A few minor test updates to cover some edge case errors

.github/workflows/ci-docker.yml

@@ -35,7 +35,7 @@ jobs:
       contents: read
       packages: write
     outputs:
-      should-push: ${{ steps.check-push.outputs.should-push }}
+      can-push: ${{ steps.check-push.outputs.can-push }}
       push-external: ${{ steps.check-push.outputs.push-external }}
       repository: ${{ steps.repo.outputs.name }}
       ref-name: ${{ steps.ref.outputs.name }}
@@ -59,28 +59,16 @@ jobs:
         env:
           REF_NAME: ${{ steps.ref.outputs.name }}
         run: |
-          # should-push: Should we push to GHCR?
-          # True for:
-          #   1. Pushes (tags/dev/beta) - filtered via the workflow triggers
-          #   2. Internal PRs where the branch name starts with 'feature-' - filtered here when a PR is synced
-
-          should_push="false"
-
-          if [[ "${{ github.event_name }}" == "push" ]]; then
-            should_push="true"
-          elif [[ "${{ github.event_name }}" == "pull_request" && "${{ github.event.pull_request.head.repo.full_name }}" == "${{ github.repository }}" ]]; then
-            if [[ "${REF_NAME}" == feature-* || "${REF_NAME}" == fix-* ]]; then
-              should_push="true"
-            fi
-          fi
-
-          echo "should-push=${should_push}"
-          echo "should-push=${should_push}" >> $GITHUB_OUTPUT
+          # can-push: Can we push to GHCR?
+          # True for: pushes, or PRs from the same repo (not forks)
+          can_push=${{ github.event_name == 'push' || github.event.pull_request.head.repo.full_name == github.repository }}
+          echo "can-push=${can_push}"
+          echo "can-push=${can_push}" >> $GITHUB_OUTPUT
 
           # push-external: Should we also push to Docker Hub and Quay.io?
           # Only for main repo on dev/beta branches or version tags
           push_external="false"
-          if [[ "${should_push}" == "true" && "${{ github.repository_owner }}" == "paperless-ngx" ]]; then
+          if [[ "${can_push}" == "true" && "${{ github.repository_owner }}" == "paperless-ngx" ]]; then
             case "${REF_NAME}" in
               dev|beta)
                 push_external="true"
@@ -137,20 +125,20 @@ jobs:
           labels: ${{ steps.docker-meta.outputs.labels }}
           build-args: |
             PNGX_TAG_VERSION=${{ steps.docker-meta.outputs.version }}
-          outputs: type=image,name=${{ env.REGISTRY }}/${{ steps.repo.outputs.name }},push-by-digest=true,name-canonical=true,push=${{ steps.check-push.outputs.should-push }}
+          outputs: type=image,name=${{ env.REGISTRY }}/${{ steps.repo.outputs.name }},push-by-digest=true,name-canonical=true,push=${{ steps.check-push.outputs.can-push }}
           cache-from: |
             type=registry,ref=${{ env.REGISTRY }}/${{ steps.repo.outputs.name }}/cache/app:${{ steps.ref.outputs.cache-ref }}-${{ matrix.arch }}
             type=registry,ref=${{ env.REGISTRY }}/${{ steps.repo.outputs.name }}/cache/app:dev-${{ matrix.arch }}
-          cache-to: ${{ steps.check-push.outputs.should-push == 'true' && format('type=registry,mode=max,ref={0}/{1}/cache/app:{2}-{3}', env.REGISTRY, steps.repo.outputs.name, steps.ref.outputs.cache-ref, matrix.arch) || '' }}
+          cache-to: ${{ steps.check-push.outputs.can-push == 'true' && format('type=registry,mode=max,ref={0}/{1}/cache/app:{2}-{3}', env.REGISTRY, steps.repo.outputs.name, steps.ref.outputs.cache-ref, matrix.arch) || '' }}
       - name: Export digest
-        if: steps.check-push.outputs.should-push == 'true'
+        if: steps.check-push.outputs.can-push == 'true'
         run: |
           mkdir -p /tmp/digests
           digest="${{ steps.build.outputs.digest }}"
           echo "digest=${digest}"
           touch "/tmp/digests/${digest#sha256:}"
       - name: Upload digest
-        if: steps.check-push.outputs.should-push == 'true'
+        if: steps.check-push.outputs.can-push == 'true'
         uses: actions/upload-artifact@v6.0.0
         with:
           name: digests-${{ matrix.arch }}
@@ -161,7 +149,7 @@ jobs:
     name: Merge and Push Manifest
     runs-on: ubuntu-24.04
     needs: build-arch
-    if: needs.build-arch.outputs.should-push == 'true'
+    if: needs.build-arch.outputs.can-push == 'true'
     permissions:
       contents: read
       packages: write

.pre-commit-config.yaml

@@ -37,7 +37,7 @@ repos:
           - json
   # See https://github.com/prettier/prettier/issues/15742 for the fork reason
   - repo: https://github.com/rbubley/mirrors-prettier
-    rev: 'v3.8.0'
+    rev: 'v3.6.2'
     hooks:
       - id: prettier
         types_or:
@@ -49,7 +49,7 @@ repos:
           - 'prettier-plugin-organize-imports@4.1.0'
   # Python hooks
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.14.13
+    rev: v0.14.5
     hooks:
       - id: ruff-check
       - id: ruff-format
@@ -76,7 +76,7 @@ repos:
     hooks:
       - id: shellcheck
   - repo: https://github.com/google/yamlfmt
-    rev: v0.21.0
+    rev: v0.20.0
     hooks:
       - id: yamlfmt
         exclude: "^src-ui/pnpm-lock.yaml"

Dockerfile

@@ -30,7 +30,7 @@ RUN set -eux \
 # Purpose: Installs s6-overlay and rootfs
 # Comments:
 #  - Don't leave anything extra in here either
-FROM ghcr.io/astral-sh/uv:0.9.26-python3.12-trixie-slim AS s6-overlay-base
+FROM ghcr.io/astral-sh/uv:0.9.15-python3.12-trixie-slim AS s6-overlay-base
 
 WORKDIR /usr/src/s6
 
@@ -196,11 +196,7 @@ RUN set -eux \
     && apt-get install --yes --quiet --no-install-recommends ${BUILD_PACKAGES} \
   && echo "Installing Python requirements" \
     && uv export --quiet --no-dev --all-extras --format requirements-txt --output-file requirements.txt \
-    && uv pip install --no-cache --system --no-python-downloads --python-preference system \
-      --index https://pypi.org/simple \
-      --index https://download.pytorch.org/whl/cpu \
-      --index-strategy unsafe-best-match \
-      --requirements requirements.txt \
+    && uv pip install --no-cache --system --no-python-downloads --python-preference system --requirements requirements.txt \
   && echo "Installing NLTK data" \
     && python3 -W ignore::RuntimeWarning -m nltk.downloader -d "/usr/share/nltk_data" snowball_data \
     && python3 -W ignore::RuntimeWarning -m nltk.downloader -d "/usr/share/nltk_data" stopwords \

docs/advanced_usage.md

@@ -501,7 +501,7 @@ The `datetime` filter formats a datetime string or datetime object using Python'
 See the [strftime format code documentation](https://docs.python.org/3.13/library/datetime.html#strftime-and-strptime-format-codes)
 for the possible codes and their meanings.
 
-##### Date Localization
+##### Date Localization {#date-localization}
 
 The `localize_date` filter formats a date or datetime object into a localized string using Babel internationalization.
 This takes into account the provided locale for translation. Since this must be used on a date or datetime object,
@@ -851,8 +851,8 @@ followed by the even pages.
 
 It's important that the scan files get consumed in the correct order, and one at a time.
 You therefore need to make sure that Paperless is running while you upload the files into
-the directory; and if you're using [polling](configuration.md#polling), make sure that
-`CONSUMER_POLLING` is set to a value lower than it takes for the second scan to appear,
+the directory; and if you're using polling, make sure that
+`CONSUMER_POLLING_INTERVAL` is set to a value lower than it takes for the second scan to appear,
 like 5-10 or even lower.
 
 Another thing that might happen is that you start a double sided scan, but then forget

docs/configuration.md

@@ -1175,21 +1175,45 @@ don't exist yet.
 
 #### [`PAPERLESS_CONSUMER_IGNORE_PATTERNS=<json>`](#PAPERLESS_CONSUMER_IGNORE_PATTERNS) {#PAPERLESS_CONSUMER_IGNORE_PATTERNS}
 
-: By default, paperless ignores certain files and folders in the
-consumption directory, such as system files created by the Mac OS
-or hidden folders some tools use to store data.
+: Additional regex patterns for files to ignore in the consumption directory. Patterns are matched against filenames only (not full paths)
+using Python's `re.match()`, which anchors at the start of the filename.
 
-    This can be adjusted by configuring a custom json array with
-    patterns to exclude.
+    See the [watchfiles documentation](https://watchfiles.helpmanual.io/api/filters/#watchfiles.BaseFilter.ignore_entity_patterns)
 
-    For example, `.DS_STORE/*` will ignore any files found in a folder
-    named `.DS_STORE`, including `.DS_STORE/bar.pdf` and `foo/.DS_STORE/bar.pdf`
+    This setting is for additional patterns beyond the built-in defaults. Common system files and directories are already ignored automatically.
+    The patterns will be compiled via Python's standard `re` module.
 
