Devops platform

See the following image for a high-level overview of Sensible's document data extraction workflow:

As the preceding image shows, Sensible's workflow contains the following steps:

1. Ingest - In this step, you upload document files to Sensible, and Sensible converts them to a standardized text representation in preparation for extracting structured data. Features include:

   • File types - Sensible supports PDFs, Microsoft Word and Excel, and image types such as JPEG. For more information, see Supported file types.
   • File upload methods - Use Sensible's SDK, API, bulk upload UI ("manual upload"), or Zapier. For more information see Integrate.
   • Optimized OCR - Sensible represents the whole document in a standardized text format. Sensible optimizes performance by choosing between OCR, including Amazon, Google, and Microsoft OCR, or direct extraction of embedded fonts. For more information, see OCR.

2. Classify - Upload the document to a document type, or API endpoint you configure. Each document type extracts data from a general category of similar documents, for example, /extract/bank_statments or extract/drivers_license. When you upload a document to a document type, Sensible automatically classifies the document into subtypes, for example, chase_statements and wells_fargo statements for the bank_statements document type. You define each subtype using a config. A config specifies how to extract data using queries, and how to populate a target output schema. Features include:

   • Multi-document files - Sensible splits documents out of a multi-document file (a "PDF portfolio"). For example, Sensible splits a mortgage-application PDF file into tax documents, bank statements, and application forms. For more information, see Multi-document extractions.
   • Classification granularity - By default, Sensible automatically classifies by document subtype during the extraction workflow. Prior to extraction, you can classify a document at the document type level using the classify endpoints.
   • Configurable classification - Configure Sensible's default subtype classification using fingerprints to improve performance and accuracy. For example, you can use fingerprints to fall back between generalized configs for high-volume documents and layout-specific configs for long-tail documents in a document type. For more information, see Fingerprints and Using fallbacks.

3. Extract - In this step, Sensible returns the extracted document data. Features include:

   • Highly configurable layout- and LLM-based extraction - Write SenseML queries to extract data from your custom documents. For more information, see SenseML reference introduction.

   • Out-of-the-box templates - Sensible provides an open-source library with out-of-the-box support for common business forms, so you can get started extracting from balance sheets, drivers licenses, rent rolls, tax documents, resumes, and more in just minutes. For more information, see Out-of-the-box extractions.

   • Structured data - Sensible returns extracted data as key-value pairs, including complex data like tables or sections.

   • Traceability for source text - Use visual overlays in the Sensible app and metadata in the returned JSON to trace extracted data to its source location (or "bounding box coordinates") in the document. For more information, see Color coding and Verbosity.

   • Validations - Quality control the data extractions in a document type by writing validations. For example, configure Sensible to throw an error if an extracted zip code is the wrong number of digits. For more information, see Validating extractions.

   • Confidence scores - For LLM-based extractions, confidence signals offer more nuanced troubleshooting than numerical confidence scores, with error messages like multiple_possible_answers and answer_may_be_incomplete. For more information, see Qualifying LLM accuracy.

4. Monitor - Quality control your extractions in production. Features include:

   • Human review UI - If extractions contain errors, for example as the result of hard-to-read handwriting, you can configure rules to flag extractions for manual review. A reviewer can correct and approve flagged extractions in the Sensible app's tab. For more information, see Human review.
   • Metrics dashboard: View and filter your real-time extraction metrics using the Sensible app's dashboard. For example, view the number of extractions in recent days or filter by which extractions are returning nulls for target data. For more information, see Monitoring extraction metrics.