How Documents Are Classified
Plomo reads each document and compares it against your deal’s categories. For each document, it determines:- Category — Which top-level group does this belong to?
- Subcategory — Which specific area within that group?
- Confidence — How certain is the classification? (0–100%)
- Evidence — What parts of the document support this classification?
Two Classification Modes
Text Mode
Works for all file types. Plomo extracts the text and sends it to the AI with your taxonomy using a chain-of-thought pipeline. The AI reasons step by step, scoring each category before making a decision. Long documents are intelligently trimmed to preserve the most relevant content while staying within model context limits.Vision Mode
For PDFs only. Plomo sends the actual PDF pages as images to the multimodal AI. This computer vision approach catches information in tables, charts, headers, and formatting that text extraction might miss — especially useful for financial statements, org charts, and scanned documents. Both modes feed into the same ensemble voting system to ensure accuracy.Ensemble Voting & Anomaly Detection
Rather than relying on a single AI call, Plomo runs multiple classification rounds with an ensemble approach:- Multi-round inference — Multiple classification attempts with diverse parameters
- Anomaly filtering — Suspicious responses are automatically removed
- Consensus voting — Final answer selected when consistent agreement is reached
Confidence and Review Status
After classification, each document gets a review status:| Confidence | Status | What It Means |
|---|---|---|
| 85%+ | Accepted | High confidence — auto-accepted, no action needed |
| 60–84% | Needs Review | Moderate confidence — take a quick look |
| Below 60% | Uncategorized | Low confidence — classify manually |
Real-Time Results
Classifications stream to the dashboard in real time over HTTP/2. As each document finishes processing, it appears in the table immediately — you don’t have to wait for the entire batch. Documents are parsed and classified in parallel across multiple worker threads.Pipeline Optimization (Opt-In)
If you provide labeled examples from past deals, the evolutionary optimizer can improve classification accuracy:- Takes your labeled examples as training data
- Evolves better classification prompts through a reflection loop
- Evaluates candidates against your data and selects top performers
- Saves an ensemble of the best prompts for higher accuracy
Plomo does not collect or store your documents for training purposes. Optimization only runs when you explicitly provide labeled data.