Skip to content

Changelog

All notable changes to this project are documented here. The format is based on Keep a Changelog and this project adheres to Semantic Versioning.

Starting with v0.2.0, each release will be documented with granular Added / Changed / Fixed / Removed / Security sections. v0.1.0 is the single foundational cut that establishes the baseline.

[0.1.3] — 2026-05-26

Fix : preserve merge span on header cells (HintKind::TableHeader now carries rowspan + colspan).

Fixed

  • Merged header cells now surface their geometric span faithfully. Prior to v0.1.3, HintKind::TableHeader only carried col: u32 ; the merge span was computed internally by the XLSX decoder (SheetContext::merge_span) and the non-origin cells were correctly skipped (is_merged_non_origin), but the public api::TableCell.rowspan / colspan for header cells were hardcoded to 1 regardless of the actual merge geometry. Downstream consumers walking Document::tables() saw (0,0) "Identité" colspan=1, (0,2) "Tarification" colspan=1, (0,4) "Stock" colspan=1 for a row-0 header band that merged (0,0)..(0,1) + (0,2)..(0,3) — the column gaps signaled the merges but the geometric span was lost at the API frontier. This made it harder to write robust consumers for XLSX with category-level header groupings (a common real-world pattern).

Changed

  • HintKind::TableHeader extended from { col: u32 } to { col: u32, rowspan: u32, colspan: u32 }. The two new fields default to 1 for the non-merged case and reflect the source document's geometric merge span for merged origins. Source-incompatible for consumers destructuring all fields (existing consumers using matches!(.., HintKind::TableHeader { col: N, .. }) keep working ; consumers using { col } need to update to { col, .. } or { col, rowspan, colspan }).
  • All 5 emit sites updated to thread the rowspan/colspan through :
  • formats/xlsx/sheets/emit.rs (heuristic header path + native_table_hint)
  • formats/pdf/tagged/semantics.rs (passes cell.rowspan / cell.colspan from active cell)
  • formats/html/walker/structures.rs (passes the parsed <th rowspan=… colspan=…> attributes)
  • formats/docx/body/tables.rs (passes cell.rowspan / cell.colspan from active cell)
  • structure/detectors/table/lattice/detect.rs (defaults to 1 / 1 — lattice-detected tables don't carry merge geometry from heuristics)
  • structure/assembler/classify.rs propagates the new fields into PrimaryRole::TableCell { rowspan, colspan, is_header: true, ... } instead of hardcoding 1 / 1.
  • structure/detectors/table/continuation.rs::promote_to_continuation preserves the span when promoting TableHeader → TableCellContinuation (instead of dropping to 1 / 1).

Downstream impact

Aliya's crates/olga-bridge (Chunk 3.1, ADR-0068 D9.6) discovered this gap empirically while exercising the messy-real-world resilience fixture : the bridge re-export surface olga_bridge::TableCell showed max_colspan = 1 even when the input xlsx had merge_cells("Identité", colspan=2). The fix unblocks the consumer-side T pipeline at Chunk 5.1 onboarding — instead of inferring merges by gap detection, the T-pipeline dispatches directly on TableCell.colspan > 1.

0.1.2 — 2026-05-25

Per-format feature flags — opt-out binary size for downstream consumers (e.g. mobile surfaces).

Added

  • [features] block in Cargo.toml exposing 4 per-format feature flags : xlsx / pdf / docx / html. Default feature set is ["xlsx", "pdf", "docx", "html"] — full backward compatibility with v0.1.1 behavior (no change for consumers who don't opt out).
  • A private umbrella _ooxml feature that pulls in the shared quick-xml + zip deps that both DOCX and XLSX need ; gated transitively by xlsx and docx features. Consumers should NOT enable _ooxml directly.
  • Compile-time feature-disabled error path : when a format's feature is opted out, Document::open() and the CLI select_decoder() dispatch return IdpError::UnsupportedFormat carrying a "feature disabled at compile time" message. The Format enum keeps all 4 variants visible so the public API stays stable across feature combinations.

Changed

  • Format-specific dependencies are now optional = true in Cargo.toml :
  • calamine → gated by xlsx
  • pdf_oxide → gated by pdf
  • scraper + ego-tree → gated by html
  • quick-xml + zip → gated by _ooxml (transitively by xlsx and docx)
  • The PDF spatial renderer modules (output::markdown + output::spatial + output::rules + helpers col_resolve / row_cluster) are now cfg-gated on the pdf feature. These modules consume pdf_oxide::layout::TextChar directly and have no meaningful rendering path without PDF support.
  • Integration test files (tests/docx_tests.rs / tests/pdf_tests.rs / tests/xlsx_tests.rs / tests/html_tests.rs) gated with #![cfg(feature = "X")] — disabled formats compile to an empty test-binary shell with zero tests.

Compatibility

Fully backward-compatible with v0.1.1 for consumers using olga = "0.1" without explicit feature manipulation : the default feature set replicates v0.1.1's all-formats behavior verbatim. Consumers opting out of formats (e.g. default-features = false, features = ["xlsx", "html"]) gain measurable binary-size reduction.

Downstream impact

Aliya's crates/olga-bridge (Chunk 3.1, ADR-0067) consumes this release to support its mobile-surface feature-flag discipline. The bridge propagates the 4 feature flags one-to-one so Aliya mobile builds can opt out of PDF + DOCX while keeping XLSX + HTML for the dominant onboarding flow (Excel catalog import + product-page web scraping).

0.1.1 — 2026-04-21

Docs patch — surface the independent v0.1.0 benchmark on PyPI and crates.io.

No engine changes. This release publishes the independent, reproducible post-release audit at olga_v0.1.0_benchmark/ and links it from the crate README, the PyPI README, the MkDocs landing page, and BENCHMARKS.md. Headline result on a 50-file mixed-format corpus: 1.62× faster and 2.62× more extracted content than a hand-routed best-of-breed pipeline. The crate metadata author field is also corrected from "Hugues Tankouo" to "Hugues Dtankouo".

0.1.0 — 2026-04-21

First public release — the end-to-end Olga pipeline in one cut.

Olga's first public release. This foundational cut ships the full intelligent document processing pipeline: a Rust core that parses PDF, DOCX, XLSX, and HTML with provenance tracking and table reconstruction, a Python distribution (olgadoc) with a strictly-typed API surface, an olga CLI for inspection, extraction, search, and page-level access, runnable examples, end-to-end regression coverage, an MkDocs site, and a full CI/CD pipeline publishing to crates.io and PyPI. The public API is stable enough for evaluation and prototyping; expect minor breaking changes on the path to 1.0.