terbium / Project
Changelog and status
terbium release history from 0.1.0 to 2.0.0: PyPI uploads, the 0.10.0 and 2.0.0 GitHub releases, and the current project status.
On this page
- Status
- 2.0.0 (2026-10-26)
- 0.10.0 (2026-09-30, GitHub release)
- 0.9.7 (PyPI, July 16, 2026)
- 0.9.6 (PyPI, July 15, 2026)
- 0.9.5 (PyPI, July 15, 2026)
- 0.9.4 (PyPI, July 9, 2026)
- 0.9.3 (not on PyPI)
- 0.9.2 (PyPI, July 3, 2026)
- 0.9.1 (PyPI, July 3, 2026)
- 0.9.0 (not on PyPI)
- 0.8.0 (not on PyPI)
- 0.7.0 (not on PyPI)
- 0.6.0 (not on PyPI)
- 0.5.0 (not on PyPI)
- 0.4.0 (not on PyPI)
- 0.3.0 (PyPI, July 3, 2026)
- 0.2.0 (PyPI, July 3, 2026)
- 0.1.0 (PyPI, July 3, 2026)
- Next
Status
- Development status: Alpha (
Development Status :: 3 - Alphainpyproject.toml). - Latest version: 2.0.0, released on 2026-10-26 on the Python Package Index (PyPI) and as a GitHub release.
- PyPI (
terbium-parse) has 0.1.0, 0.2.0, 0.3.0, 0.9.1, 0.9.2, 0.9.4, 0.9.5, 0.9.6, 0.9.7 and 2.0.0. 0.10.0 was published on GitHub only, so on PyPI 2.0.0 follows 0.9.7. - GitHub releases: v0.10.0 and v2.0.0, each with a wheel and a source distribution (sdist) attached. v0.10.0 is the first tagged release.
- The repository’s CHANGELOG.md covers 0.10.0 and 2.0.0. Entries before 0.10.0 are rebuilt from commit messages and PyPI upload dates.
2.0.0 (2026-10-26)
The first PyPI release since 0.9.7, published on PyPI and as a GitHub release. 0.10.0 was published on GitHub only, so pip install terbium-parse now brings everything listed under 0.10.0 as well as the fixes below. 2.0.0 follows 0.10.0 directly: the jump in the version number is not an application programming interface (API) break, and code written for 0.10.0 runs unchanged. The release has 51 passing tests.
- Fix: no
fitzwarning. With PyMuPDF 2.0, every import and every command run printedFutureWarning: The fitz API is deprecated. terbium now imports PyMuPDF aspymupdf, and falls back tofitzonly on PyMuPDF 1.24.0 to 1.24.2, which predate that name. - Fix: resume lines that mention a section word are kept. A line such as “Led a team with deep Python experience” was read as a heading: it was dropped from the content and started a new section. A heading is now a line that is exactly a section word, or a line of at most four words in which a word starts with a section word or phrase, such as “Work Experience”, “Academics” or “Educational Background”. A section word inside another word, as in “Inexperienced”, no longer counts, and bulleted lines are always content. The trade-off: a heading of five words or more, such as “Education, Training and Professional Development”, is now a content line of the section before it.
- Fix: image manifest colour.
dominant_colorno longer usesImage.getdata, which Pillow 12.1 deprecates and Pillow 14 removes. On Pillow 14 the colour would have come back empty with no error. - Documentation. The README leads with the terbium app tile, names NodeMaven as primary sponsor, and no longer says that PyPI serves 0.9.7. Its links to repository files are absolute, so they work on PyPI. The
parseandbuild_catalogdocstrings and the developer notes indocs/now match the code.
Known issues: the 0.10.0 known issues below still apply. The changelog also lists these, each described in Troubleshooting and limits:
build_catalog(..., ocr=True)without Tesseract prints the advice for an optical character recognition (OCR) pass that already ran, so it points at the vision lane instead of telling you to install Tesseract. Leaveocr="auto".build_catalog(..., ai=True)skips the artificial intelligence (AI) fill for photo rows; passterbium.AI().- A phone number written
(555) 123-4567comes out as555) 123-4567. - When an invoice or receipt is read from its page layout rather than a table, a label line that ends in digits, such as
Receipt No: R-8891orDate: 03/22/2024, also becomes a line item, and the total can appear as twototalrecords. - A resume line of four words or fewer that names a section is still read as a heading, so
Skills: Python, SQLstarts the skills section and the text after the colon is dropped, and a short line such as “Experienced Python developer” starts the experience section.
