Skip to main content
As soon as a document lands in your deal, Plomo’s classification pipeline gets to work. It reads the file — whether that’s a PDF, a Word document, a spreadsheet, or a scanned image — identifies what the document is, and maps it to the most appropriate category and subcategory in your deal taxonomy. You see results stream in as each file is processed, without waiting for the entire batch to finish.

​
What a classification produces

Every document receives a classification record that contains the following fields:
  • Category — the top-level taxonomy grouping the document belongs to (e.g. Financial Overview)
  • Subcategory — the mid-level grouping within that category (e.g. Financial Statements)
  • Confidence score — a calibrated probability between 0 and 1 indicating how certain the model is
  • Evidence — specific text spans or page regions from the document that support the classification, so you can verify the AI’s reasoning
  • Review status — whether the classification was auto-accepted, needs your review, or is uncategorized

​
Confidence thresholds and review status

Not every classification is equally certain. Plomo uses the confidence score to decide what happens next:
The confidence score is calibrated, meaning a score of 0.70 genuinely corresponds to roughly 70% likelihood of a correct classification. It is not an arbitrary internal score.

​
Two classification modes

Plomo applies different analysis strategies depending on the type of file it’s processing.

Text mode

Works with all supported file types (PDF, Word, Excel, PowerPoint, plain text, CSV). Plomo extracts the document’s text content, combines it with the deal taxonomy, and sends both to the AI. This is the default path for most documents.

Vision mode

PDF-only. Instead of extracting text, Plomo renders each page as an image and passes those images to a multimodal AI model. This is particularly effective for scanned documents, tables, charts, and PDFs where the visual layout carries meaning that plain text extraction would lose.
Plomo selects the appropriate mode automatically based on the file type and content characteristics. For PDFs, it may use both modes and combine the results.

​
Real-time streaming

Classifications stream to your dashboard as each document finishes processing. You don’t need to wait for an entire upload batch to complete before you can start reviewing results. The coverage tracker and document list update live as classifications arrive.

​
Recovery and robustness

The first classification pass is deterministic — given the same document and taxonomy, the model produces the same initial output. If that output is malformed or fails a validation check, Plomo runs bounded recovery rounds: a small number of additional attempts where results are combined using confidence-weighted voting to reach a final answer. This means a single bad model response won’t leave a document permanently uncategorized.

​
Manual overrides

You have full control over every classification. If the AI assigns a document to the wrong category, select the document in your deal and choose Re-classify to pick a different category and subcategory yourself. Manual overrides are recorded and will be reflected in coverage tracking immediately.

​
Improving accuracy with labeled examples

If your team has completed previous deals with Plomo, you can provide labeled examples from those engagements to improve classification accuracy on future deals with similar taxonomies. Plomo uses these examples during a pipeline optimization step to tune prompts and scoring without changing the underlying model. Talk to your account team if you’re interested in setting this up for your organization.
Your documents are never used to train or fine-tune any foundation model. Classification is performed using your documents as input at inference time only — they remain private to your deal and your organization.