~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.
Python academiaanthropicarxivbravedeep-researchencryptionhome-automationhomeserverlocallocal-deep-researchlocal-llmmistralollamaopenaipubmedresearchresearch-toolretrieval-augmented-generationsearxngself-hosted
"""Formatter failures must not hide synthesis errors from the worker."""
from contextlib import ExitStack
from unittest.mock import MagicMock, patch
import pytest
from local_deep_research.advanced_search_system.findings.repository import (
FindingsRepository,
)
from local_deep_research.advanced_search_system.strategies.source_based_strategy import (
SourceBasedSearchStrategy,
)
from local_deep_research.database.models import ResearchHistory
from local_deep_research.error_handling.report_generator import (
ErrorReportGenerator,
)
from local_deep_research.storage.database import DatabaseReportStorage
from tests.web.services.helpers import (
MODULE,
_base_run_patches,
_egress_and_search_patches,
_get_raw_run_research_process,
)
@pytest.mark.parametrize("formatting_fails", [False, True])
@pytest.mark.parametrize("outcome", ["answer", "partial", "all_errors"])
def test_worker_saves_safe_content_after_formatting(
report_db, formatting_fails, outcome
):
marker = "/srv/private/provider-internals.txt"
answer = "A complete and useful synthesized answer."
partial = "Useful safe partial finding."
synthesized = (
answer
if outcome == "answer"
else f"Error: Failed to synthesize final answer. LLM error: {marker}"
)
findings = []
if outcome == "partial":
findings.append(
{"phase": "Search", "content": partial, "search_results": []}
)
findings.append(
{
"phase": "Final synthesis",
"content": synthesized,
"search_results": [],
}
)
repository = FindingsRepository(MagicMock())
# Exercise the real formatter's failure path with a malformed question
# group. Mocking formatted_findings missed the old prefix-changing wrapper.
repository.set_questions_by_iteration(
{1: None} if formatting_fails else {1: ["test question"]}
)
formatted = repository.format_findings_to_text(findings, synthesized)
system = MagicMock()
system.analyze_topic.return_value = {
"findings": findings,
"iterations": 1,
"formatted_findings": formatted,
"current_knowledge": synthesized,
}
system.all_links_of_system = []
patches = _base_run_patches()
patches[f"{MODULE}.get_llm"] = MagicMock(return_value=MagicMock())
patches[f"{MODULE}.AdvancedSearchSystem"] = MagicMock(return_value=system)
patches[f"{MODULE}.ErrorReportGenerator"] = ErrorReportGenerator
# Keep the real worker, error renderer, citation formatter and storage.
# Read the committed row so this checks the report actually saved for users.
with ExitStack() as stack:
for cm in _egress_and_search_patches():
stack.enter_context(cm)
for name, value in patches.items():
stack.enter_context(patch(name, value))
stack.enter_context(
patch(
"local_deep_research.storage.get_report_storage",
return_value=DatabaseReportStorage(report_db),
)
)
_get_raw_run_research_process()(
"report-errors",
"test query",
"quick",
username="user1",
search_engine="searxng",
settings_snapshot={
"llm.provider": "ollama",
"llm.model": "m",
"search.tool": "searxng",
},
)
report_db.expire_all()
report = report_db.get(ResearchHistory, "report-errors").report_content
assert report
assert marker not in report
if outcome == "answer":
assert report == answer
elif outcome == "partial":
assert partial in report
assert "Fallback Mode" in report
else:
assert "Research Failed" in report
assert "What happened:" in report
@pytest.mark.parametrize(
"failure,expected_hint,expected_category",
[
(
"Connection refused [Errno 111]",
"Cannot connect to the LLM service",
"Connection Issue",
),
(
"No auth credentials found",
"API key is missing",
"LLM Service Error",
),
(
"database is locked",
"database is temporarily locked",
"File System Error",
),
],
)
def test_all_error_results_keep_safe_category_and_hint(
report_db, failure, expected_hint, expected_category
):
marker = "/srv/private/provider-internals.txt"
strategy = SourceBasedSearchStrategy(
search=MagicMock(),
model=MagicMock(),
citation_handler=MagicMock(),
use_cross_engine_filter=False,
include_text_content=False,
max_iterations=1,
questions_per_iteration=1,
settings_snapshot={"app.max_user_query_length": 300},
)
# Let the real strategy create its error result from a failed LLM call.
with patch(
"local_deep_research.advanced_search_system.questions.standard_question.invoke_llm_sync",
side_effect=RuntimeError(f"{failure}: {marker}"),
):
results = strategy.analyze_topic("test query")
assert results["findings"][0]["phase"] == "Error"
assert failure in results["formatted_findings"]
system = MagicMock()
system.analyze_topic.return_value = results
system.all_links_of_system = []
patches = _base_run_patches()
patches[f"{MODULE}.get_llm"] = MagicMock(return_value=MagicMock())
patches[f"{MODULE}.AdvancedSearchSystem"] = MagicMock(return_value=system)
patches[f"{MODULE}.ErrorReportGenerator"] = ErrorReportGenerator
with ExitStack() as stack:
for cm in _egress_and_search_patches():
stack.enter_context(cm)
for name, value in patches.items():
stack.enter_context(patch(name, value))
stack.enter_context(
patch(
"local_deep_research.storage.get_report_storage",
return_value=DatabaseReportStorage(report_db),
)
)
_get_raw_run_research_process()(
"report-errors",
"test query",
"quick",
username="user1",
search_engine="searxng",
settings_snapshot={
"llm.provider": "ollama",
"llm.model": "m",
"search.tool": "searxng",
},
)
report_db.expire_all()
report = report_db.get(ResearchHistory, "report-errors").report_content
assert expected_hint in report
assert f"**Error Type:** {expected_category}" in report
assert marker not in report