~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
"""Run the real ASGI app with only model and search egress stubbed for UI CI.
The browser test creates ``release/<research_id>`` after it has subscribed or
enabled polling. The model writes ``entered/<research_id>`` and waits, so the
worker cannot finish before the browser has observed the transport state.
This module is executed only by the release browser shard, never by ldr-web.
"""
from __future__ import annotations
import os
import time
import uuid
from pathlib import Path
import uvicorn
from langchain_core.language_models import BaseChatModel
from langchain_core.messages import AIMessage
from langchain_core.outputs import ChatGeneration, ChatResult
from local_deep_research.llm import register_llm
from local_deep_research.utilities.thread_context import get_search_context
from local_deep_research.web.fastapi_app import app
from local_deep_research.web_search_engines.engines.search_engine_wikipedia import (
WikipediaSearchEngine,
)
SMOKE_DIR = Path(os.environ["LDR_RESEARCH_SMOKE_DIR"])
ENTERED = SMOKE_DIR / "entered"
RELEASE = SMOKE_DIR / "release"
ANSWER = "RELEASEBROWSERSYNTHESISMARKER"
SOURCE_URL = "https://research-smoke.invalid/source"
class _GatedChatModel(BaseChatModel):
@property
def _llm_type(self) -> str:
return "release-browser-stub"
def _generate(self, messages, stop=None, run_manager=None, **kwargs):
context = get_search_context() or {}
research_id = str(uuid.UUID(str(context["research_id"])))
(ENTERED / research_id).touch()
deadline = time.monotonic() + 90
while not (RELEASE / research_id).exists():
if time.monotonic() >= deadline:
raise TimeoutError(
f"browser did not release research {research_id}"
)
time.sleep(0.05) # allow: unmarked-sleep -- test-only gate
return ChatResult(
generations=[ChatGeneration(message=AIMessage(content=ANSWER))]
)
def _make_llm(model_name=None, temperature=None, settings_snapshot=None):
return _GatedChatModel()
def _previews(self, query):
return [
{
"id": "release-browser-source",
"title": "Release Browser Source",
"link": SOURCE_URL,
"snippet": "RELEASEBROWSERSEARCHMARKER",
}
]
def _full_content(self, relevant_items):
return [
dict(
item,
content="RELEASEBROWSERSEARCHMARKER body",
full_content="RELEASEBROWSERSEARCHMARKER body",
)
for item in relevant_items
]
def main() -> None:
ENTERED.mkdir(parents=True, exist_ok=True)
RELEASE.mkdir(parents=True, exist_ok=True)
# fastapi_app has already imported llm_config and discovered built-ins.
# Override only the test server's Ollama boundary; the browser still
# selects the normal Ollama provider and submits through production code.
register_llm("ollama", _make_llm)
WikipediaSearchEngine._get_previews = _previews
WikipediaSearchEngine._get_full_content = _full_content
uvicorn.run(
app,
host="0.0.0.0",
port=5000,
workers=1,
server_header=False,
access_log=False,
# Mirrors MAX_INBOUND_MESSAGE_BYTES in web/services/socketio_asgi.py.
ws_max_size=65536,
)
if __name__ == "__main__":
main()