Document Classification

Before running the full extraction pipeline, Cell can classify a document to determine whether it's a policy or a quote. This drives which extraction flow to use.

Basic usage

import { classifyDocumentType } from "@claritylabs-inc/cell";

const { documentType, confidence, signals } = await classifyDocumentType(pdfBase64);

console.log(documentType);  // "policy" | "quote"
console.log(confidence);    // 0.95
console.log(signals);       // ["declarations_page", "policy_number_present", ...]

Options

const result = await classifyDocumentType(pdfBase64, {
  log: async (msg) => console.log(msg),
  models: customModelConfig,
});

Option	Type	Description
log	LogFn	Callback for progress logging
models	ModelConfig	Custom model configuration. Uses classification role.

How it works

Classification is Pass 0 in the pipeline. It uses the classification model (Claude Haiku by default) to analyze the PDF and return:

• documentType — "policy" or "quote". Defaults to "policy" on parse failure.
• confidence — numeric confidence score (0-1). Defaults to 0.5 on parse failure.
• signals — array of strings describing what signals the model detected.

Routing extraction

A typical flow uses classification to choose the right extraction function:

import {
  classifyDocumentType,
  extractFromPdf,
  extractQuoteFromPdf,
  applyExtracted,
  applyExtractedQuote,
} from "@claritylabs-inc/cell";

const { documentType } = await classifyDocumentType(pdfBase64);

if (documentType === "quote") {
  const { extracted } = await extractQuoteFromPdf(pdfBase64);
  return applyExtractedQuote(extracted);
} else {
  const { extracted } = await extractFromPdf(pdfBase64);
  return applyExtracted(extracted);
}