from pathlib import Path
from types import SimpleNamespace

import pytest
from bs4 import BeautifulSoup
from jinja2 import ChoiceLoader, DictLoader
from starlette.requests import Request

from local_deep_research.web.routers import library


@pytest.fixture
def render_document(monkeypatch):
    env = library.templates.env.overlay(
        loader=ChoiceLoader(
            [
                DictLoader({"base.html": "{% block content %}{% endblock %}"}),
                library.templates.env.loader,
            ]
        )
    )
    monkeypatch.setattr(
        library.templates,
        "TemplateResponse",
        lambda **kwargs: env.get_template(kwargs["name"]).render(
            kwargs["context"]
        ),
    )

    def render(url):
        document = {
            "id": "doc-1",
            "document_title": "Example document",
            "original_url": url,
            "has_pdf": True,
            "has_text_db": True,
        }
        monkeypatch.setattr(
            library,
            "LibraryService",
            lambda username: SimpleNamespace(
                get_document_by_id=lambda document_id: document
            ),
        )
        request = Request({"type": "http", "method": "GET", "path": "/"})
        html = library.document_details_page(request, "doc-1", "alice")
        return BeautifulSoup(html, "html.parser")

    return render


@pytest.mark.parametrize(
    "url",
    [
        "javascript:alert(1)",
        "JaVaScRiPt:alert(1)",
        "java\nscript:alert(1)",
        "\tjavascript:alert(1)",
        "data:text/html,<svg onload=alert(1)>",
        "vbscript:msgbox(1)",
        "file:///etc/passwd",
        "//example.com/paper",
        "/relative/paper",
        "#fragment",
        "https:",
        "https://[invalid/paper",
        "https://:443/paper",
        "https://example.org:invalid/paper",
        "https://example.org:65536/paper",
        "https://example.org/\tpaper",
        "javascript&#58;alert(1)",
        "javascript%3Aalert(1)",
        'javascript:alert(1)"><img src=x onerror=alert(1)>',
    ],
)
def test_rejected_source_url_is_visible_but_never_linked(render_document, url):
    soup = render_document(url)
    assert url in soup.get_text()
    assert not soup.find(
        "a", string=lambda text: text and "Original Source" in text
    )
    assert [a["href"] for a in soup.find_all("a")] == [
        "/library",
        "/library/document/doc-1/pdf",
        "/library/document/doc-1/txt",
    ]
    assert not soup.select("script, img, svg, [onerror], [onload]")


@pytest.mark.parametrize(
    "url",
    [
        "https://example.org/paper?x=1&y=2#results",
        "http://localhost:8080/paper",
        "https://[2001:db8::1]/paper",
        'https://example.org/paper?q="quoted"',
        # A literal percent followed by hexadecimal-looking filename/query
        # text is not double encoding. These were rejected by callback policy.
        "https://example.org/100%25Efficiency.pdf",
        "https://example.org/paper.pdf?q=50%25effect",
        "https://example.org/paper.pdf?formula=A&B;C",
        "https://example.org./paper.pdf",
        "https://example.org/paper.pdf?next=https%3A%2F%2Fexample.org%2Fa%2520b",
        "HTTPS://example.org/paper.pdf",
        'https://example.org/paper?q="><img src=x onerror=alert(1)>&lt;script&gt;',
    ],
)
def test_http_source_links_and_download_links_are_preserved(
    render_document, url
):
    soup = render_document(url)
    source_links = soup.find_all("a", href=url)
    assert len(source_links) == 2
    assert source_links[1].get_text() == url
    links = soup.select('a[target="_blank"]')
    assert len(links) == 4
    assert all({"noopener", "noreferrer"} <= set(a["rel"]) for a in links)
    assert not soup.select("[onerror], [onload]")


@pytest.mark.parametrize("url", [None, ""])
def test_absent_source_keeps_document_downloads(render_document, url):
    soup = render_document(url)
    assert len(soup.select('a[target="_blank"]')) == 2
    assert "Source URL" not in soup.get_text()


def test_benchmark_new_tab_link_is_isolated():
    # Inspect the actual anchor; rendering the benchmark form requires its
    # unrelated provider/model context.
    template_dir = Path(library.templates.env.loader.searchpath[0])
    soup = BeautifulSoup(
        (template_dir / "pages/benchmark.html").read_text(), "html.parser"
    )
    link = soup.find("a", href="/", target="_blank")
    assert {"noopener", "noreferrer"} <= set(link["rel"])