-    A pattern like `._*` will ignore anything starting with `._`, including:
-    `._foo.pdf` and `._bar/foo.pdf`
+    Example custom patterns:
 
-    Defaults to
-    `[".DS_Store", ".DS_STORE", "._*", ".stfolder/*", ".stversions/*", ".localized/*", "desktop.ini", "@eaDir/*", "Thumbs.db"]`.
+    ```json
+    ["^temp_", "\\.bak$", "^~"]
+    ```
+
+    This would ignore:
+
+    - Files starting with `temp_` (e.g., `temp_scan.pdf`)
+    - Files ending with `.bak` (e.g., `document.pdf.bak`)
+    - Files starting with `~` (e.g., `~$document.docx`)
+
+    Defaults to `[]` (empty list, uses only built-in defaults).
+
+    The default ignores are `[.DS_Store, .DS_STORE, ._*, desktop.ini, Thumbs.db]` and cannot be overridden.
+
+#### [`PAPERLESS_CONSUMER_IGNORE_DIRS=<json>`](#PAPERLESS_CONSUMER_IGNORE_DIRS) {#PAPERLESS_CONSUMER_IGNORE_DIRS}
+
+: Additional directory names to ignore in the consumption directory. Directories matching these names (and all their contents) will be skipped.
+
+    This setting is for additional directories beyond the built-in defaults. Matching is done by directory name only, not full path.
+
+    Example:
+
+    ```json
+    ["temp", "incoming", ".hidden"]
+    ```
+
+    Defaults to `[]` (empty list, uses only built-in defaults).
+
+    The default ignores are `[.stfolder, .stversions, .localized, @eaDir, .Spotlight-V100, .Trashes, __MACOSX]` and cannot be overridden.
 
 #### [`PAPERLESS_CONSUMER_BARCODE_SCANNER=<string>`](#PAPERLESS_CONSUMER_BARCODE_SCANNER) {#PAPERLESS_CONSUMER_BARCODE_SCANNER}
 
@@ -1288,48 +1312,24 @@ within your documents.
 
     Defaults to false.
 
-### Polling {#polling}
+#### [`PAPERLESS_CONSUMER_POLLING_INTERVAL=<num>`](#PAPERLESS_CONSUMER_POLLING_INTERVAL) {#PAPERLESS_CONSUMER_POLLING_INTERVAL}
 
-#### [`PAPERLESS_CONSUMER_POLLING=<num>`](#PAPERLESS_CONSUMER_POLLING) {#PAPERLESS_CONSUMER_POLLING}
+: Configures how the consumer detects new files in the consumption directory.
 
-: If paperless won't find documents added to your consume folder, it
-might not be able to automatically detect filesystem changes. In
-that case, specify a polling interval in seconds here, which will
-then cause paperless to periodically check your consumption
-directory for changes. This will also disable listening for file
-system changes with `inotify`.
+    When set to `0` (default), paperless uses native filesystem notifications for efficient, immediate detection of new files.
 
-    Defaults to 0, which disables polling and uses filesystem
-    notifications.
+    When set to a positive number, paperless polls the consumption directory at that interval in seconds. Use polling for network filesystems (NFS, SMB/CIFS) where native notifications may not work reliably.
 
-#### [`PAPERLESS_CONSUMER_POLLING_RETRY_COUNT=<num>`](#PAPERLESS_CONSUMER_POLLING_RETRY_COUNT) {#PAPERLESS_CONSUMER_POLLING_RETRY_COUNT}
+    Defaults to 0.
 
-: If consumer polling is enabled, sets the maximum number of times
-paperless will check for a file to remain unmodified. If a file's
-modification time and size are identical for two consecutive checks, it
-will be consumed.
+#### [`PAPERLESS_CONSUMER_STABILITY_DELAY=<num>`](#PAPERLESS_CONSUMER_STABILITY_DELAY) {#PAPERLESS_CONSUMER_STABILITY_DELAY}
 
-    Defaults to 5.
+: Sets the time in seconds that a file must remain unchanged (same size and modification time) before paperless will begin consuming it.
 
-#### [`PAPERLESS_CONSUMER_POLLING_DELAY=<num>`](#PAPERLESS_CONSUMER_POLLING_DELAY) {#PAPERLESS_CONSUMER_POLLING_DELAY}
+    Increase this value if you experience issues with files being consumed before they are fully written, particularly on slower network storage or
+    with certain scanner quirks
 
-: If consumer polling is enabled, sets the delay in seconds between
-each check (above) paperless will do while waiting for a file to
-remain unmodified.
-
-    Defaults to 5.
-
-### iNotify {#inotify}
-
-#### [`PAPERLESS_CONSUMER_INOTIFY_DELAY=<num>`](#PAPERLESS_CONSUMER_INOTIFY_DELAY) {#PAPERLESS_CONSUMER_INOTIFY_DELAY}
-
-: Sets the time in seconds the consumer will wait for additional
-events from inotify before the consumer will consider a file ready
-and begin consumption. Certain scanners or network setups may
-generate multiple events for a single file, leading to multiple
-consumers working on the same file. Configure this to prevent that.
-
-    Defaults to 0.5 seconds.
+    Defaults to 5.0 seconds.
 
 ## Workflow webhooks
 
@@ -1873,7 +1873,7 @@ using the OpenAI API. This setting is required to be set to use the AI features.
 #### [`PAPERLESS_AI_LLM_MODEL=<str>`](#PAPERLESS_AI_LLM_MODEL) {#PAPERLESS_AI_LLM_MODEL}
 
 : The model to use for the AI backend, i.e. "gpt-3.5-turbo", "gpt-4" or any of the models supported by the
-current backend. If not supplied, defaults to "gpt-3.5-turbo" for OpenAI and "llama3.1" for Ollama.
+current backend. If not supplied, defaults to "gpt-3.5-turbo" for OpenAI and "llama3" for Ollama.
 
     Defaults to None.
 

docs/migration.md

@@ -0,0 +1,19 @@
+# v3 Migration Guide
+
+## Consumer Settings Changes
+
+The v3 consumer command uses a [different library](https://watchfiles.helpmanual.io/) to unify
+the watching for new files in the consume directory. For the user, this removes several configuration options related to delays and retries
+and replaces with a single unified setting. It also adjusts how the consumer ignore filtering happens, replaced `fnmatch` with `regex` and
+separating the directory ignore from the file ignore.
+
+### Summary
+
+| Old Setting                    | New Setting                                                                         | Notes                                                                                |
+| ------------------------------ | ----------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------ |
+| `CONSUMER_POLLING`             | [`CONSUMER_POLLING_INTERVAL`](configuration.md#PAPERLESS_CONSUMER_POLLING_INTERVAL) | Renamed for clarity                                                                  |
+| `CONSUMER_INOTIFY_DELAY`       | [`CONSUMER_STABILITY_DELAY`](configuration.md#PAPERLESS_CONSUMER_STABILITY_DELAY)   | Unified for all modes                                                                |
+| `CONSUMER_POLLING_DELAY`       | _Removed_                                                                           | Use `CONSUMER_STABILITY_DELAY`                                                       |
+| `CONSUMER_POLLING_RETRY_COUNT` | _Removed_                                                                           | Automatic with stability tracking                                                    |
+| `CONSUMER_IGNORE_PATTERNS`     | [`CONSUMER_IGNORE_PATTERNS`](configuration.md#PAPERLESS_CONSUMER_IGNORE_PATTERNS)   | **Now regex, not fnmatch**; user patterns are added to (not replacing) default ones  |
+| _New_                          | [`CONSUMER_IGNORE_DIRS`](configuration.md#PAPERLESS_CONSUMER_IGNORE_DIRS)           | Additional directories to ignore; user entries are added to (not replacing) defaults |

docs/setup.md

@@ -124,8 +124,7 @@ account. The script essentially automatically performs the steps described in [D
         system notifications with `inotify`. When storing the consumption
         directory on such a file system, paperless will not pick up new
         files with the default configuration. You will need to use
-        [`PAPERLESS_CONSUMER_POLLING`](configuration.md#PAPERLESS_CONSUMER_POLLING), which will disable inotify. See
-        [here](configuration.md#polling).
+        [`PAPERLESS_CONSUMER_POLLING_INTERVAL`](configuration.md#PAPERLESS_CONSUMER_POLLING_INTERVAL), which will disable inotify.
 
 5.  Run `docker compose pull`. This will pull the image from the GitHub container registry
     by default but you can change the image to pull from Docker Hub by changing the `image`

docs/troubleshooting.md

@@ -46,7 +46,7 @@ run:
 If you notice that the consumer will only pickup files in the
 consumption directory at startup, but won't find any other files added
 later, you will need to enable filesystem polling with the configuration
-option [`PAPERLESS_CONSUMER_POLLING`](configuration.md#PAPERLESS_CONSUMER_POLLING).
+option [`PAPERLESS_CONSUMER_POLLING_INTERVAL`](configuration.md#PAPERLESS_CONSUMER_POLLING_INTERVAL).
 
