RAG 與文件處理
Docling 2.122 Brings Compact Table Output to the Conversion API and Restores Legacy iWork Document Structure
Docling 2.122.0 allows batch conversions to directly output Markdown tables without column-padding whitespace and begins restoring headings, hierarchy, and tables from iWork ’09 files. The release also fixes OCR, PDF hyphenation, and layout-overlap issues, although the project has not quantified changes in performance or extraction accuracy.
The open-source document conversion tool Docling released version 2.122.0 on August 25. The change with the most direct impact on RAG pipelines is the addition of `compact_tables` to `ConvertDocumentsOptions`. Compact Markdown tables were already available through the underlying serialization interface, but they can now be configured consistently through the document conversion entry point. When enabled, the option removes only the padding spaces used for column alignment; it does not change table detection or cell-matching results. Its primary benefit is reducing meaningless tokens before content is sent to embedding models, rerankers, or LLMs. A lower token count should not be interpreted as improved extraction quality.
The second change targets Apple documents. The updated iWork backend can restore titles, section headings, and tables from iWork ’09 files, expanding the structural information preserved when legacy Pages documents are converted into the unified `DoclingDocument` representation. However, the current “Supported formats” documentation still states that Pages extraction is limited to body text and cannot yet restore headings, lists, or tables, indicating a temporary mismatch between the documentation and implementation. Adopters should rely on actual 2.122.0 output and regression samples, with particular attention to the fact that the new version promises titles, heading hierarchy, and tables only; it does not claim to restore all list and layout semantics.
The remaining fixes focus on details that can easily contaminate retrieval corpora: preserving hyphens at PDF line endings when they do not represent word breaks; skipping native segmented-page decoding in full-page OCR mode; strengthening overlap cleanup across different element types during layout post-processing; using stable node identification for DOCX text boxes; and accelerating the matching of table cells to layout elements. These changes may alter existing chunk content and hashes. Even though the API introduces no breaking changes, teams should rerun their golden documents after upgrading and compare element counts, table structures, and citation locations. The release notes provide no figures for speed, token savings, or accuracy, so production benefits must be measured independently.