~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
test(close_base_llm): scale FD-growth test for Py 3.14 asyncio drift (#4224)
LearningCircuitc88c1ffe864d
diff --git a/tests/utilities/test_close_base_llm.py b/tests/utilities/test_close_base_llm.pyindex 40158e5a9a..a0ba1122cf 100644--- a/tests/utilities/test_close_base_llm.py+++ b/tests/utilities/test_close_base_llm.py@@ -368,8 +368,21 @@ class TestCloseBaseLLMRealHttpxAsync:def test_no_fd_growth_across_repeated_close_cycles(self):# Unit-level analogue of the PR's manual `lsof | grep -c eventpoll`# smoke. A real per-cycle leak (~1 FD/iter as observed in #3816)- # would push the delta well past the +2 slack we leave for ambient- # logging/pytest churn.+ # would push the delta well past the +8 slack.+ #+ # Sizing rationale (#4214 + #3816 follow-up):+ # Each cycle invokes ``asyncio.run`` twice — once to construct the+ # client in a (then-closed) loop A, once inside ``_close_base_llm``+ # to spin loop B for ``aclose()``. On Py 3.14 these transient loops+ # leave ambient FDs from selector/self-pipe lazy release. CI on the+ # 5-iter version of this test showed a deterministic +3 delta (58+ # of 58 recent failures), i.e. ~0.6 FDs/cycle steady-state — not+ # absorbable by a single warmup since it's per-cycle, not init-only.+ # At 20 iterations the expected ceiling is ~0.6 × 20 = ~12 in the+ # worst case (observed ~3–4 locally); the +8 slack covers the+ # typical range without masking a real per-iteration leak, which+ # would land at ≥+20 (1 FD/iter, #3816 shape) + ambient → far+ # above the threshold.async def _make():return httpx.AsyncClient()@@ -387,22 +400,20 @@ class TestCloseBaseLLMRealHttpxAsync:del llm, async_ollama, async_httpx# Warmup: one cycle before measuring absorbs one-time init drift- # (lazy imports, logging handlers, asyncio internals first-init)- # that would otherwise eat the +2 steady-state slack. A real- # per-iteration leak still shows as ≥+5 across the 5 measured iters.+ # (lazy imports, logging handlers, asyncio internals first-init)._one_cycle()gc.collect()before = _open_fd_count()- for _ in range(5):+ for _ in range(20):_one_cycle()gc.collect()gc.collect()after = _open_fd_count()- assert after - before <= 2, (+ assert after - before <= 8, (f"FD count climbed across close cycles: "f"before={before}, after={after}")