 This will disable listening to filesystem changes with inotify and
 paperless will manually check the consumption directory for changes
@@ -234,47 +234,9 @@ FileNotFoundError: [Errno 2] No such file or directory: '/tmp/ocrmypdf.io.yhk3zb
 
 This probably indicates paperless tried to consume the same file twice.
 This can happen for a number of reasons, depending on how documents are
-placed into the consume folder. If paperless is using inotify (the
-default) to check for documents, try adjusting the
-[inotify configuration](configuration.md#inotify). If polling is enabled, try adjusting the
-[polling configuration](configuration.md#polling).
-
-## Consumer fails waiting for file to remain unmodified.
-
-You might find messages like these in your log files:
-
-```
-[ERROR] [paperless.management.consumer] Timeout while waiting on file /usr/src/paperless/src/../consume/SCN_0001.pdf to remain unmodified.
-```
-
-This indicates paperless timed out while waiting for the file to be
-completely written to the consume folder. Adjusting
-[polling configuration](configuration.md#polling) values should resolve the issue.
-
-!!! note
-
-    The user will need to manually move the file out of the consume folder
-    and back in, for the initial failing file to be consumed.
-
-## Consumer fails reporting "OS reports file as busy still".
-
-You might find messages like these in your log files:
-
-```
-[WARNING] [paperless.management.consumer] Not consuming file /usr/src/paperless/src/../consume/SCN_0001.pdf: OS reports file as busy still
-```
-
-This indicates paperless was unable to open the file, as the OS reported
-the file as still being in use. To prevent a crash, paperless did not
-try to consume the file. If paperless is using inotify (the default) to
-check for documents, try adjusting the
-[inotify configuration](configuration.md#inotify). If polling is enabled, try adjusting the
-[polling configuration](configuration.md#polling).
-
-!!! note
-
-    The user will need to manually move the file out of the consume folder
-    and back in, for the initial failing file to be consumed.
+placed into the consume folder, such as how a scanner may modify a file multiple times as it scans.
+Try adjusting the
+[file stability delay](configuration.md#PAPERLESS_CONSUMER_STABILITY_DELAY) to a larger value.
 
 ## Log reports "Creating PaperlessTask failed".
 

docs/usage.md

@@ -565,7 +565,7 @@ This allows for complex logic to be used to generate the title, including [logic
 and [filters](https://jinja.palletsprojects.com/en/3.1.x/templates/#id11).
 The template is provided as a string.
 
-Using Jinja2 Templates is also useful for [Date localization](advanced_usage.md#Date-Localization) in the title.
+Using Jinja2 Templates is also useful for [Date localization](advanced_usage.md#date-localization) in the title.
 
 The available inputs differ depending on the type of workflow trigger.
 This is because at the time of consumption (when the text is to be set), no automatic tags etc. have been

mkdocs.yml

@@ -69,8 +69,9 @@ nav:
   - development.md
   - 'FAQs': faq.md
   - troubleshooting.md
+  - 'Migration to v3': migration.md
   - changelog.md
-copyright: Copyright © 2016 - 2023 Daniel Quinn, Jonas Winkler, and the Paperless-ngx team
+copyright: Copyright © 2016 - 2026 Daniel Quinn, Jonas Winkler, and the Paperless-ngx team
 extra:
   social:
     - icon: fontawesome/brands/github

paperless.conf.example

@@ -55,10 +55,10 @@
 #PAPERLESS_TASK_WORKERS=1
 #PAPERLESS_THREADS_PER_WORKER=1
 #PAPERLESS_TIME_ZONE=UTC
-#PAPERLESS_CONSUMER_POLLING=10
+#PAPERLESS_CONSUMER_POLLING_INTERVAL=10
 #PAPERLESS_CONSUMER_DELETE_DUPLICATES=false
 #PAPERLESS_CONSUMER_RECURSIVE=false
-#PAPERLESS_CONSUMER_IGNORE_PATTERNS=[".DS_STORE/*", "._*", ".stfolder/*", ".stversions/*", ".localized/*", "desktop.ini"]
+#PAPERLESS_CONSUMER_IGNORE_PATTERNS=[]  # Defaults are built in; add filename regexes, e.g. ["^\\.DS_Store$", "^desktop\\.ini$"]
 #PAPERLESS_CONSUMER_SUBDIRS_AS_TAGS=false
 #PAPERLESS_CONSUMER_ENABLE_BARCODES=false
 #PAPERLESS_CONSUMER_BARCODE_STRING=PATCHT

pyproject.toml

@@ -19,16 +19,16 @@ dependencies = [
   "azure-ai-documentintelligence>=1.0.2",
   "babel>=2.17",
   "bleach~=6.3.0",
-  "celery[redis]~=5.6.2",
+  "celery[redis]~=5.5.1",
   "channels~=4.2",
   "channels-redis~=4.2",
   "concurrent-log-handler~=0.9.25",
   "dateparser~=1.2",
   # WARNING: django does not use semver.
   #          Only patch versions are guaranteed to not introduce breaking changes.
-  "django==5.2.10",
-  "django-allauth[mfa,socialaccount]~=65.13.0",
-  "django-auditlog~=3.4.1",
+  "django~=5.2.5",
+  "django-allauth[mfa,socialaccount]~=65.12.1",
+  "django-auditlog~=3.3.0",
   "django-cachalot~=2.8.0",
   "django-celery-results~=2.6.0",
   "django-compression-middleware~=0.5.0",
@@ -47,20 +47,19 @@ dependencies = [
   "faiss-cpu>=1.10",
   "filelock~=3.20.0",
   "flower~=2.0.1",
-  "gotenberg-client~=0.13.1",
+  "gotenberg-client~=0.12.0",
   "httpx-oauth~=0.16",
   "imap-tools~=1.11.0",
-  "inotifyrecursive~=0.3",
   "jinja2~=3.1.5",
   "langdetect~=1.0.9",
-  "llama-index-core>=0.14.12",
-  "llama-index-embeddings-huggingface>=0.6.1",
-  "llama-index-embeddings-openai>=0.5.1",
-  "llama-index-llms-ollama>=0.9.1",
-  "llama-index-llms-openai>=0.6.13",
-  "llama-index-vector-stores-faiss>=0.5.2",
+  "llama-index-core>=0.12.33.post1",
+  "llama-index-embeddings-huggingface>=0.5.3",
+  "llama-index-embeddings-openai>=0.3.1",
+  "llama-index-llms-ollama>=0.5.4",
+  "llama-index-llms-openai>=0.3.38",
+  "llama-index-vector-stores-faiss>=0.3",
   "nltk~=3.9.1",
-  "ocrmypdf~=16.13.0",
+  "ocrmypdf~=16.12.0",
   "openai>=1.76",
   "pathvalidate~=3.3.1",
   "pdf2image~=1.17.0",
@@ -77,10 +76,9 @@ dependencies = [
   "sentence-transformers>=4.1",
   "setproctitle~=1.3.4",
   "tika-client~=0.10.0",
-  "torch~=2.9.1",
   "tqdm~=4.67.1",
-  "watchdog~=6.0",
-  "whitenoise~=6.11",
+  "watchfiles>=1.1.1",
+  "whitenoise~=6.9",
   "whoosh-reloaded>=2.7.5",
   "zxing-cpp~=2.3.0",
 ]
@@ -89,13 +87,13 @@ optional-dependencies.mariadb = [
   "mysqlclient~=2.2.7",
 ]
 optional-dependencies.postgres = [
-  "psycopg[c,pool]==3.3",
+  "psycopg[c,pool]==3.2.12",
   # Direct dependency for proper resolution of the pre-built wheels
-  "psycopg-c==3.3",
-  "psycopg-pool==3.3",
+  "psycopg-c==3.2.12",
+  "psycopg-pool==3.2.7",
 ]
 optional-dependencies.webserver = [
-  "granian[uvloop]~=2.6.0",
+  "granian[uvloop]~=2.5.1",
 ]
 
 [dependency-groups]
@@ -127,7 +125,7 @@ testing = [
 ]
 
 lint = [
-  "pre-commit~=4.5.1",
+  "pre-commit~=4.4.0",
   "pre-commit-uv~=4.2.0",
   "ruff~=0.14.0",
 ]
@@ -152,7 +150,7 @@ typing = [
 ]
 
