### Description
As we add more and more steps to the pipeline (i.e. chunking, embedding,
table manipulation), it would help seperate the responsibility of each
of these into their own processes, running each in parallel using json
files to share data across. This will also help guarantee data is
serializable if this code was used in an actual pipeline. Following is a
flow diagram of the proposed changes. As part of this change:
* A parent pipeline class will be responsible for running each `node`,
which can optionally be run via multiprocessing if it supports it, or
not. Possible nodes at this moment:
* Doc factory: creates all the ingest docs via the source connector
* Source: reads/downloads all of the content to process to the local
filesystem to the location set by the `download_dir` parameter.
* Partition: runs partition on all of the downloaded content in json
format.
* Any number of reformat nodes that modify the partitioned content. This
can include chunking, embedding, etc.
* Write: push the final json into the destination via the destination
connector
* This pipeline relies on the information of the ingest docs to be
available via their serialization. An optimization was introduced with
the `IngestDocJsonMixin` which adds in all the `@property` fields to the
serialized json already being created via the `DataClassJsonMixin`
* For all intermediate steps (partitioning, reformatting), the content
is saved to a dedicated location on the local filesystem. Right now it's
set to `$HOME/.cache/unstructured/ingest/pipeline/STEP_NAME/`.
* Minor changes: made sense to move some of the config parameters
between the read and partition configs when I explicitly divided the
responsibility to download vs partition the content in the pipeline.
* The pipeline class only makes the doc factory, source and partition
nodes required, keeping with the logic that has been supported so far.
All reformatting nodes and write node are optional.
* Long term, there should also be some changes to the base configs
supported by the CLI to support pipeline specific configs, but for now
what exists was used to minimize changes in this PR.
* Final step to copy the final output to the location designated by the
`_output_filename` value of the ingest doc.
* Hashing occurs at each step by hashing the parameters of that step
(i.e. partition configs) along with the previous step via the filename
used. This allows each step to be the same _if_ all the parameters for
it have not changed and the content so far is the same.
* The only data that is shared and has writes to across processes is the
dictionary of ingest json data. This dict is created using the
`multiprocessing.manager.DictProxy` to make sure any interaction with it
is behind a lock.
### Minor refactors included:
* Utility methods added to extract configs from the click options
* Utility method to add common options to click commands.
* All writers moved to using the class approach which extracts a lot of
the common code so there's less copy-paste when new runners are added.
* Use `@property` for source metadata on base ingest doc to add logic to
call `update_source_metadata` if it's still `None` at the time it's
fetched.
### Additional bug fixes included
* Fsspec connectors were not serializable due to the `ingest_doc_cls`.
This was removed from the fields captured by the `@dataclass` decorator
and added in a `__post_init__` method.
* Various reddit connector params were missing. This doesn't have an
explicit ingest test at the moment so was never caught.
* Fsspec connector had the parent `update_source_metadata` misnamed as
`update_source_metadata_metadata` so it was never being called.
### Flow Diagram
Fixes
docker exec unstructured-smoke-test /bin/bash -c
/home/notebook-user/test_unstructured_ingest/test-ingest-wikipedia.sh
/home/notebook-user/test_unstructured_ingest/test-ingest-wikipedia.sh:
line 10: python: command not found
in
https://github.com/Unstructured-IO/unstructured/blob/6ad4971/scripts/docker-smoke-test.sh#L43
that was preventing docker images from being built.
## **Summary**
By adding hierarchy to unstructured elements, users will have more
information for implementing vector db/LLM chunking strategies. For
example, text elements could be queried by their preceding title
element. The hierarchy is implemented by a parent_id tag in the
element's metadata.
### Features
- Introduces a parent_id to ElementMetadata (The id of the parent
element, not a pointer)
- Creates a rule set for assigning hierarchies. Sensible default is
assigned, with an optional override parameter
- Sets element parent ids if there isn't an existing parent id or
matches the ruleset
### How it works
Hierarchies are assigned via a parent id field in element metadata.
Elements are read sequentially and evaluated against a ruleset. For
example take the following elements:
1. Title, "This is the Title"
2. Text, "this is the text"
And the ruleset: `{"title": ["text"]}`. When evaluated, the parent_id of
2 will be the id of 1. The algorithm for determining this is more
complex and resolves several edge cases, so please read the code for
further details.
### Schema Changes
```
@dataclass
class ElementMetadata:
coordinates: Optional[CoordinatesMetadata] = None
data_source: Optional[DataSourceMetadata] = None
filename: Optional[str] = None
file_directory: Optional[str] = None
last_modified: Optional[str] = None
filetype: Optional[str] = None
attached_to_filename: Optional[str] = None
+ parent_id: Optional[Union[str, uuid.UUID, NoID, UUID]] = None
+ category_depth: Optional[int] = None
...
```
### Testing
```
from unstructured.partition.auto import partition
from typing import List
elements = partition(filename="./unstructured/example-docs/fake-html.html", strategy="auto")
for element in elements:
print(
f"Category: {getattr(element, 'category', '')}\n"\
f"Text: {getattr(element, 'text', '')}\n"
f"ID: {element.id}\n" \
f"Parent ID: {element.metadata.parent_id}\n"\
f"Depth: {element.metadata.category_depth}\n" \
)
```
### Additional Notes
Implementing this feature revealed a possibly undesired side-effect in
how element metadata are processed. In
`unstructured/partition/common.py` the `_add_element_metadata` is
invoked as part of the `add_metadata_with_filetype` decorator for
filetype partitioning. This method is intended to add additional
information to the metadata generated with the element including
filename and filetype, however the existing metadata is merged into a
newly created metadata object rather than the other way around. Because
of the way it's structured, new metadata fields can easily be forgotten
and pose debugging challenges to developers. This likely warrants a new
issue.
