Most of what we process is financial PDFs, such as annual reports, factsheets, prospectuses, regulatory filings. In those documents it is important that a figure is connected to the row it sits in, the column heading above it, and the footnote that qualifies it. PrimeCut extracts tables according to the needs, so the LLM can give answers with the right context. It improves the search quality especially in similar document categories, so the information does not get mixed up, because it keeps the heading path and the footnote attached to the passage they belong to.

In terms of investor and earnings presentations, industry reports, which are full of charts, graphs and diagrams, our old pipeline either dropped those visuals or left a gap where the numbers should have been, so we kept retrieving the text around a chart while the data in the chart itself stayed out of reach. PrimeCut converts visuals into described, searchable text in the same chunk as the body copy, which significantly lifted our retrieval accuracy. It also allows us to reinject the original images extracted as base64 by their reference, if necessary. From a developer's side, integration itself was seamless. Review collected by and hosted on G2.com.

The documentation was extensive, though some areas could be clearer or easier to find, but maybe this is something to be expected from a startup. Review collected by and hosted on G2.com.