 [tool.uv]
-required-version = ">=0.9.0"
+required-version = ">=0.5.14"
 package = false
 environments = [
   "sys_platform == 'darwin'",
@@ -162,23 +160,14 @@ environments = [
 [tool.uv.sources]
 # Markers are chosen to select these almost exclusively when building the Docker image
 psycopg-c = [
-  { url = "https://github.com/paperless-ngx/builder/releases/download/psycopg-trixie-3.3.0/psycopg_c-3.3.0-cp312-cp312-linux_x86_64.whl", marker = "sys_platform == 'linux' and platform_machine == 'x86_64' and python_version == '3.12'" },
-  { url = "https://github.com/paperless-ngx/builder/releases/download/psycopg-trixie-3.3.0/psycopg_c-3.3.0-cp312-cp312-linux_aarch64.whl", marker = "sys_platform == 'linux' and platform_machine == 'aarch64' and python_version == '3.12'" },
+  { url = "https://github.com/paperless-ngx/builder/releases/download/psycopg-bookworm-3.2.12/psycopg_c-3.2.12-cp312-cp312-linux_x86_64.whl", marker = "sys_platform == 'linux' and platform_machine == 'x86_64' and python_version == '3.12'" },
+  { url = "https://github.com/paperless-ngx/builder/releases/download/psycopg-bookworm-3.2.12/psycopg_c-3.2.12-cp312-cp312-linux_aarch64.whl", marker = "sys_platform == 'linux' and platform_machine == 'aarch64' and python_version == '3.12'" },
 ]
 zxing-cpp = [
   { url = "https://github.com/paperless-ngx/builder/releases/download/zxing-2.3.0/zxing_cpp-2.3.0-cp312-cp312-linux_x86_64.whl", marker = "sys_platform == 'linux' and platform_machine == 'x86_64' and python_version == '3.12'" },
   { url = "https://github.com/paperless-ngx/builder/releases/download/zxing-2.3.0/zxing_cpp-2.3.0-cp312-cp312-linux_aarch64.whl", marker = "sys_platform == 'linux' and platform_machine == 'aarch64' and python_version == '3.12'" },
 ]
 
-torch = [
-  { index = "pytorch-cpu" },
-]
-
-[[tool.uv.index]]
-name = "pytorch-cpu"
-url = "https://download.pytorch.org/whl/cpu"
-explicit = true
-
 [tool.ruff]
 target-version = "py310"
 line-length = 88

src-ui/src/app/components/common/edit-dialog/workflow-edit-dialog/workflow-edit-dialog.component.spec.ts

@@ -252,7 +252,7 @@ describe('WorkflowEditDialogComponent', () => {
     expect(component.object.actions.length).toEqual(2)
   })
 
-  it('should update order on drag n drop', () => {
+  it('should update order and remove ids from actions on drag n drop', () => {
     const action1 = workflow.actions[0]
     const action2 = workflow.actions[1]
     component.object = workflow
@@ -261,6 +261,8 @@ describe('WorkflowEditDialogComponent', () => {
       WorkflowAction[]
     >)
     expect(component.object.actions).toEqual([action2, action1])
+    expect(action1.id).toBeNull()
+    expect(action2.id).toBeNull()
   })
 
   it('should not include auto matching in algorithms', () => {

src-ui/src/app/components/common/edit-dialog/workflow-edit-dialog/workflow-edit-dialog.component.ts

@@ -1283,6 +1283,11 @@ export class WorkflowEditDialogComponent
     const actionField = this.actionFields.at(event.previousIndex)
     this.actionFields.removeAt(event.previousIndex)
     this.actionFields.insert(event.currentIndex, actionField)
+    // removing id will effectively re-create the actions in this order
+    this.object.actions.forEach((a) => (a.id = null))
+    this.actionFields.controls.forEach((c) =>
+      c.get('id').setValue(null, { emitEvent: false })
+    )
   }
 
   save(): void {

src/documents/management/commands/document_consumer.py

@@ -1,135 +1,343 @@
+"""
+Document consumer management command.
+
+Watches a consumption directory for new documents and queues them for processing.
+Uses watchfiles for efficient file system monitoring with support for both
+native OS notifications and polling fallback.
+"""
+
+from __future__ import annotations
+
 import logging
-import os
-from concurrent.futures import ThreadPoolExecutor
-from fnmatch import filter
+from dataclasses import dataclass
 from pathlib import Path
-from pathlib import PurePath
 from threading import Event
 from time import monotonic
-from time import sleep
+from typing import TYPE_CHECKING
 from typing import Final
 
 from django import db
 from django.conf import settings
 from django.core.management.base import BaseCommand
 from django.core.management.base import CommandError
-from watchdog.events import FileSystemEventHandler
-from watchdog.observers.polling import PollingObserver
+from watchfiles import Change
+from watchfiles import DefaultFilter
+from watchfiles import watch
 
 from documents.data_models import ConsumableDocument
 from documents.data_models import DocumentMetadataOverrides
 from documents.data_models import DocumentSource
 from documents.models import Tag
-from documents.parsers import is_file_ext_supported
+from documents.parsers import get_supported_file_extensions
 from documents.tasks import consume_file
 
-try:
-    from inotifyrecursive import INotify
-    from inotifyrecursive import flags
-except ImportError:  # pragma: no cover
-    INotify = flags = None
+if TYPE_CHECKING:
+    from collections.abc import Iterator
+
 
 logger = logging.getLogger("paperless.management.consumer")
 
 
-def _tags_from_path(filepath: Path) -> list[int]:
+@dataclass
+class TrackedFile:
+    """Represents a file being tracked for stability."""
+
+    path: Path
+    last_event_time: float
+    last_mtime: float | None = None
+    last_size: int | None = None
+
+    def update_stats(self) -> bool:
+        """
+        Update file stats. Returns True if file exists and stats were updated.
+        """
+        try:
+            stat = self.path.stat()
+            self.last_mtime = stat.st_mtime
+            self.last_size = stat.st_size
+            return True
+        except OSError:
+            return False
+
+    def is_unchanged(self) -> bool:
+        """
+        Check if file stats match the previously recorded values.
+        Returns False if file doesn't exist or stats changed.
+        """
+        try:
+            stat = self.path.stat()
+            return stat.st_mtime == self.last_mtime and stat.st_size == self.last_size
+        except OSError:
+            return False
+
+
+class FileStabilityTracker:
     """
-    Walk up the directory tree from filepath to CONSUMPTION_DIR
+    Tracks file events and determines when files are stable for consumption.
+
+    A file is considered stable when:
+    1. No new events have been received for it within the stability delay
+    2. Its size and modification time haven't changed
+    3. It still exists as a regular file
+
+    This handles various edge cases:
+    - Network copies that write in chunks
+    - Scanners that open/close files multiple times
+    - Temporary files that get renamed
+    - Files that are deleted before becoming stable
+    """
+
+    def __init__(self, stability_delay: float = 1.0) -> None:
+        """
+        Initialize the tracker.
+
+        Args:
+            stability_delay: Time in seconds a file must remain unchanged
+                           before being considered stable.
+        """
+        self.stability_delay = stability_delay
+        self._tracked: dict[Path, TrackedFile] = {}
+
+    def track(self, path: Path, change: Change) -> None:
+        """
+        Register a file event.
+
+        Args:
+            path: The file path that changed.
+            change: The type of change (added, modified, deleted).
+        """
+        path = path.resolve()
+
+        match change:
+            case Change.deleted:
+                self._tracked.pop(path, None)
+                logger.debug(f"Stopped tracking deleted file: {path}")
+            case Change.added | Change.modified:
+                current_time = monotonic()
+                if path in self._tracked:
+                    tracked = self._tracked[path]
+                    tracked.last_event_time = current_time
+                    tracked.update_stats()
+                    logger.debug(f"Updated tracking for: {path}")
+                else:
+                    tracked = TrackedFile(path=path, last_event_time=current_time)
+                    if tracked.update_stats():
+                        self._tracked[path] = tracked
+                        logger.debug(f"Started tracking: {path}")
+                    else:
+                        logger.debug(f"Could not stat file, not tracking: {path}")
+
+    def get_stable_files(self) -> Iterator[Path]:
+        """
+        Yield files that have been stable for the configured delay.
+
+        Files are removed from tracking once yielded or determined to be invalid.
+        """
+        current_time = monotonic()
+        to_remove: list[Path] = []
+        to_yield: list[Path] = []
+
+        for path, tracked in self._tracked.items():
+            time_since_event = current_time - tracked.last_event_time
+
+            if time_since_event < self.stability_delay:
+                continue
+
+            # File has waited long enough, verify it's unchanged
+            if not tracked.is_unchanged():
+                # Stats changed or file gone - update and wait again
+                if tracked.update_stats():
+                    tracked.last_event_time = current_time
+                    logger.debug(f"File changed during stability check: {path}")
+                else:
+                    # File no longer exists, remove from tracking
+                    to_remove.append(path)
+                    logger.debug(f"File disappeared during stability check: {path}")
+                continue
+
+            # File is stable, we can return it
+            to_yield.append(path)
+            logger.info(f"File is stable: {path}")
+
+        # Remove files that are no longer valid
+        for path in to_remove:
+            self._tracked.pop(path, None)
+
+        # Remove and yield stable files
+        for path in to_yield:
+            self._tracked.pop(path, None)
+            yield path
+
+    def has_pending_files(self) -> bool:
+        """Check if there are files waiting for stability check."""
+        return len(self._tracked) > 0
+
+    @property
+    def pending_count(self) -> int:
+        """Number of files being tracked."""
+        return len(self._tracked)
+
+
+class ConsumerFilter(DefaultFilter):
+    """
+    Filter for watchfiles that accepts only supported document types
+    and ignores system files/directories.
+
+    Extends DefaultFilter leveraging its built-in filtering:
+    - `ignore_dirs`: Directory names to ignore (and all their contents)
+    - `ignore_entity_patterns`: Regex patterns matched against filename/dirname only
+
+    We add custom logic for file extension filtering (only accept supported
+    document types), which the library doesn't provide.
+    """
+
+    # Regex patterns for files to always ignore (matched against filename only)
+    # These are passed to DefaultFilter.ignore_entity_patterns
+    DEFAULT_IGNORE_PATTERNS: Final[tuple[str, ...]] = (
+        r"^\.DS_Store$",
+        r"^\.DS_STORE$",
+        r"^\._.*",
+        r"^desktop\.ini$",
+        r"^Thumbs\.db$",
+    )
+
+    # Directories to always ignore (passed to DefaultFilter.ignore_dirs)
+    # These are matched by directory name, not full path
+    DEFAULT_IGNORE_DIRS: Final[tuple[str, ...]] = (
+        ".stfolder",  # Syncthing
+        ".stversions",  # Syncthing
+        ".localized",  # macOS
+        "@eaDir",  # Synology NAS
+        ".Spotlight-V100",  # macOS
+        ".Trashes",  # macOS
+        "__MACOSX",  # macOS archive artifacts
+    )
+
+    def __init__(
+        self,
+        *,
+        supported_extensions: frozenset[str] | None = None,
+        ignore_patterns: list[str] | None = None,
+        ignore_dirs: list[str] | None = None,
+    ) -> None:
+        """
+        Initialize the consumer filter.
+
+        Args:
+            supported_extensions: Set of file extensions to accept (e.g., {".pdf", ".png"}).
+                If None, uses get_supported_file_extensions().
+            ignore_patterns: Additional regex patterns to ignore (matched against filename).
+            ignore_dirs: Additional directory names to ignore (merged with defaults).
+        """
+        # Get supported extensions
+        if supported_extensions is None:
+            supported_extensions = frozenset(get_supported_file_extensions())
+        self._supported_extensions = supported_extensions
+
+        # Combine default and user patterns
+        all_patterns: list[str] = list(self.DEFAULT_IGNORE_PATTERNS)
+        if ignore_patterns:
+            all_patterns.extend(ignore_patterns)
+
+        # Combine default and user ignore_dirs
+        all_ignore_dirs: list[str] = list(self.DEFAULT_IGNORE_DIRS)
+        if ignore_dirs:
+            all_ignore_dirs.extend(ignore_dirs)
+
+        # Let DefaultFilter handle all the pattern and directory filtering
+        super().__init__(
+            ignore_dirs=tuple(all_ignore_dirs),
+            ignore_entity_patterns=tuple(all_patterns),
+            ignore_paths=(),
+        )
+
+    def __call__(self, change: Change, path: str) -> bool:
+        """
+        Filter function for watchfiles.
+
+        Returns True if the path should be watched, False to ignore.
+
+        The parent DefaultFilter handles:
+        - Hidden files/directories (starting with .)
+        - Directories in ignore_dirs
+        - Files/directories matching ignore_entity_patterns
+
+        We additionally filter files by extension.
+        """
+        # Let parent filter handle directory ignoring and pattern matching
+        if not super().__call__(change, path):
+            return False
+
+        path_obj = Path(path)
+
+        # For directories, parent filter already handled everything
+        if path_obj.is_dir():
+            return True
+
+        # For files, check extension
+        return self._has_supported_extension(path_obj)
+
+    def _has_supported_extension(self, path: Path) -> bool:
+        """Check if the file has a supported extension."""
+        suffix = path.suffix.lower()
+        return suffix in self._supported_extensions
+
+
+def _tags_from_path(filepath: Path, consumption_dir: Path) -> list[int]:
+    """
+    Walk up the directory tree from filepath to consumption_dir
     and get or create Tag IDs for every directory.
 