I'm guessing that the implementation is done this way to avoid issues
with deserializing elements, but could be wrong.
---------
Co-authored-by: Benjamin Torres <benjats07@users.noreply.github.com>
### Description
Update all other connectors to use the new downstream architecture that
was recently introduced for the s3 connector.
Closes#1313 and #1311
If a layout model is used from unstructured-inference, you get back
class probabilities in the element metadata from partition.
extra-pdf-image-in in requirements already has the newest version of
unstructured-inference in there without a pinned version. Is there any
place else that the unstructured-inference version needs to be updated
to the required release version, 0.5.22?
* Pull out s3 code as subcommand
* Pull out dropbox code as subcommand
* Pull out azure code as subcommand
* Pull out fsspec code as subcommand
* Pull out github code as subcommand
* Pull out gitlab code as subcommand
* Pull out reddit code as subcommand
* Pull out slack code as subcommand
* Pull out discord code as subcommand
* Pull out wikipedia code as subcommand
* Pull out gdrive code as subcommand
* Pull out biomed code as subcommand
* rename parameters
* Pull out onedrive code as subcommand
* Pull out outlook code as subcommand
* Pull out local code as subcommand
* Pull out elasticsearch code as subcommand
* Pull out confluence code as subcommand
* Drop previous main file
* update changelog
* Add back in mp.Pool
* Fix mypy issues with click
* Make sure all tests run with verbose flag
* refactor approach to dynamically add common options to each subcommand, scrub logging of options for sensitive data
* Pull out some more shared options
* Support running code via python as well as cli
* update ingest readme and move it to the ingest folder
* update usage in connector docs
* move local command arg in test
* Seperate out cli code from logic running unstructured
* Make some cli fields required rather than optional
* rename process -> processor
* Improve logger to avoid duplicate handlers
---------
Co-authored-by: Ryan Nikolaidis <1208590+ryannikolaidis@users.noreply.github.com>
- Adds reusable validation scripts (check-x.sh) to minimize repeated (or near-repeated) code and create one source of truth
- Restructures the location of download and output folders such that they are nested in the test_unstructured_ingest directory
- Adds gitignore for output folders / files to avoid them accidentally getting checked into the repository
- Construct paths as reusable variables declared at top of scripts
- Sort order of flag for ingest calls, across all tests (this makes it easier to parse at a glance)
- OVERWRITE_FIXTURES removes all old fixtures for path to guarantee no stale results are left behind
- Bonus: don't check/exit on expected number of expected outputs when OVERWRITE_FIXTURES is true
- Bonus: exclude file_directory from Slack and Discord test scripts (match convention in all others)
Summary
* Updates "fast" PDF output element ordering to be consistent across Python versions by using the X,Y coordinates of elements extracted
* Added PDFs ingest tests with fast strategy with new script ./test_unstructured_ingest/test-ingest-pdf-fast-reprocess.sh
Updated ingest tests procedure:
* Processing files with hi_res strategy, and preserve downloads to repo files-ingest-download/<ingest_test_name>
* Reprocessing all PDFs with fast strategy from local file files-ingest-download, the partition outputs are stored at expected-structured-output/pdf-fast-reprocess/<ingest_test_name>
Test
* Reproduce tests with ./scripts/ingest-test-fixtures-update.sh , should expect no update. Also don't need any secret tokens since relevant tests won't produce PDFs.
* change strategy arg defalut to auto in partition
* passing --partition-strategy down
* add strategy="hi_res" to test (default changed)
* made an error on param name, added note
Previously, if there was an error (non-zero exit code) in an ingest test script,
the script would still complete and echo a warning about mismatched outputs
and how to regenerate the fixtures. However, this statement is irrelevant and
misleading: if the ingest failed with a non-zero exit code in the first place,
that is the failure that should be debugged -- don't confuse the user with
a comment about outputs.
- Updates CI to install tesseract version 5.3.0 (better than 4.x in various ways incl. perf.).
- Adds azure expected output fixtures for more useful reference points and as a repro for Some PDF's with scanned images return empty elements #346 .
- Adds a script to regenerate ingest test fixtures that is run in an ubuntu docker container (like CI), with the same version of tesseract. See the comments in scripts/ingest-test-fixtures-update.sh for details.
- Updates expected outputs with above script.
- Updates individual test-ingest scripts to update expected .json output if OVERWRITE_FIXTURES=true.
Closes#200. Fixes the failing test for label_studio_sdk>0.0.17 using the suggestion found in this comment. The vcr fixture on the test needed allow_playback_repeats=True. Unpinned label_studio_sdk and pip-compiled.
Update versions of dependencies, including unpinning the unstructured-inference dependency that's causing conflicts in repos like pipeline-oer that want the newer version.
Add GitLab data connector for ingest.
Involves more general Git functionality that is shared between the GitHub and GitLab data connectors.
Prevent code duplication for functionality between GitHub and GitLab ingest connectors.
Renamed github-access-token, github-branch and github-file-glob to git-access-token, git-branch and git-file-glob, respectively.
These work for GitHub and GitLab.
