File Extractors

File extractors are Druid-specific tools that pull raw content out of uploaded files before the Knowledge Base (KB) engine indexes it. For each supported file type, you choose:

• Extractor — how text, structure, images, and media are read from the file.

• Content Chunker — how extracted content is split into articles (see Content Chunkers).

By default, most file types use the Standard extractor and the LLM content chunker. You can change either setting per file type on the File Extractors section or override them at the data source, node, or leaf level.

Once you've selected the preferred file extractor(s), click Save to apply your changes.

Specific file extractors are available per file type, as follows:

CSV

Word Files

Druid supports the following extractors for Word files (.doc, .docx):

Powerpoint

Druid supports the following extractors for PowerPoint files (.pptx, .ppsx):

HTML

Druid supports the Standard extractor for HTML. It extracts text while preserving headings, paragraphs, lists, and links, and removes unnecessary web formatting.

Image extraction is enabled by default if the img tag is not excluded in the HTML Settings. Images are stored in Druid storage and linked with a 30-minute authentication token. The link is embedded in the extracted article paragraph so users can view the image temporarily in chat.

To make text within images searchable in the Knowledge Base, set Use OCR for pictures to true. The extractor performs OCR on images found in the HTML code, stores the image in Druid storage, and links it with a 30-minute authentication token. This link is embedded in the extracted article paragraph, allowing users to also view the image temporarily in chat.

PDF

Druid supports multiple file extractors for PDF files. Each extractor is designed for different types of PDF documents, ensuring optimal content extraction based on document structure and format.

For the Standard (formerly known as Elpis) and Daguerre extractors, image extraction is enabled by default. Images are stored in Druid storage and linked with a 30-minute authentication token. The link is embedded in the extracted article paragraph so users can view the image temporarily in chat.

To make text within images searchable in the Knowledge Base, set Use OCR for pictures to true. The extractor performs OCR on images found in the file, stores the image in Druid storage, and links it with a 30-minute authentication token. This link is embedded in the extracted article paragraph, allowing users to also view the image temporarily in chat.

For the Elpis and Daguerre extractors, you can configure Use OCR for pictures using one of three modes:

• False (Default): It extracts the image, stores it in Druid storage, and embeds a link with a 30-minute authentication token in the paragraph, allowing users to view the original image within the chat.
• True: The extractor performs OCR on all images in the file to convert them into searchable text, stores the images in Druid storage, and links them with a 30-minute authentication token. These links are embedded in the extracted article paragraph, allowing users to also view the original images temporarily in chat.
• Auto: A hybrid intelligence mode. If the extractor finds both images and text on the same page, it skips OCR for the image. Instead, it extracts the image, stores it in Druid storage, and embeds a link with a 30-minute authentication token in the paragraph, allowing users to view the original image within the chat.

Excel Files

Druid provides multiple extractors for Excel files (.xls, .xlsx, .xlsm), each designed for different extraction needs. Choose the appropriate extractor based on your need for table structure, formatting, or bulk data extraction.

JSON

Druid supports the Structured extractor for JSON data processing.

Use structured JSON files to ingest content from third-party systems (such as Salesforce or Confluence) directly into the Knowledge Base. This extraction method is compatible with the following data source types:

• Unstructured
• File Repository
• Custom

The JSON file should follow this format:

[
  {
    "Title": "Sample Title",
    "Content": "Content 1"
  },
  {
    "Title": "Sample Title 2",
    "Content": "Content 2",
    "PageNumber": "3"
  },
  {
    "Title": "Sample Title 3",
    "Content": "Sample Content 3",
    "SheetName": "Sheet1"
  }
]

The following table provides the description of each JSON property:

Video

Druid can ingest video files from SharePoint, Custom Data Sources, file repository, shared drive, and websites where the video is hosted directly (not embedded from YouTube, Vimeo, or similar platforms). The KB Agent discovers the file, converts it to audio, generates a transcript with ASR, then applies the selected Extractor and Content Chunker to build searchable KB content from that transcript.

For more information, on how to extract data from video content, making it searchable and usable within your knowledge base, see Extracting Data from Video Files.

Audio

Druid supports the Standard extractor for audio files. The KB Agent transcribes the file using automatic speech recognition (ASR), then extracts and chunks content from the transcript for indexing.

Choose Content Chunker: Basic for fixed-size chunks, or Llm for context-aware chunking. You can override these settings at the data source, node, or leaf level, as with other file types.