-    Returns set of Tag models
+    Returns list of Tag primary keys.
     """
     db.close_old_connections()
-    tag_ids = set()
-    path_parts = filepath.relative_to(settings.CONSUMPTION_DIR).parent.parts
+    tag_ids: set[int] = set()
+    path_parts = filepath.relative_to(consumption_dir).parent.parts
+
     for part in path_parts:
-        tag_ids.add(
-            Tag.objects.get_or_create(name__iexact=part, defaults={"name": part})[0].pk,
+        tag, _ = Tag.objects.get_or_create(
+            name__iexact=part,
+            defaults={"name": part},
         )
+        tag_ids.add(tag.pk)
 
     return list(tag_ids)
 
 
-def _is_ignored(filepath: Path) -> bool:
+def _consume_file(
+    filepath: Path,
+    consumption_dir: Path,
+    *,
+    subdirs_as_tags: bool,
+) -> None:
     """
-    Checks if the given file should be ignored, based on configured
-    patterns.
+    Queue a file for consumption.
 
-    Returns True if the file is ignored, False otherwise
+    Args:
+        filepath: Path to the file to consume.
+        consumption_dir: Base consumption directory.
+        subdirs_as_tags: Whether to create tags from subdirectory names.
     """
-    # Trim out the consume directory, leaving only filename and it's
-    # path relative to the consume directory
-    filepath_relative = PurePath(filepath).relative_to(settings.CONSUMPTION_DIR)
-
-    # March through the components of the path, including directories and the filename
-    # looking for anything matching
-    # foo/bar/baz/file.pdf -> (foo, bar, baz, file.pdf)
-    parts = []
-    for part in filepath_relative.parts:
-        # If the part is not the name (ie, it's a dir)
-        # Need to append the trailing slash or fnmatch doesn't match
-        # fnmatch("dir", "dir/*") == False
-        # fnmatch("dir/", "dir/*") == True
-        if part != filepath_relative.name:
-            part = part + "/"
-        parts.append(part)
-
-    for pattern in settings.CONSUMER_IGNORE_PATTERNS:
-        if len(filter(parts, pattern)):
-            return True
-
-    return False
-
-
-def _consume(filepath: Path) -> None:
-    # Check permissions early
+    # Verify file still exists and is accessible
     try:
-        filepath.stat()
-    except (PermissionError, OSError):
-        logger.warning(f"Not consuming file {filepath}: Permission denied.")
+        if not filepath.is_file():
+            logger.debug(f"Not consuming {filepath}: not a file or doesn't exist")
+            return
+    except OSError as e:
+        logger.warning(f"Not consuming {filepath}: {e}")
         return
 
-    if filepath.is_dir() or _is_ignored(filepath):
-        return
-
-    if not filepath.is_file():
-        logger.debug(f"Not consuming file {filepath}: File has moved.")
-        return
-
-    if not is_file_ext_supported(filepath.suffix):
-        logger.warning(f"Not consuming file {filepath}: Unknown file extension.")
-        return
-
-    # Total wait time: up to 500ms
-    os_error_retry_count: Final[int] = 50
-    os_error_retry_wait: Final[float] = 0.01
-
-    read_try_count = 0
-    file_open_ok = False
-    os_error_str = None
-
-    while (read_try_count < os_error_retry_count) and not file_open_ok:
+    # Get tags from path if configured
+    tag_ids: list[int] | None = None
+    if subdirs_as_tags:
         try:
-            with filepath.open("rb"):
-                file_open_ok = True
-        except OSError as e:
-            read_try_count += 1
-            os_error_str = str(e)
-            sleep(os_error_retry_wait)
+            tag_ids = _tags_from_path(filepath, consumption_dir)
+        except Exception:
+            logger.exception(f"Error creating tags from path for {filepath}")
 
-    if read_try_count >= os_error_retry_count:
-        logger.warning(f"Not consuming file {filepath}: OS reports {os_error_str}")
-        return
-
-    tag_ids = None
+    # Queue for consumption
     try:
-        if settings.CONSUMER_SUBDIRS_AS_TAGS:
-            tag_ids = _tags_from_path(filepath)
-    except Exception:
-        logger.exception("Error creating tags from path")
-
-    try:
-        logger.info(f"Adding {filepath} to the task queue.")
+        logger.info(f"Adding {filepath} to the task queue")
         consume_file.delay(
             ConsumableDocument(
                 source=DocumentSource.ConsumeFolder,
@@ -138,228 +346,209 @@ def _consume(filepath: Path) -> None:
             DocumentMetadataOverrides(tag_ids=tag_ids),
         )
     except Exception:
-        # Catch all so that the consumer won't crash.
-        # This is also what the test case is listening for to check for
-        # errors.
-        logger.exception("Error while consuming document")
-
-
-def _consume_wait_unmodified(file: Path) -> None:
-    """
-    Waits for the given file to appear unmodified based on file size
-    and modification time.  Will wait a configured number of seconds
-    and retry a configured number of times before either consuming or
-    giving up
-    """
-    if _is_ignored(file):
-        return
-
-    logger.debug(f"Waiting for file {file} to remain unmodified")
-    mtime = -1
-    size = -1
-    current_try = 0
-    while current_try < settings.CONSUMER_POLLING_RETRY_COUNT:
-        try:
-            stat_data = file.stat()
-            new_mtime = stat_data.st_mtime
-            new_size = stat_data.st_size
-        except FileNotFoundError:
-            logger.debug(
-                f"File {file} moved while waiting for it to remain unmodified.",
-            )
-            return
-        if new_mtime == mtime and new_size == size:
-            _consume(file)
-            return
-        mtime = new_mtime
-        size = new_size
-        sleep(settings.CONSUMER_POLLING_DELAY)
-        current_try += 1
-
-    logger.error(f"Timeout while waiting on file {file} to remain unmodified.")
-
-
-class Handler(FileSystemEventHandler):
-    def __init__(self, pool: ThreadPoolExecutor) -> None:
-        super().__init__()
-        self._pool = pool
-
-    def on_created(self, event):
-        self._pool.submit(_consume_wait_unmodified, Path(event.src_path))
-
-    def on_moved(self, event):
-        self._pool.submit(_consume_wait_unmodified, Path(event.dest_path))
+        logger.exception(f"Error while queuing document {filepath}")
 
 
 class Command(BaseCommand):
     """
-    On every iteration of an infinite loop, consume what we can from the
-    consumption directory.
+    Watch a consumption directory and queue new documents for processing.
+
+    Uses watchfiles for efficient file system monitoring. Supports both
+    native OS notifications (inotify on Linux, FSEvents on macOS) and
+    polling for network filesystems.
     """
 