0.10.0 (2026-09-30, GitHub release)
The first tagged release, published on GitHub only, with a wheel and an sdist attached; PyPI went from 0.9.7 to 2.0.0. The release has 46 passing tests. It includes the work merged in pull requests #11 (July 26, 2026) and #12 (July 27, 2026), plus the fixes listed below.
- Document classification (
classify.py): after adapting a file, terbium detectscatalog,lookbook,transaction,resume,table,deckorunknownand picks the schema.--typeanddoc_type=override it. - Transaction schema for invoices, bills, receipts, purchase orders and quotes, with key-value form extraction and summary records.
- Resume schema with a candidate header and sectioned records.
- Universal output: comma-separated values (CSV), self-contained HyperText Markup Language (HTML) and terminal tables for every lane, with fallback chains that never raise. New command line flags
--type,--htmland--open. The command now auto-classifies and routes invoices and resumes to the records view. - Image adapter for PNG, JPG, JPEG, WEBP and TIFF through the Tesseract OCR bridge.
- Picture-heavy catalogues: per-page lookbook detection, proximity-scored captions (below, right, above), and support for dense grids and one-photo pages. The catalog table gains a Dimensions column.
- Transaction AI lane modelled on the catalog AI fill. See Invoices, receipts and resumes.
- More AI providers: GPT (OpenAI), Kimi (Moonshot) and Grok (xAI), all through the OpenAI-compatible client, with new extras
openai,kimiandgrok. Claude stays the default when several keys are set;AI(provider=...)pins one. - New exports:
terbium.classify,terbium.records_to_csv,terbium.render_html,terbium.render_terminal_table. - Fix: invoice and receipt AI step. It now runs. It fills missing summary fields and adds line items only when the parser found none.
- Fix:
used_ai. It isTrueonly when a model call returned a response. - Fix: terminal tables. Header cells are truncated and the columns line up.
- Fix: Tesseract message. The catalog message without Tesseract now tells you to install it, or to pass
ai=terbium.AI(...)for the vision lane. - Fix: OpenAI o-series models. They get
max_completion_tokens. Theopenai,kimi,grokandaiextras needopenai>=1.45.
Known issues that remain: image input to o3-mini is unverified, model identifiers (IDs) are not verified against the provider application programming interfaces (APIs), and the Gemini lane uses google-generativeai.
0.9.7 (PyPI, July 16, 2026)
- Catalog: rate prices such as
2.25/SQFTare no longer taken as stock keeping unit (SKU) codes; product names fall back to a caption line when there is no heading. - Image extraction: thumbnail sheets retry with swatch-sized images when the photo pass finds nothing.
0.9.6 (PyPI, July 15, 2026)
- Release to publish the tracked sponsor link in the PyPI project description.
0.9.5 (PyPI, July 15, 2026)
- README: NodeMaven sponsor section.
0.9.4 (PyPI, July 9, 2026)
- Local OCR lane for image-only pages through Tesseract.
- PowerPoint (PPTX) image anchoring: each slide picture becomes a catalog row.
- Catalogue escalation only when data is genuinely unread.
0.9.3 (not on PyPI)
- Catalog escalation message for half-blank catalog tables.
0.9.2 (PyPI, July 3, 2026)
terbium <file>prints the catalog table and writes CSV, with no AI.
0.9.1 (PyPI, July 3, 2026)
- Stricter SKU detection and validation on real vendor catalogues in six categories.
- The cross-category benchmark (
eval/category_bench.py), name, SKU, materials and image 24 of 24.
0.9.0 (not on PyPI)
- The catalog table: name, SKU, materials and image per product, plus CSV output.
0.8.0 (not on PyPI)
- Feed exporters: Shopify CSV and product information management (PIM) JavaScript Object Notation (JSON).
0.7.0 (not on PyPI)
- Deterministic normalization of dimensions, colours and materials.
0.6.0 (not on PyPI)
- AI product enrichment layer: implicit attributes and vision reads.
0.5.0 (not on PyPI)
- Universal, category-agnostic
productschema.
0.4.0 (not on PyPI)
- Content-agnostic table detector; the generic detector becomes the default.
0.3.0 (PyPI, July 3, 2026)
- Image extraction: product photos exported and named after the product.
0.2.0 (PyPI, July 3, 2026)
- Label-grid extraction for lookbooks and reporting of image-only pages.
0.1.0 (PyPI, July 3, 2026)
- First release: algorithmic parser for Portable Document Format (PDF), PPTX, Excel (XLSX) and CSV files with confidence scoring and opt-in AI escalation.
- Private-use icon glyphs are stripped from PDF text.
Next
The repository roadmap (docs/ROADMAP.md) lists one open phase: publishing a public benchmark and golden sets per product category. It has no date.