~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
"""Gate output and linear work; measure the real raw-context construction."""
import pytest
from local_deep_research.chat.context import ChatContextManager
def history(count, kind, steps=False):
messages, findings = [], []
for number in range(count):
prefix = f"Finding {number}: "
if kind == "short":
body = prefix + "Evidence. " * 8
expected = body
else:
marker = (
"[[7]](https://example.test/paper)"
if kind == "citations"
else ""
)
body = prefix + marker + "x" * 450 + "\n\n" + "Support. " * 60
expected = prefix + "x" * 450 + "..."
messages.extend(
[
dict(role="user", content=f"User [9] {number}"),
dict(
role="assistant",
research_id=f"research-{number}",
content=body,
),
]
)
if steps:
messages.append(
dict(
message_type="step",
role="assistant",
research_id="progress",
content="Excluded progress",
)
)
findings.append(expected)
return messages, findings
@pytest.mark.parametrize("kind", ["short", "long", "citations"])
@pytest.mark.parametrize("steps", [False, True], ids=["turns", "with_steps"])
def test_history_growth_preserves_recent_findings_and_bounds_work(
chat_workload, kind, steps
):
for size in (
chat_workload.settings["size"],
chat_workload.settings["size"] * 2,
):
messages, findings = history(size, kind, steps)
expected = "\n\n---\n\n".join(findings[-5:])
def prepare():
manager = ChatContextManager("performance-session", messages)
manager.MAX_FINDINGS_TO_INCLUDE = 5
actual = manager._strip_citation_markers
calls = []
def counted(text):
calls.append(len(text))
return actual(text)
manager._strip_citation_markers = counted
return manager, calls
def verify(state, output):
manager, calls = state
try:
assert output == {
"session_id": "performance-session",
"original_query": "User [9] 0",
"accumulated_findings": expected,
"past_findings": expected,
"key_entities": [],
"topics": [],
"is_multi_turn": True,
"turn_count": size * 2,
}
# Allows future window-first processing while rejecting repeated
# processing beyond one pass through qualifying messages.
assert len(calls) <= size
return dict(
history_findings=size,
retained_findings=min(5, size),
stripping_calls=len(calls),
stripped_input_chars=sum(calls),
output_chars=len(expected),
)
finally:
del manager._strip_citation_markers
chat_workload.measure(
f"raw_{kind}_{'steps' if steps else 'turns'}",
size,
prepare,
lambda state: state[0].build_research_context(),
verify,
units=size,
unit="history_finding",
parameters=dict(
kind=kind,
retained_findings=5,
progress_steps=steps,
instrumented_work_count=True,
constructor_timed=False,
),
)
@pytest.mark.parametrize("limit", [0, -1, 1, 5, 20])
def test_findings_window_boundaries_keep_existing_selection_semantics(limit):
messages, expected = history(8, "long", steps=True)
manager = ChatContextManager("boundaries", messages)
manager.MAX_FINDINGS_TO_INCLUDE = limit
assert manager._extract_findings_from_history() == "\n\n---\n\n".join(
expected[-limit:]
)
def test_changed_latest_response_is_observed_on_next_build():
messages, _ = history(8, "short")
manager = ChatContextManager("freshness", messages)
first = manager.build_research_context()["past_findings"]
manager.messages[-1]["content"] = "Updated latest finding"
second = manager.build_research_context()["past_findings"]
assert first != second and second.endswith("Updated latest finding")