~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
"""Cover isolated NoteAIService branches without databases or model backends."""
import asyncio
from contextlib import contextmanager
from types import SimpleNamespace
from unittest.mock import AsyncMock, MagicMock
import pytest
from local_deep_research.research_library.notes.services.note_ai_service import (
NoteAIService,
)
def _llm_returning(payload: str) -> MagicMock:
# The AI methods are async (#6232) and reach the model through
# ``ainvoke``; ``invoke`` stays wired for the sync change-summary twin.
llm = MagicMock()
response = SimpleNamespace(content=payload)
llm.invoke.return_value = response
llm.ainvoke = AsyncMock(return_value=response)
return llm
def test_get_llm_temperature_override_is_not_cached(monkeypatch):
service = NoteAIService("alice")
first = MagicMock(name="first")
second = MagicMock(name="second")
get_llm = MagicMock(side_effect=[first, second])
monkeypatch.setattr(
"local_deep_research.config.llm_config.get_llm", get_llm
)
monkeypatch.setattr(service, "_build_settings_snapshot", lambda: {"x": 1})
assert service._get_llm(temperature=0) is first
assert service._get_llm(temperature=0) is second
assert service._llm is None
@pytest.mark.parametrize("value", [None, "not-a-number", []])
def test_coerce_confidence_rejects_unusable_values(value):
assert NoteAIService._coerce_confidence(value) == 0
def test_common_prefix_scans_complete_equal_block():
# The divergence must sit INSIDE the second 64 KiB block, not on the
# block boundary itself (e.g. `shared + "a"` vs `shared + "b"`, which
# diverges at the very first byte of block two): at the boundary, a
# pure per-character scan and a block-then-char scan agree by
# construction, so deleting the block loop or the char scan entirely
# still returns the right answer. Fifty matching characters into the
# second block forces both stages to actually run.
shared = "x" * (1 << 16)
a = shared + "y" * 50 + "p"
b = shared + "y" * 50 + "q"
assert NoteAIService._common_prefix_len(a, b) == len(shared) + 50
def test_fetch_and_project_note_edge_branches(monkeypatch):
session = MagicMock()
@contextmanager
def fake_session(_username):
yield session
monkeypatch.setattr(
"local_deep_research.research_library.notes.services.note_ai_service.get_user_db_session",
fake_session,
)
service = NoteAIService("alice")
monkeypatch.setattr(
service, "_get_note_source_type_id", lambda _session: None
)
assert service._fetch_document("missing") is None
document = SimpleNamespace(
id="n1", title="Title", text_content="Body", tags=None
)
monkeypatch.setattr(service, "_fetch_document", lambda _note_id: document)
assert service._get_note_content("n1") == "Body"
assert service._get_note("n1") == {
"id": "n1",
"title": "Title",
"content": "Body",
"tags": [],
}
def test_simple_ai_methods_cover_empty_and_failure_paths(monkeypatch):
service = NoteAIService("alice")
monkeypatch.setattr(service, "_get_note_content", lambda _note_id: None)
assert asyncio.run(service.extract_research_questions("missing")) == []
assert asyncio.run(service.summarize_changes("", "new")) == "Note created"
assert asyncio.run(service.summarize_changes("old", "")) == "Changes made"
monkeypatch.setattr(service, "_get_note_content", lambda _note_id: "body")
monkeypatch.setattr(
service,
"semantic_search",
MagicMock(side_effect=RuntimeError("search unavailable")),
)
with pytest.raises(RuntimeError, match="search unavailable"):
service.find_similar_notes("n1")
def test_grade_rejects_non_list_source_refs():
# A scalar (unwrapped) source_refs must be discarded rather than
# iterated. `source_refs: "0"` doesn't distinguish this: iterating a
# string yields characters, which the `isinstance(r, int)` filter
# already drops, so the anti-rubber-stamp downgrade fires either way
# and masks whether the non-list guard did anything. An unwrapped int
# is unmasked: without the guard, `for r in 1` raises TypeError instead
# of degrading to the same "unverified" verdict.
service = NoteAIService("alice")
service._get_llm = lambda temperature=None: _llm_returning(
'[{"verdict":"supported","confidence":80,'
'"reasoning":"r","source_refs":1}]'
)
verdict = asyncio.run(
service.grade_all_claims(["claim"], "report", ["source label"])
)[0]
assert verdict["verdict"] == "unverified"
assert verdict["confidence"] == 0
assert verdict["sources"] == []
def test_grade_positional_fallback_for_unlabelled_item():