-    # This is here primarily for the tests and is irrelevant in production.
-    stop_flag = Event()
-    # Also only for testing, configures in one place the timeout used before checking
-    # the stop flag
-    testing_timeout_s: Final[float] = 0.5
-    testing_timeout_ms: Final[float] = testing_timeout_s * 1000.0
+    help = "Watch the consumption directory for new documents"
 
-    def add_arguments(self, parser):
+    # For testing - allows tests to stop the consumer
+    stop_flag: Event = Event()
+
+    # Testing timeout in seconds
+    testing_timeout_s: Final[float] = 0.5
+
+    def add_arguments(self, parser) -> None:
         parser.add_argument(
             "directory",
-            default=settings.CONSUMPTION_DIR,
+            default=None,
             nargs="?",
-            help="The consumption directory.",
+            help="The consumption directory (defaults to CONSUMPTION_DIR setting)",
+        )
+        parser.add_argument(
+            "--oneshot",
+            action="store_true",
+            help="Process existing files and exit without watching",
         )
-        parser.add_argument("--oneshot", action="store_true", help="Run only once.")
-
-        # Only use during unit testing, will configure a timeout
-        # Leaving it unset or false and the consumer will exit when it
-        # receives SIGINT
         parser.add_argument(
             "--testing",
             action="store_true",
-            help="Flag used only for unit testing",
+            help="Enable testing mode with shorter timeouts",
             default=False,
         )
 
-    def handle(self, *args, **options):
-        directory = options["directory"]
-        recursive = settings.CONSUMER_RECURSIVE
-
+    def handle(self, *args, **options) -> None:
+        # Resolve consumption directory
+        directory = options.get("directory")
         if not directory:
-            raise CommandError("CONSUMPTION_DIR does not appear to be set.")
+            directory = getattr(settings, "CONSUMPTION_DIR", None)
+        if not directory:
+            raise CommandError("CONSUMPTION_DIR is not configured")
 
         directory = Path(directory).resolve()
 
-        if not directory.is_dir():
-            raise CommandError(f"Consumption directory {directory} does not exist")
+        if not directory.exists():
+            raise CommandError(f"Consumption directory does not exist: {directory}")
 
-        # Consumer will need this
+        if not directory.is_dir():
+            raise CommandError(f"Consumption path is not a directory: {directory}")
+
+        # Ensure scratch directory exists
         settings.SCRATCH_DIR.mkdir(parents=True, exist_ok=True)
 
-        if recursive:
-            for dirpath, _, filenames in os.walk(directory):
-                for filename in filenames:
-                    filepath = Path(dirpath) / filename
-                    _consume(filepath)
-        else:
-            for filepath in directory.iterdir():
-                _consume(filepath)
+        # Get settings
+        recursive: bool = settings.CONSUMER_RECURSIVE
+        subdirs_as_tags: bool = settings.CONSUMER_SUBDIRS_AS_TAGS
+        polling_interval: float = settings.CONSUMER_POLLING_INTERVAL
+        stability_delay: float = settings.CONSUMER_STABILITY_DELAY
+        ignore_patterns: list[str] = settings.CONSUMER_IGNORE_PATTERNS
+        ignore_dirs: list[str] = settings.CONSUMER_IGNORE_DIRS
+        is_testing: bool = options.get("testing", False)
+        is_oneshot: bool = options.get("oneshot", False)
 
-        if options["oneshot"]:
+        # Create filter
+        consumer_filter = ConsumerFilter(
+            ignore_patterns=ignore_patterns,
+            ignore_dirs=ignore_dirs,
+        )
+
+        # Process existing files
+        self._process_existing_files(
+            directory=directory,
+            recursive=recursive,
+            subdirs_as_tags=subdirs_as_tags,
+            consumer_filter=consumer_filter,
+        )
+
+        if is_oneshot:
+            logger.info("Oneshot mode: processed existing files, exiting")
             return
 
-        if settings.CONSUMER_POLLING == 0 and INotify:
-            self.handle_inotify(directory, recursive, is_testing=options["testing"])
+        # Start watching
+        self._watch_directory(
+            directory=directory,
+            recursive=recursive,
+            subdirs_as_tags=subdirs_as_tags,
+            consumer_filter=consumer_filter,
+            polling_interval=polling_interval,
+            stability_delay=stability_delay,
+            is_testing=is_testing,
+        )
+
+        logger.debug("Consumer exiting")
+
+    def _process_existing_files(
+        self,
+        *,
+        directory: Path,
+        recursive: bool,
+        subdirs_as_tags: bool,
+        consumer_filter: ConsumerFilter,
+    ) -> None:
+        """Process any existing files in the consumption directory."""
+        logger.info(f"Processing existing files in {directory}")
+
+        glob_pattern = "**/*" if recursive else "*"
+
+        for filepath in directory.glob(glob_pattern):
+            # Use filter to check if file should be processed
+            if not filepath.is_file():
+                continue
+
+            if not consumer_filter(Change.added, str(filepath)):
+                continue
+
+            _consume_file(
+                filepath=filepath,
+                consumption_dir=directory,
+                subdirs_as_tags=subdirs_as_tags,
+            )
+
+    def _watch_directory(
+        self,
+        *,
+        directory: Path,
+        recursive: bool,
+        subdirs_as_tags: bool,
+        consumer_filter: ConsumerFilter,
+        polling_interval: float,
+        stability_delay: float,
+        is_testing: bool,
+    ) -> None:
+        """Watch directory for changes and process stable files."""
+        use_polling = polling_interval > 0
+        poll_delay_ms = int(polling_interval * 1000) if use_polling else 0
+
+        if use_polling:
+            logger.info(
+                f"Watching {directory} using polling (interval: {polling_interval}s)",
+            )
         else:
-            if INotify is None and settings.CONSUMER_POLLING == 0:  # pragma: no cover
-                logger.warning("Using polling as INotify import failed")
-            self.handle_polling(directory, recursive, is_testing=options["testing"])
+            logger.info(f"Watching {directory} using native file system events")
 
-        logger.debug("Consumer exiting.")
+        # Create stability tracker
+        tracker = FileStabilityTracker(stability_delay=stability_delay)
 
-    def handle_polling(self, directory, recursive, *, is_testing: bool):
-        logger.info(f"Polling directory for changes: {directory}")
+        # Calculate timeouts
+        stability_timeout_ms = int(stability_delay * 1000)
+        testing_timeout_ms = int(self.testing_timeout_s * 1000)
 
-        timeout = None
-        if is_testing:
-            timeout = self.testing_timeout_s
-            logger.debug(f"Configuring timeout to {timeout}s")
+        # Start with no timeout (wait indefinitely for first event)
+        # unless in testing mode
+        timeout_ms = testing_timeout_ms if is_testing else 0
 
-        polling_interval = settings.CONSUMER_POLLING
-        if polling_interval == 0:  # pragma: no cover
-            # Only happens if INotify failed to import
-            logger.warning("Using polling of 10s, consider setting this")
-            polling_interval = 10
+        self.stop_flag.clear()
 
