Garbage In, Garbage Out

Retrieval-augmented generation lives or dies on what enters the index. Yet most RAG stacks pour their engineering into retrieval and generation and treat ingestion as a solved problem — strip the text, drop the layout, discard the page, move on. The hardest documents and the most decision-relevant context are lost before a single query runs. GoodMem is built the other way around — it does the hard work at ingestion, where the quality of every later answer is decided. Enterprise deployments add on-prem document OCR that keeps every page inside your network while parsing within reach of the best frontier models, and page-faithful capture — a rendered image of every page, across PDF and all three Microsoft Office formats, Excel included — so answers can be checked against the source and read directly by multimodal models.

Retrieval can't recover what ingestion discards

A RAG system is judged on retrieval and generation, but both inherit a hard ceiling from ingestion: a model can only retrieve and reason over what was captured in the first place. The default pipeline — extract a stream of text, split into chunks, embed — quietly throws away three things that enterprise documents depend on. Layout: which row belongs to which header, which number to which column, the reading order of a multi-column page. Visual context: the rendered page a human uses to trust an answer. And the hardest documents entirely: scanned contracts, image-only PDFs, dense financial tables, slide decks that carry their meaning in their composition. What is dropped at ingestion cannot be retrieved at query time, no matter how good the embeddings.

On-prem OCR: frontier-class parsing, nothing leaves the building

For visual and scanned documents, GoodMem's enterprise deployments run document OCR locally, on a self-hosted model that stays on your own hardware. Parsing is layout-aware: titles, tables, formulas, and reading order are detected in a single pass, with tables emitted as HTML and formulas as LaTeX rather than flattened into a text blob. Two properties make it enterprise-grade. First, it is on-prem: the model runs behind your firewall, images are passed to a local endpoint, and no document or page is sent to a third-party cloud OCR service. Second, it is accurate: on recent public document-parsing benchmarks it ranks just behind the leading frontier system and ahead of every other model tested (Figure 1) — frontier-class quality without frontier-cloud data exposure.

Pages, not just text — and Excel, not just PDF

Beyond text, GoodMem captures a rendered image of every page at ingestion: a faithful raster of the page exactly as a person would see it. This is rare — most ingestion keeps only extracted text, or at best crops of embedded figures — and it pays off twice. Verification: a retrieved answer can be shown against its actual source page, so a user sees where it came from and trusts it. Multimodal reading: the preserved page image can be handed directly to a multimodal model, which often pulls more from the rendered page — a chart, a stamped signature, a merged-cell table — than any text chunk carries. The differentiator is breadth and method: GoodMem renders page images not only for PDF but for all three Microsoft Office formats — Word, PowerPoint, and Excel — and does it in pure Java, in process, with no LibreOffice or external office engine in the deployment. The few ingestion products that render Office pages at all shell out to a headless LibreOffice install and commonly stop short of spreadsheets — one prominent multimodal RAG engine renders Word and PowerPoint but deliberately excludes Excel. We are not aware of another ingestion product that returns page images for Excel specifically: the densest, most table-heavy, and least-served format in the enterprise.

Ingestion capability, side by side

Why PAIR Systems

GoodMem is built by PAIR Systems, which develops the information-retrieval layer for enterprise AI — bringing techniques once limited to specialized search teams into any product as self-hostable software. Its founder helped pioneer practical zero-shot neural retrieval at Google and productized enterprise retrieval at Vectara. For the heaviest token spenders, GoodMem extends into model distillation and self-tuning that compound the gains shown here.