KB Extraction Tools

DRUID provides you with two extraction tools that operate as content extraction tools during conversations, particularly valuable for document analysis using Large Language Models (LLMs).

The KB extraction tools during conversations use the content extraction functionality found in the DRUID KB Engine but operate in isolation from the KB processing pipeline, focused on content extraction without introducing changes to the underlying KB infrastructure.

Extract content from documents

Extract content from a document (structured or unstructured) using the internal action KBExtractDocumentContent. This action accepts a document from a [[File]] entity field and provides output in the following format:

• For unstructured documents: content is returned in [[KBOperation]].ExtractedDocument[i].Content.
• For structured documents: content is divided into [[KBOperation]].ExtractedDocument[i].Question and [[KBOperation]].ExtractedDocument[i].Answer.

{
  "SourceFile": "[[Entity]].FileProperty",
  "DocumentType": "structured" //optional, default value = unstructured
}

{
  "SourceFile": "[[Employee]].Resume",
  "DocumentType": "structured" 
}

When extracting data from EML documents, the response JSON includes dedicated fields for email metadata, making the extracted data easier to process and use in business applications:

• To - The recipient email address(es).
• From - The sender email address.
• Subject - The email subject line.
• Date - The email timestamp (ISO 8601 format).

The full email content, including all metadata and the complete message body, is also saved in both the answer and content fields within the paragraphs array.

{
  "to": "recipient@example.com",
  "from": "sender@example.com",
  "subject": "Email subject line",
  "date": "2024-01-15T10:30:00Z",
  "paragraphs": [
    {
      "question": "string",
      "answer": "Complete email body content including To: recipient@example.com, From: sender@example.com, Subject: Email subject line, Date: 2024-01-15T10:30:00Z, and the full message body text",
      "content": "Complete email body content including To: recipient@example.com, From: sender@example.com, Subject: Email subject line, Date: 2024-01-15T10:30:00Z, and the full message body text"
    }
  ],
  "error": "string"
}

This structured response allows you to easily access specific email fields for filtering, sorting, or display purposes, while maintaining the complete email content for comprehensive processing.

Extract content from websites

Crawl a specific website and extract paragraphs (unstructured data) using the internal action KBExtractUrlContent. internal action crawls a specified website and extracts unstructured content as paragraphs. The extracted content is returned as a collection of strings in [[KBOperation]].ExtractedDocument[i].Question and [[KBOperation]].ExtractedDocument[i].Answer.

{
  "Url": "",
  "CrawlHttpRequests": "true|false"
}