tryworks

Turn PDFs, Word, PowerPoint, Excel, HTML, Markdown, email and text into LLM-ready elements, with zero required dependencies.

tryworks implements the document-partitioning call path of unstructured: the same partition() function, the same element types and metadata, the same chunking, the same JSON. It leaves out what most pipelines never call, namely OCR, layout models and ingest connectors, and the 2 GB of dependencies they bring.

from tryworks.partition.auto import partition
from tryworks.chunking.title import chunk_by_title

elements = partition("quarterly-report.docx")
chunks = chunk_by_title(elements, max_characters=1000)

for chunk in chunks:
    print(chunk.metadata.page_number, chunk.text[:80])

The name: tryworks were the brick furnaces on whaling ships that rendered raw blubber into oil.

Why

Measured on Linux x86-64 with Python 3.12, reading the same document types:

Base tryworks without the PDF extra is a single 80 KB wheel with no dependencies.

In practice that means:

• It fits in AWS Lambda. The zip deployment limit is 250 MB unpacked, and it cannot be raised.
• Cold starts are fast, and since August 2025 AWS bills the Lambda init phase.
• Security review is short. Two packages to approve, pin and scan, where unstructured needs 149. Air-gapped and regulated environments review each dependency individually.
• The supply-chain surface is small. Most advisories in a large tree come from its transitive dependencies, not from the library you asked for.

Numbers come from scripts/measure.py (exact resolution with uv pip compile, advisory counts from OSV.dev) and from real installs; see Reproducing the numbers. Neither tree had an open advisory against the versions resolved when this was measured; the 701 are historical, which is what a scanner and a triage process work through over time.

Install

pip install "tryworks @ git+https://github.com/ezzcodeezzlife/tryworks"
pip install "tryworks[pdf] @ git+https://github.com/ezzcodeezzlife/tryworks"   # PDF support

Python 3.10 or newer. PDF support adds pypdfium2, bindings to PDFium, the PDF engine used in Chrome. Everything else uses the standard library.

Use

Partition

from tryworks.partition.auto import partition

elements = partition("report.pdf")                      # type from extension or content
elements = partition(file=open("memo.docx", "rb"))      # file objects work too
elements = partition(url="https://example.com/post")    # http(s) URLs

for el in elements:
    print(el.category, el.metadata.page_number, el.text)
# Title 1 Quarterly Operations Memo
# NarrativeText 1 Revenue grew eleven percent year over year, ...
# ListItem 1 Expand the Rotterdam warehouse
# Table 1 Warehouse totals Units Rotterdam KR-1001 1200 ...

Each format also has its own function with unstructured's signature: partition_pdf, partition_docx, partition_pptx, partition_xlsx, partition_html, partition_md, partition_text, partition_csv, partition_tsv, partition_email, partition_json.

Chunk

from tryworks.chunking.title import chunk_by_title
from tryworks.chunking.basic import chunk_elements

chunks = chunk_by_title(elements, max_characters=1000, overlap=100)
chunks = partition("report.pdf", chunking_strategy="by_title", max_characters=1000)

Sections start at each Title, small sections are combined, long text is split at newlines or spaces, and tables get chunks of their own. Tables too large for one chunk are split by row into TableChunk elements that keep valid HTML in metadata.text_as_html.

Serialize

from tryworks.staging.base import elements_to_json, elements_from_json, elements_to_md

json_text = elements_to_json(elements)         # same JSON as unstructured, readable by it
elements = elements_from_json(text=json_text)
markdown = elements_to_md(elements)

Command line

tryworks report.pdf                                   # JSON
tryworks memo.docx --format markdown
tryworks https://example.com/post --format text --chunking-strategy by_title --max-characters 800

Moving from unstructured

Change the package name in your imports:

- from unstructured.partition.auto import partition
- from unstructured.chunking.title import chunk_by_title
+ from tryworks.partition.auto import partition
+ from tryworks.chunking.title import chunk_by_title

For code you can't edit, such as a third-party library that imports unstructured itself, register tryworks under the unstructured import names before that code is imported:

import tryworks
tryworks.alias_as_unstructured()

This only affects imports in the current process. It installs nothing named unstructured.

Formats

Compatibility with unstructured

A generated corpus with one realistic document per format is partitioned by both libraries and compared element by element (report, script). A workflow repeats the comparison weekly against the latest unstructured release.

Against unstructured 0.27.5:

• 97% of upstream elements have a tryworks element with identical text,
• 97% of those have the same element type,
• 94% have the same element id (ids hash filename, text, page and position, so this also checks ordering),
• parent_id, filetype, page_number, page_name, link_urls, emphasized_text_contents, header_footer_type, sent_from and subject agree on every aligned element.

Every remaining difference in the report is one of the deliberate differences listed below.

Deliberate differences

Not included

• OCR and layout models. strategy="hi_res" and "ocr_only" raise NotImplementedError, and scanned PDFs without embedded text produce a warning and no elements. Use unstructured or an OCR service for those.
• Formats: images, audio, .doc, .ppt, .xls, .msg, .rtf, .odt, .epub, .rst, .org, .xml. Unsupported files raise UnsupportedFileFormatError, with a conversion hint.
• PDF table structure and image extraction. PDF tables come out as text blocks.
• Pictures in DOCX and PPTX are not emitted as Image elements.
• Token-based chunking (max_tokens) raises NotImplementedError; use max_characters.
• Ingest connectors, the hosted API client, and staging bricks for labeling tools.

Text classification uses the same rules as unstructured, but detects verbs with a compact lexicon instead of NLTK. On borderline text such as short fragments, a paragraph can come out as UncategorizedText where unstructured says NarrativeText, or the reverse.

Untrusted documents

Documents are parsed defensively: zip-bomb checks on Office files, rejection of XML entity declarations, and caps on spreadsheet cells, PDF pages, HTML nesting and chunk payloads, all adjustable through environment variables. Only http(s) URLs are fetched. See SECURITY.md.

Reproducing the numbers

# install size, packages and advisories (needs uv)
python scripts/measure.py "unstructured[docx,pptx,xlsx,md,csv,pdf]" "tryworks[pdf] @ ."

# element-by-element comparison
python tests/corpus.py compat/out/corpus
python scripts/compat.py dump tryworks compat/out/corpus compat/out/tryworks.json
python scripts/compat.py dump unstructured compat/out/corpus compat/out/upstream.json   # in an env with unstructured
python scripts/compat.py compare compat/out/upstream.json compat/out/tryworks.json --markdown compat/REPORT.md

Development

pip install -e ".[dev]"
python -m pytest

See CONTRIBUTING.md. The one hard rule: the base install stays dependency-free.

How this was built

tryworks was written with Claude Code under human direction. It is an independent implementation of unstructured's public API; no unstructured source code was copied. Confidence in it comes from things you can check yourself: the test suite, which runs on Linux, macOS and Windows, and the differential comparison against unstructured above.

License

Apache-2.0. See LICENSE and NOTICE. tryworks is not affiliated with or endorsed by Unstructured Technologies, Inc.