~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
ci+test: retry transient network installs, fix patch.object race (#4302)
LearningCircuite1bc52904e2b
diff --git a/.github/workflows/playwright-webkit-tests.yml b/.github/workflows/playwright-webkit-tests.ymlindex 440add070f..2bc9fdc2e8 100644--- a/.github/workflows/playwright-webkit-tests.yml+++ b/.github/workflows/playwright-webkit-tests.yml@@ -50,12 +50,28 @@ jobs:npm cinpm run build- - name: Install Playwright dependencies+ - name: Install Playwright dependencies (npm)working-directory: tests/ui_tests/playwright- run: |- npm ci+ run: npm ci++ # ``playwright install --with-deps`` downloads ~150 MB of browser+ # binaries from the playwright CDN per attempt. Single CDN+ # hiccups have failed the whole release pipeline in the past+ # (e.g. transient 5xx mid-download); a single retry costs little+ # and removes the dominant transient-network failure mode for+ # this workflow. Two attempts only — three attempts of a 150 MB+ # download starts hiding real outages.+ - name: Install Playwright browsers+ uses: nick-fields/retry@ad984534de44a9489a53aefd81eb77f87c70dc60 # v4.0.0+ with:+ timeout_minutes: 10+ max_attempts: 2+ retry_on: error+ shell: bash# Install chromium for auth setup step + webkit for Safari tests- npx playwright install chromium webkit --with-deps+ command: |+ cd tests/ui_tests/playwright+ npx playwright install chromium webkit --with-deps- name: Setup test directoriesrun: |@@ -190,12 +206,28 @@ jobs:npm cinpm run build- - name: Install Playwright dependencies+ - name: Install Playwright dependencies (npm)working-directory: tests/ui_tests/playwright- run: |- npm ci+ run: npm ci++ # ``playwright install --with-deps`` downloads ~150 MB of browser+ # binaries from the playwright CDN per attempt. Single CDN+ # hiccups have failed the whole release pipeline in the past+ # (e.g. transient 5xx mid-download); a single retry costs little+ # and removes the dominant transient-network failure mode for+ # this workflow. Two attempts only — three attempts of a 150 MB+ # download starts hiding real outages.+ - name: Install Playwright browsers+ uses: nick-fields/retry@ad984534de44a9489a53aefd81eb77f87c70dc60 # v4.0.0+ with:+ timeout_minutes: 10+ max_attempts: 2+ retry_on: error+ shell: bash# Install chromium for auth setup step + webkit for Safari tests- npx playwright install chromium webkit --with-deps+ command: |+ cd tests/ui_tests/playwright+ npx playwright install chromium webkit --with-deps- name: Setup test directoriesrun: |diff --git a/.github/workflows/pre-commit.yml b/.github/workflows/pre-commit.ymlindex c98766046d..984784d34f 100644--- a/.github/workflows/pre-commit.yml+++ b/.github/workflows/pre-commit.yml@@ -56,7 +56,37 @@ jobs:sudo apt-get updatesudo apt-get install -y libsqlcipher-dev+ # Replicate pre-commit/[email protected] manually so we can wrap the+ # actual `pre-commit run` invocation in nick-fields/retry. The+ # upstream action is a thin composite (set PY for cache key →+ # cache `~/.cache/pre-commit` → `pip install pre-commit` →+ # `pre-commit run`), and the failure that motivates the retry+ # always happens in that last step: pre-commit lazily downloads+ # hook environments from external sources (PyPI for ruff, GitHub+ # release binaries for shellcheck, etc.), and a single HTTP 5xx+ # from any one of them fails the whole job. See run #2524 where+ # `Building wheel for shellcheck_py` hit HTTP 502 fetching the+ # shellcheck binary during the wheel build. A second attempt+ # benefits from the partially-populated cache and almost always+ # succeeds. Two attempts is enough: a hook environment that+ # fails to install twice in a row is not a transient outage.+ - name: Compute pre-commit cache key+ run: echo "PY=$(python -VV | sha256sum | cut -d' ' -f1)" >> "$GITHUB_ENV"++ - name: Cache pre-commit hook environments+ uses: actions/cache@27d5ce7f107fe9357f9df03efb73ab90386fccae # v5.0.5+ with:+ path: ~/.cache/pre-commit+ key: pre-commit|${{ env.PY }}|${{ hashFiles('.pre-commit-config.yaml') }}++ - name: Install pre-commit+ run: python -m pip install pre-commit+- name: Run pre-commit- uses: pre-commit/action@2c7b3805fd2a0fd8c1884dcaebf91fc102a13ecd # v3.0.1+ uses: nick-fields/retry@ad984534de44a9489a53aefd81eb77f87c70dc60 # v4.0.0with:- extra_args: --all-files+ timeout_minutes: 15+ max_attempts: 2+ retry_on: error+ shell: bash+ command: pre-commit run --show-diff-on-failure --color=always --all-filesdiff --git a/tests/research_library/services/test_library_rag_service_coverage.py b/tests/research_library/services/test_library_rag_service_coverage.pyindex f5bfcd81c3..e90f37138f 100644--- a/tests/research_library/services/test_library_rag_service_coverage.py+++ b/tests/research_library/services/test_library_rag_service_coverage.py@@ -1467,26 +1467,46 @@ class TestMergeAndPersistLocked:svc.faiss_index.docstore = MagicMock()svc.faiss_index.docstore._dict = {}start_event.wait() # Race start- with patch.object(- mod.FAISS,- "load_local",- side_effect=lambda *a, **kw: make_index_proxy(),- ):- svc._merge_and_persist_locked(- index_path,- [LangchainDocument(page_content=f"chunk-{cid}")],- [cid],- force_reindex=False,- )+ svc._merge_and_persist_locked(+ index_path,+ [LangchainDocument(page_content=f"chunk-{cid}")],+ [cid],+ force_reindex=False,+ )go = threading.Event()t1 = threading.Thread(target=worker, args=("A", go))t2 = threading.Thread(target=worker, args=("B", go))- t1.start()- t2.start()- go.set()- t1.join(timeout=5.0)- t2.join(timeout=5.0)++ # Patch FAISS.load_local ONCE, around both threads. Patching+ # inside each worker (the original shape) is a hidden race:+ # ``patch.object`` rewrites a module attribute on the FAISS+ # class and is not thread-safe — when the first worker's+ # ``with patch.object(...)`` block exits, it restores whatever+ # it captured as "the original" at entry, which can be the+ # second worker's lambda OR the real ``load_local`` (depending+ # on which thread entered patch first). If the real+ # ``load_local`` becomes active mid-test, the second worker's+ # reload step raises on the empty touched ``shared.faiss``,+ # the production code's ``except`` branch falls back to the+ # caller's stale in-memory MagicMock, and the worker's save is+ # a no-op (the setup MagicMock has no ``save_local`` side+ # effect). Result: that worker's chunk never reaches+ # ``disk_state``, the assertion fails, and the test wrongly+ # blames a regression in the production lock. Applying the+ # patch around both threads removes the race entirely — the+ # patched value is the only value visible while either worker+ # runs.+ with patch.object(+ mod.FAISS,+ "load_local",+ side_effect=lambda *a, **kw: make_index_proxy(),+ ):+ t1.start()+ t2.start()+ go.set()+ t1.join(timeout=5.0)+ t2.join(timeout=5.0)# Both writers' chunks present — read-modify-write under the# lock did its job. Without the fix, only one of {A, B} would