-        with ThreadPoolExecutor(max_workers=4) as pool:
-            observer = PollingObserver(timeout=polling_interval)
-            observer.schedule(Handler(pool), directory, recursive=recursive)
-            observer.start()
+        while not self.stop_flag.is_set():
             try:
-                while observer.is_alive():
-                    observer.join(timeout)
-                    if self.stop_flag.is_set():
-                        observer.stop()
-            except KeyboardInterrupt:
-                observer.stop()
-            observer.join()
-
-    def handle_inotify(self, directory, recursive, *, is_testing: bool):
-        logger.info(f"Using inotify to watch directory for changes: {directory}")
-
-        timeout_ms = None
-        if is_testing:
-            timeout_ms = self.testing_timeout_ms
-            logger.debug(f"Configuring timeout to {timeout_ms}ms")
-
-        inotify = INotify()
-        inotify_flags = flags.CLOSE_WRITE | flags.MOVED_TO | flags.MODIFY
-        if recursive:
-            inotify.add_watch_recursive(directory, inotify_flags)
-        else:
-            inotify.add_watch(directory, inotify_flags)
-
-        inotify_debounce_secs: Final[float] = settings.CONSUMER_INOTIFY_DELAY
-        inotify_debounce_ms: Final[int] = inotify_debounce_secs * 1000
-
-        finished = False
-
-        notified_files = {}
-
-        try:
-            while not finished:
-                try:
-                    for event in inotify.read(timeout=timeout_ms):
-                        path = inotify.get_path(event.wd) if recursive else directory
-                        filepath = Path(path) / event.name
-                        if flags.MODIFY in flags.from_mask(event.mask):
-                            notified_files.pop(filepath, None)
-                        else:
-                            notified_files[filepath] = monotonic()
-
-                    # Check the files against the timeout
-                    still_waiting = {}
-                    # last_event_time is time of the last inotify event for this file
-                    for filepath, last_event_time in notified_files.items():
-                        # Current time - last time over the configured timeout
-                        waited_long_enough = (
-                            monotonic() - last_event_time
-                        ) > inotify_debounce_secs
-
-                        # Also make sure the file exists still, some scanners might write a
-                        # temporary file first
-                        try:
-                            file_still_exists = filepath.exists() and filepath.is_file()
-                        except (PermissionError, OSError):  # pragma: no cover
-                            # If we can't check, let it fail in the _consume function
-                            file_still_exists = True
+                for changes in watch(
+                    directory,
+                    watch_filter=consumer_filter,
+                    rust_timeout=timeout_ms,
+                    yield_on_timeout=True,
+                    force_polling=use_polling,
+                    poll_delay_ms=poll_delay_ms,
+                    recursive=recursive,
+                    stop_event=self.stop_flag,
+                ):
+                    # Process each change
+                    for change_type, path in changes:
+                        path = Path(path).resolve()
+                        if not path.is_file():
                             continue
+                        logger.debug(f"Event: {change_type.name} for {path}")
+                        tracker.track(path, change_type)
 
-                        if waited_long_enough and file_still_exists:
-                            _consume(filepath)
-                        elif file_still_exists:
-                            still_waiting[filepath] = last_event_time
+                    # Check for stable files
+                    for stable_path in tracker.get_stable_files():
+                        _consume_file(
+                            filepath=stable_path,
+                            consumption_dir=directory,
+                            subdirs_as_tags=subdirs_as_tags,
+                        )
 
-                    # These files are still waiting to hit the timeout
-                    notified_files = still_waiting
+                    # Exit watch loop to reconfigure timeout
+                    break
 
-                    # If files are waiting, need to exit read() to check them
-                    # Otherwise, go back to infinite sleep time, but only if not testing
-                    if len(notified_files) > 0:
-                        timeout_ms = inotify_debounce_ms
-                    elif is_testing:
-                        timeout_ms = self.testing_timeout_ms
-                    else:
-                        timeout_ms = None
+                # Determine next timeout
+                if tracker.has_pending_files():
+                    # Check pending files at stability interval
+                    timeout_ms = stability_timeout_ms
+                elif is_testing:
+                    # In testing, use short timeout to check stop flag
+                    timeout_ms = testing_timeout_ms
+                else:  # pragma: nocover
+                    # No pending files, wait indefinitely
+                    timeout_ms = 0
 
-                    if self.stop_flag.is_set():
-                        logger.debug("Finishing because event is set")
-                        finished = True
-
-                except KeyboardInterrupt:
-                    logger.info("Received SIGINT, stopping inotify")
-                    finished = True
-        finally:
-            inotify.close()
+            except KeyboardInterrupt:  # pragma: nocover
+                logger.info("Received interrupt, stopping consumer")
+                self.stop_flag.set()

src/documents/migrations/1076_workflowaction_order.py

@@ -1,28 +0,0 @@
-# Generated by Django 5.2.7 on 2026-01-14 16:53
-
-from django.db import migrations
-from django.db import models
-from django.db.models import F
-
-
-def populate_action_order(apps, schema_editor):
-    WorkflowAction = apps.get_model("documents", "WorkflowAction")
-    WorkflowAction.objects.all().update(order=F("id"))
-
-
-class Migration(migrations.Migration):
-    dependencies = [
-        ("documents", "1075_alter_paperlesstask_task_name"),
-    ]
-
-    operations = [
-        migrations.AddField(
-            model_name="workflowaction",
-            name="order",
-            field=models.PositiveIntegerField(default=0, verbose_name="order"),
-        ),
-        migrations.RunPython(
-            populate_action_order,
-            reverse_code=migrations.RunPython.noop,
-        ),
-    ]

src/documents/models.py

@@ -1295,8 +1295,6 @@ class WorkflowAction(models.Model):
         default=WorkflowActionType.ASSIGNMENT,
     )
 
-    order = models.PositiveIntegerField(_("order"), default=0)
-
     assign_title = models.TextField(
         _("assign title"),
         null=True,

src/documents/serialisers.py

@@ -2577,8 +2577,7 @@ class WorkflowSerializer(serializers.ModelSerializer):
                 set_triggers.append(trigger_instance)
 
         if actions is not None and actions is not serializers.empty:
-            for index, action in enumerate(actions):
-                action["order"] = index
+            for action in actions:
                 assign_tags = action.pop("assign_tags", None)
                 assign_view_users = action.pop("assign_view_users", None)
                 assign_view_groups = action.pop("assign_view_groups", None)
@@ -2705,16 +2704,6 @@ class WorkflowSerializer(serializers.ModelSerializer):
 
         return instance
 
-    def to_representation(self, instance):
-        data = super().to_representation(instance)
-        actions = instance.actions.order_by("order", "pk")
-        data["actions"] = WorkflowActionSerializer(
-            actions,
-            many=True,
-            context=self.context,
-        ).data
-        return data
-
 
 class TrashSerializer(SerializerWithPerms):
     documents = serializers.ListField(

src/documents/signals/handlers.py

@@ -781,7 +781,7 @@ def run_workflows(
 
         if matching.document_matches_workflow(document, workflow, trigger_type):
             action: WorkflowAction
-            for action in workflow.actions.order_by("order", "pk"):
+            for action in workflow.actions.all():
                 message = f"Applying {action} from {workflow}"
                 if not use_overrides:
                     logger.info(message, extra={"group": logging_group})

src/documents/workflows/utils.py

@@ -20,6 +20,9 @@ def get_workflows_for_trigger(
     wrap it in a list; otherwise fetch enabled workflows for the trigger with
     the prefetches used by the runner.
     """
+    if workflow_to_run is not None:
+        return [workflow_to_run]
+
     annotated_actions = (
         WorkflowAction.objects.select_related(
             "assign_correspondent",
@@ -102,25 +105,10 @@ def get_workflows_for_trigger(
         )
     )
 
-    action_prefetch = Prefetch(
-        "actions",
-        queryset=annotated_actions.order_by("order", "pk"),
-    )
-
-    if workflow_to_run is not None:
-        return (
-            Workflow.objects.filter(pk=workflow_to_run.pk)
-            .prefetch_related(
-                action_prefetch,
-                "triggers",
-            )
-            .distinct()
-        )
-
     return (
         Workflow.objects.filter(enabled=True, triggers__type=trigger_type)
         .prefetch_related(
-            action_prefetch,
+            Prefetch("actions", queryset=annotated_actions),
             "triggers",
         )
         .order_by("order")

src/locale/en_US/LC_MESSAGES/django.po

@@ -2,7 +2,7 @@ msgid ""
 msgstr ""
 "Project-Id-Version: paperless-ngx\n"
 "Report-Msgid-Bugs-To: \n"
-"POT-Creation-Date: 2026-01-15 23:01+0000\n"
+"POT-Creation-Date: 2026-01-13 16:26+0000\n"
 "PO-Revision-Date: 2022-02-17 04:17\n"
 "Last-Translator: \n"
 "Language-Team: English\n"
@@ -89,7 +89,7 @@ msgstr ""
 msgid "Automatic"
 msgstr ""
 
-#: documents/models.py:64 documents/models.py:456 documents/models.py:1529
+#: documents/models.py:64 documents/models.py:456 documents/models.py:1527
 #: paperless_mail/models.py:23 paperless_mail/models.py:143
 msgid "name"
 msgstr ""
@@ -264,7 +264,7 @@ msgid "The position of this document in your physical document archive."
 msgstr ""
 
 #: documents/models.py:318 documents/models.py:700 documents/models.py:754
-#: documents/models.py:1572
+#: documents/models.py:1570
 msgid "document"
 msgstr ""
 
@@ -1047,180 +1047,179 @@ msgstr ""
 msgid "Workflow Action Type"
 msgstr ""
 
-#: documents/models.py:1298 documents/models.py:1531
-#: paperless_mail/models.py:145
-msgid "order"
-msgstr ""
-
-#: documents/models.py:1301
+#: documents/models.py:1299
 msgid "assign title"
 msgstr ""
 
-#: documents/models.py:1305
+#: documents/models.py:1303
 msgid "Assign a document title, must  be a Jinja2 template, see documentation."
 msgstr ""
 
