Searching Documents as Pages: What NeoMME Changes for Retrieval

Hcompany introduced NeoMME on September 3, 2026, describing multilingual multimodal encoders in 260-million and 800-million-parameter sizes. The architecture processes text and image patches through a shared bidirectional Transformer rather than assembling a separate pretrained vision tower and generative decoder. Its retrieval variants produce dense and late-interaction representations, and the release includes Apache 2.0 checkpoints.

For document applications, the interesting question is not whether an encoder can replace every component of a retrieval system. It is whether preserving the visual page helps find evidence that text extraction misses. A diagram, footnote or relationship between table columns can be important even when all the individual words were extracted correctly.

Recognise when layout carries meaning

Imagine searching an illustrated equipment manual for a connector location. A text-only index may retrieve a page mentioning the connector, but the useful answer depends on an arrow pointing to the correct part. Similarly, a financial table may contain the right number several times while its meaning depends on the adjacent heading and reporting period.

A page-image approach gives a retrieval model access to these visual relationships. That does not eliminate the need for readable source documents or accurate downstream interpretation. It changes what information is available during the search stage. Poor scans, rotated pages and tiny labels remain conditions worth testing rather than assuming away.

Create a small collection of representative questions before selecting a model. Include ordinary prose queries as well as layout-dependent questions. Otherwise an evaluation built entirely around attractive diagrams could overstate the value of visual retrieval for a corpus that mostly contains straightforward text.

Separate finding evidence from answering questions

First evaluate the ranked pages without generating an answer. For each query, mark which pages contain enough information to answer it and which merely discuss the same subject. Measure how often sufficient evidence appears near the top, and inspect the cases where it does not appear at all.

This separation makes failures easier to diagnose. If the correct page was never retrieved, changing the answer-writing prompt is unlikely to fix the root problem. If the page was retrieved but the answer misunderstood a table, the retrieval model should not receive all the blame. Each stage needs its own observable inputs and outputs.

Allow questions with no answer in the collection. A useful search system should not present a vaguely related page as definitive evidence. In an evaluation interface, let reviewers identify insufficient evidence explicitly. That creates a basis for later abstention behaviour rather than training the entire application to produce an answer at any cost.

Understand the indexing tradeoff

A dense representation compresses a page into a compact vector for efficient comparison. Late-interaction approaches preserve multiple representations and perform a more detailed comparison with the query. Conceptually, the latter can retain local matches that a single summary vector smooths away, but it also changes storage and search costs.

A practical design can use an inexpensive first stage to gather candidates and a more detailed stage to reorder them. Evaluate the whole pipeline rather than assuming a more sophisticated final stage automatically helps. If the first stage omits the relevant page, later scoring cannot recover evidence it never receives.

Measure index size, indexing time and query latency using the page resolutions you actually intend to store. A collection of short clean documents is different from a large archive of high-resolution scans. Include the cost of updating or removing documents, since an index that is cheap to build once may still be awkward to maintain.

Design a grounded reading experience

Show users where an answer came from. A page citation should open the actual document at a useful location, not merely display a filename. Where possible, keep enough surrounding context visible for a reader to check whether the evidence supports the claim. This is especially important when several versions of a document contain similar wording.

Version the corpus along with the retrieval configuration. If a manual changes, record which edition was indexed and when. A technically accurate answer from an obsolete page can be operationally wrong. Access controls must also be enforced before results are shown, not treated as a cosmetic filter after answer generation.

NeoMME is a reason to revisit which parts of a document your search system can see. The adoption decision should still rest on a measured improvement in evidence retrieval, sensible operating costs and a reading experience that lets users verify the result for themselves.