# The verdict object carries no "index" key and by_index is empty, so
# only the positional fallback (parsed[i] matched to claim i by
# position) can supply it. The citation is valid and non-empty, so the
# anti-rubber-stamp downgrade never fires — removing the fallback
# leaves `item == {}`, which flips verdict/confidence/sources rather
# than converging on the same downgraded result.
service = NoteAIService("alice")
service._get_llm = lambda temperature=None: _llm_returning(
'[{"verdict":"supported","confidence":90,'
'"reasoning":"ok","source_refs":[0]}]'
)
verdict = asyncio.run(
service.grade_all_claims(["claim"], "report", [{"title": "Source A"}])
)[0]
assert verdict["verdict"] == "supported"
assert verdict["confidence"] == 90
assert verdict["sources"] == [{"title": "Source A", "url": ""}]
def test_grade_skips_non_dict_shown_source():
# Two refs: one resolves to a dict source (kept, so the anti-rubber-
# stamp downgrade never triggers) and one to a non-dict source, which
# `isinstance(shown_sources[r], dict)` must filter out. Without that
# guard the code calls `.get()` on a plain string and raises
# AttributeError instead of silently producing the same citation list.
service = NoteAIService("alice")
service._get_llm = lambda temperature=None: _llm_returning(
'[{"verdict":"supported","confidence":70,'
'"reasoning":"ok","source_refs":[0,1]}]'
)
verdict = asyncio.run(
service.grade_all_claims(
["claim"], "report", [{"title": "Source A"}, "plain label"]
)
)[0]
assert verdict["verdict"] == "supported"
assert verdict["sources"] == [{"title": "Source A", "url": ""}]
def test_similar_passages_empty_engine_result_and_failure(monkeypatch):
from local_deep_research.research_library.notes.services.note_service import (
NoteService,
)
@contextmanager
def fake_session(_username):
yield MagicMock()
class EmptyEngine:
def __init__(self, **_kwargs):
pass
def search(self, _query, *, limit):
return []
monkeypatch.setattr(
"local_deep_research.research_library.notes.services.note_ai_service.get_user_db_session",
fake_session,
)
monkeypatch.setattr(
NoteService,
"_get_or_create_notes_collection",
lambda _self, _session: "notes",
)
monkeypatch.setattr(
"local_deep_research.web_search_engines.engines.search_engine_collection.CollectionSearchEngine",
EmptyEngine,
)
service = NoteAIService("alice")
assert service.find_similar_passages("query") == []
monkeypatch.setattr(
EmptyEngine, "search", MagicMock(side_effect=RuntimeError("index down"))
)
with pytest.raises(RuntimeError, match="index down"):
service.find_similar_passages("query")
def test_suggest_links_missing_success_and_failure(monkeypatch):
service = NoteAIService("alice")
monkeypatch.setattr(service, "_get_note", lambda _note_id: None)
assert service.suggest_links("missing") == []
monkeypatch.setattr(
service,
"_get_note",
lambda _note_id: {"title": "T", "content": "C"},
)
monkeypatch.setattr(
service,
"semantic_search",
lambda *_args, **_kwargs: [{"id": "n1"}, {"id": "n2"}],
)
assert service.suggest_links("n1") == [{"id": "n2"}]
monkeypatch.setattr(
service,
"semantic_search",
MagicMock(side_effect=RuntimeError("index down")),
)
with pytest.raises(RuntimeError, match="index down"):
service.suggest_links("n1")
def test_semantic_diff_reraises_model_failure():
service = NoteAIService("alice")
service._llm = MagicMock()
service._llm.ainvoke = AsyncMock(side_effect=RuntimeError("model down"))
with pytest.raises(RuntimeError, match="model down"):
asyncio.run(service.semantic_diff("old", "new"))
def test_summarize_synthesis_selects_summary_prompt_and_title(monkeypatch):
documents = [
SimpleNamespace(
id="a", title="Alpha", text_content="alpha source content"
),
SimpleNamespace(
id="b", title="Beta", text_content="beta source content"
),
]
session = MagicMock()
session.query.return_value.filter.return_value.all.return_value = documents
@contextmanager
def fake_session(_username):
yield session
monkeypatch.setattr(
"local_deep_research.research_library.notes.services.note_ai_service.get_user_db_session",
fake_session,
)
service = NoteAIService("alice")
monkeypatch.setattr(
service, "_get_note_source_type_id", lambda _session: "note"
)
llm = _llm_returning("combined")
service._llm = llm
result = asyncio.run(service.synthesize_notes(["a", "b"], "summarize"))
assert result["success"] is True
assert result["content"] == "combined"
assert result["suggested_title"] == "Summary: Alpha + 1 more"
llm.ainvoke.assert_called_once()
prompt = llm.ainvoke.call_args.args[0]
assert prompt.startswith("Summarize the key points")
assert "alpha source content" in prompt
assert "beta source content" in prompt