Unstructured Transform vs LlamaParse: Can annual-report answers link numbers back to the original table?
Teams with an existing element-based ingestion workflow can evaluate Transform first; teams needing data from multipage tables and precise positioning can start with LlamaParse. Compare output contracts, source-page relationships, and review signals rather than treating successful parsing or a high score as proof that an answer is correct.
When a research assistant answers “How much revenue did this region generate?” from an annual report, colleagues still need to be able to click through to the original table to check it. Both Unstructured Transform and LlamaParse can prepare complex documents for downstream retrieval, but comparing only how tidy the Markdown looks during procurement misses this step. Teams that already ingest content by element can first examine Transform’s elements, page numbers, and failed-page information; teams that need to pass multipage tables to data analysis or locate individual cells can prioritize evaluating LlamaParse’s table output. This selection guidance is based on official API documentation, not on a hands-on ranking of the two tools’ recognition accuracy.
First, establish which page the answer needs to link back to
Transform currently uses the v2 Parse API and outputs Markdown by default. According to the official documentation, explicitly requesting elements lets callers read each element’s type, text, and metadata.page_number; in that example, markdown is null. Connecting the default request and then expecting a complete set of element data automatically may therefore conflict with the response contract. If an existing vector database ingests paragraphs or table elements as units, preserve the links among the original file identifier, page number, and element before handling chunking and embedding.
Do not assume that position data is always available either. Transform element coordinates may be missing or null; coordinates for an element spanning pages may consist of several boxes, each with a page number. The citation interface should highlight evidence when coordinates are available and fall back to the original page when they are not, rather than inventing a position to fill out the interface. This fallback behavior belongs in procurement acceptance tests.
LlamaParse’s expand=items returns a tree of page elements organized in reading order, including page numbers and page dimensions, with different fields for tables, text, images, and other element types. Bounding boxes use page points, not pixels or normalized coordinates; different pages in a document can have different dimensions. A team drawing evidence locations on screenshots should convert coordinates using the dimensions of the corresponding page, rather than treating API coordinates directly as image pixels. It also allows bbox to be null, so the application still needs to preserve and check the complete citation chain.
A multipage table needs more than its first screenshot
Imagine a revenue table that spans two pages: the first contains region and year headers, while the second continues the data rows. Readers need a row’s value together with its year, currency, and source, not just a text block that looks like a table. Acceptance testing should use original files containing merged headers, continued tables, and footnotes to check whether headers still match the correct columns and whether the answer can locate the page that actually contains the value. This is a suggested test scenario, not an existing test result.
LlamaParse’s table documentation offers outputs such as rows, CSV, and HTML; HTML can represent merged cells, whereas Markdown pipe tables cannot. When merge_continued_tables is enabled, continued tables are merged onto the starting page, and merged_from_pages and merged_into_page can preserve the merge relationships. This is useful for teams passing tables to analysis programs, but an application that saves only the first page containing the merged table may lose source-page information. Choosing the format and preserving page-number relationships should be treated as a single integration decision.
For Transform, start by testing against the way elements are already consumed: can the actual Table elements, page numbers, and cross-page coordinates meet downstream ingestion and citation requirements? Do not treat examples from its older partition library as responses from the current Transform, or assume that your merged headers will be handled correctly just because a marketing page advertises table capabilities. If the final requirement is complete row-and-column data or precise highlighting, inspect the actual response using the same sample, rather than checking only that the returned string is nonempty.
Quality signals belong in the review queue
A completed result from Transform may still omit unreadable pages or include pages that were not fully recognized. The official documentation tells callers to check failed_pages, degraded_pages, and warnings. An element’s parse_confidence measures agreement between its text and OCR text, not the “probability that the content is correct”; null means unavailable, not a score of zero. Teams can use this information to organize review, but should not present a high score as direct evidence that a business fact has been verified.
LlamaParse’s table parse_concerns flags issues such as unusual row or column counts and mismatches between headers and value types. The official documentation also states that flags are signals for review: a flag does not prove that a table is wrong, and the absence of flags does not guarantee correctness. Page output also requires checking success. Both tools provide information that can feed a review workflow, but neither replaces checking against the original. Important amounts or quantities should still be checked by matching parsed results to source pages.
Pilots of both tools should therefore deliver the same set of results: searchable content, relationships between original files and source pages, passages requiring human review, and a list of omitted pages. Evaluate Transform first when an element-based ingestion contract is already in place and the team wants to retain it; evaluate LlamaParse first when the data format of multipage tables and fine-grained positioning are core requirements. Ultimately, compare which output helps the team lose less evidence and how much adaptation and review work it requires, rather than replacing the trustworthiness of an entire answer with a single confidence score.
Content note:This article is based on publicly available information and is intended as a reference for evaluating tools. Features, pricing, and related policies may change at any time; please refer to the latest official information.
Original source:docs.unstructured.io