-#: documents/models.py:1313 paperless_mail/models.py:274
+#: documents/models.py:1311 paperless_mail/models.py:274
 msgid "assign this tag"
 msgstr ""
 
-#: documents/models.py:1322 paperless_mail/models.py:282
+#: documents/models.py:1320 paperless_mail/models.py:282
 msgid "assign this document type"
 msgstr ""
 
-#: documents/models.py:1331 paperless_mail/models.py:296
+#: documents/models.py:1329 paperless_mail/models.py:296
 msgid "assign this correspondent"
 msgstr ""
 
-#: documents/models.py:1340
+#: documents/models.py:1338
 msgid "assign this storage path"
 msgstr ""
 
-#: documents/models.py:1349
+#: documents/models.py:1347
 msgid "assign this owner"
 msgstr ""
 
-#: documents/models.py:1356
+#: documents/models.py:1354
 msgid "grant view permissions to these users"
 msgstr ""
 
-#: documents/models.py:1363
+#: documents/models.py:1361
 msgid "grant view permissions to these groups"
 msgstr ""
 
-#: documents/models.py:1370
+#: documents/models.py:1368
 msgid "grant change permissions to these users"
 msgstr ""
 
-#: documents/models.py:1377
+#: documents/models.py:1375
 msgid "grant change permissions to these groups"
 msgstr ""
 
-#: documents/models.py:1384
+#: documents/models.py:1382
 msgid "assign these custom fields"
 msgstr ""
 
-#: documents/models.py:1388
+#: documents/models.py:1386
 msgid "custom field values"
 msgstr ""
 
-#: documents/models.py:1392
+#: documents/models.py:1390
 msgid "Optional values to assign to the custom fields."
 msgstr ""
 
-#: documents/models.py:1401
+#: documents/models.py:1399
 msgid "remove these tag(s)"
 msgstr ""
 
-#: documents/models.py:1406
+#: documents/models.py:1404
 msgid "remove all tags"
 msgstr ""
 
-#: documents/models.py:1413
+#: documents/models.py:1411
 msgid "remove these document type(s)"
 msgstr ""
 
-#: documents/models.py:1418
+#: documents/models.py:1416
 msgid "remove all document types"
 msgstr ""
 
-#: documents/models.py:1425
+#: documents/models.py:1423
 msgid "remove these correspondent(s)"
 msgstr ""
 
-#: documents/models.py:1430
+#: documents/models.py:1428
 msgid "remove all correspondents"
 msgstr ""
 
-#: documents/models.py:1437
+#: documents/models.py:1435
 msgid "remove these storage path(s)"
 msgstr ""
 
-#: documents/models.py:1442
+#: documents/models.py:1440
 msgid "remove all storage paths"
 msgstr ""
 
-#: documents/models.py:1449
+#: documents/models.py:1447
 msgid "remove these owner(s)"
 msgstr ""
 
-#: documents/models.py:1454
+#: documents/models.py:1452
 msgid "remove all owners"
 msgstr ""
 
-#: documents/models.py:1461
+#: documents/models.py:1459
 msgid "remove view permissions for these users"
 msgstr ""
 
-#: documents/models.py:1468
+#: documents/models.py:1466
 msgid "remove view permissions for these groups"
 msgstr ""
 
-#: documents/models.py:1475
+#: documents/models.py:1473
 msgid "remove change permissions for these users"
 msgstr ""
 
-#: documents/models.py:1482
+#: documents/models.py:1480
 msgid "remove change permissions for these groups"
 msgstr ""
 
-#: documents/models.py:1487
+#: documents/models.py:1485
 msgid "remove all permissions"
 msgstr ""
 
-#: documents/models.py:1494
+#: documents/models.py:1492
 msgid "remove these custom fields"
 msgstr ""
 
-#: documents/models.py:1499
+#: documents/models.py:1497
 msgid "remove all custom fields"
 msgstr ""
 
-#: documents/models.py:1508
+#: documents/models.py:1506
 msgid "email"
 msgstr ""
 
-#: documents/models.py:1517
+#: documents/models.py:1515
 msgid "webhook"
 msgstr ""
 
-#: documents/models.py:1521
+#: documents/models.py:1519
 msgid "workflow action"
 msgstr ""
 
-#: documents/models.py:1522
+#: documents/models.py:1520
 msgid "workflow actions"
 msgstr ""
 
-#: documents/models.py:1537
+#: documents/models.py:1529 paperless_mail/models.py:145
+msgid "order"
+msgstr ""
+
+#: documents/models.py:1535
 msgid "triggers"
 msgstr ""
 
-#: documents/models.py:1544
+#: documents/models.py:1542
 msgid "actions"
 msgstr ""
 
-#: documents/models.py:1547 paperless_mail/models.py:154
+#: documents/models.py:1545 paperless_mail/models.py:154
 msgid "enabled"
 msgstr ""
 
-#: documents/models.py:1558
+#: documents/models.py:1556
 msgid "workflow"
 msgstr ""
 
-#: documents/models.py:1562
+#: documents/models.py:1560
 msgid "workflow trigger type"
 msgstr ""
 
-#: documents/models.py:1576
+#: documents/models.py:1574
 msgid "date run"
 msgstr ""
 
-#: documents/models.py:1582
+#: documents/models.py:1580
 msgid "workflow run"
 msgstr ""
 
-#: documents/models.py:1583
+#: documents/models.py:1581
 msgid "workflow runs"
 msgstr ""
 

src/paperless/settings.py

@@ -1044,29 +1044,30 @@ IGNORABLE_FILES: Final[list[str]] = [
     "Thumbs.db",
 ]
 
-CONSUMER_POLLING = int(os.getenv("PAPERLESS_CONSUMER_POLLING", 0))
+CONSUMER_POLLING_INTERVAL = float(os.getenv("PAPERLESS_CONSUMER_POLLING_INTERVAL", 0))
 
-CONSUMER_POLLING_DELAY = int(os.getenv("PAPERLESS_CONSUMER_POLLING_DELAY", 5))
-
-CONSUMER_POLLING_RETRY_COUNT = int(
-    os.getenv("PAPERLESS_CONSUMER_POLLING_RETRY_COUNT", 5),
-)
-
-CONSUMER_INOTIFY_DELAY: Final[float] = __get_float(
-    "PAPERLESS_CONSUMER_INOTIFY_DELAY",
-    0.5,
-)
+CONSUMER_STABILITY_DELAY = float(os.getenv("PAPERLESS_CONSUMER_STABILITY_DELAY", 5))
 
 CONSUMER_DELETE_DUPLICATES = __get_boolean("PAPERLESS_CONSUMER_DELETE_DUPLICATES")
 
 CONSUMER_RECURSIVE = __get_boolean("PAPERLESS_CONSUMER_RECURSIVE")
 
-# Ignore glob patterns, relative to PAPERLESS_CONSUMPTION_DIR
+# Ignore regex patterns, matched against filename only
 CONSUMER_IGNORE_PATTERNS = list(
     json.loads(
         os.getenv(
             "PAPERLESS_CONSUMER_IGNORE_PATTERNS",
-            json.dumps(IGNORABLE_FILES),
+            json.dumps([]),
+        ),
+    ),
+)
+
+# Directories to always ignore.  These are matched by directory name, not full path
+CONSUMER_IGNORE_DIRS = list(
+    json.loads(
+        os.getenv(
+            "PAPERLESS_CONSUMER_IGNORE_DIRS",
+            json.dumps([]),
         ),
     ),
 )

src/paperless_ai/client.py

@@ -23,7 +23,7 @@ class AIClient:
     def get_llm(self) -> Ollama | OpenAI:
         if self.settings.llm_backend == "ollama":
             return Ollama(
-                model=self.settings.llm_model or "llama3.1",
+                model=self.settings.llm_model or "llama3",
                 base_url=self.settings.llm_endpoint or "http://localhost:11434",
                 request_timeout=120,
             )
@@ -52,7 +52,7 @@ class AIClient:
         )
         tool_calls = self.llm.get_tool_calls_from_response(
             result,
-            error_on_no_tool_call=True,
+            error_on_no_tool_calls=True,
         )
         logger.debug("LLM query result: %s", tool_calls)
         parsed = DocumentClassifierSchema(**tool_calls[0].tool_kwargs)

src/paperless_ai/tests/test_chat.py

@@ -11,12 +11,14 @@ from paperless_ai.chat import stream_chat_with_documents
 @pytest.fixture(autouse=True)
 def patch_embed_model():
     from llama_index.core import settings as llama_settings
-    from llama_index.core.embeddings.mock_embed_model import MockEmbedding
 
-    # Use a real BaseEmbedding subclass to satisfy llama-index 0.14 validation
-    llama_settings.Settings.embed_model = MockEmbedding(embed_dim=1536)
+    mock_embed_model = MagicMock()
+    mock_embed_model._get_text_embedding_batch.return_value = [
+        [0.1] * 1536,
+    ]  # 1 vector per input
+    llama_settings.Settings._embed_model = mock_embed_model
     yield
-    llama_settings.Settings.embed_model = None
+    llama_settings.Settings._embed_model = None
 
 
 @pytest.fixture(autouse=True)