Skip to content

Business knowledge preparation guide · By Yes AI

Which file types can AI knowledge search read?

Published 5 min read

AI knowledge search can only answer from text it can extract. Most office formats read cleanly, scanned files need an extra step, and some formats are skipped entirely. This guide sets out what to expect from each type before you decide what goes into the first index.

Check which of your files AI search can readCall (03) 9003 0111Book a free knowledge search consultation

Quick answer

Text-based Word, PowerPoint, PDF, Excel, CSV, HTML, Markdown, email and plain text files are read by mainstream indexing tools. Scanned PDFs and photos need optical character recognition, a separate step that is often billed. Drawings, audio and video are usually left out unless converted. Tables, size limits and files the indexer cannot open cause many of the remaining gaps, so test a sample of each file type.

Administration officer sorting printed manuals, scanned forms and a laptop of folders on a desk in a Newcastle office
Count your files by type before deciding what the first index will read.

Which formats read cleanly?

Microsoft's Azure AI Search blob indexer, one building block used for custom knowledge search, lists the formats it extracts text from, including Word (DOCX, DOC, DOCM), Excel (XLSX, XLS, XLSM), PowerPoint (PPTX, PPT, PPTM), Outlook MSG, EML, PDF, HTML, Markdown, CSV, JSON, XML, RTF, EPUB, OpenDocument files, plain text, and ZIP and GZ archives. Other tools publish their own lists. ChatGPT's SharePoint app, for example, searches .txt, .pdf, .docx, .pptx, .xlsx and .csv files. A format on a supported list tells you text can be extracted. Whether tables, headings and footnotes keep their meaning still needs testing.

What happens to scanned PDFs and photos?

A scanned page is an image, so a tool must run optical character recognition (OCR) before there is any text to search. In Azure AI Search this is a separate enrichment step: image extraction has to be switched on and incurs an additional charge, and OCR beyond a small daily free allowance runs through a billable AI resource. Azure supports OCR on JPEG, PNG, BMP and TIF images and extracts up to 1,000 images from a single document. In Google Drive's search index, OCR in PDF files is limited to 80 pages. Check recognised text against the page before relying on it.

How are spreadsheets and tables handled?

Spreadsheets are extracted as text, which keeps the values and can lose the relationships between them. A price or tolerance means little once it is separated from its row and column labels. Azure's blob indexer can treat each CSV row as its own search document, which suits lists such as part catalogues. For tables inside PDFs and Word files, layout-aware extraction keeps more of the structure, and Microsoft's Document Intelligence layout model, which extracts text, tables, section headings, page headers and footers, is one example. Test questions that depend on a table cell before trusting answers drawn from it.

What about email, archives and attachments?

Outlook MSG and EML files can be indexed, but Azure indexes a message with its attachments, or a ZIP archive with its contents, as a single document. A long email thread with three attached procedures becomes one large item, and an answer may cite the email instead of the procedure. Where attachments hold the real content, save the current versions into the document library and index those. Mailboxes also hold a lot of personal information, so leave them out of a first project unless there is a clear reason to include them.

Which files are usually skipped?

Formats outside a tool's supported list are skipped or reported as errors. Microsoft says the blob indexer stops by default when it meets an unsupported content type, such as an audio file, unless it is configured to continue, and its supported list does not include CAD drawings, audio or video. Oversized files are treated as errors by default. Check password-protected files as well, because an indexer can only read what it can open. Read the indexing error report after the first run, since it lists the files the system could not read.

Do file size limits matter?

Yes, and they differ by tool. Google lists Drive search index limits of 10 MB for most files, 20 MB for XLSX, 30 MB for PDF and 100 MB for text files, and says only 1 MB of text and formatting data per file is extracted for search. ChatGPT's SharePoint app has a limit of 100 MiB per file for search and retrieval. Azure AI Search sets maximum file sizes and extracted character counts by pricing tier. A long manual may pass a size limit and still exceed a text limit, so split long manuals by chapter.

What to expect from common file types

File typeUsually read asCommon problemPreparation step
Word, PowerPoint and text PDFsText with headingsOld versions and duplicatesKeep one current version per topic
Scanned PDFs and photosNothing until OCR runsMisread numbers and names, page limitsRun OCR, check against the page, add a text layer
Excel and CSVCell values as textValues separated from row and column labelsTest table questions and parse lists row by row
Email (MSG, EML)One item including attachmentsAnswers cite the email instead of the procedureSave current attachments to the library
ZIP archivesOne combined itemMixed content in one search resultUnzip and index the files you need
CAD drawings, audio and videoSkipped or reported as errorsContent invisible to searchExport drawing notes or transcripts if needed
Google Docs, Sheets and SlidesText through a Drive connectorFiles over size or text limitsSplit long documents

Practical checklist

  • A count of files by type exists for the document set in scope.
  • A sample of each file type has been indexed and its extracted text checked against the original.
  • Scanned documents have been through OCR and spot-checked for numbers, names and dates.
  • At least three test questions depend on tables, and their answers have been checked.
  • Files over the tool's size or text limits have been split or excluded.
  • The indexing error report from the first run has been read, and each skipped file type has a decision.

Common questions

Can AI search read handwriting?

Sometimes. OCR tools can read some handwriting, and results get worse with faint scans and unusual writing. Treat handwritten notes on forms as unreliable until a person has checked the extracted text against a sample of pages. Where a handwritten note changes the meaning of a procedure, type it into the controlled document.

Should we convert everything to PDF first?

No. Converting Word files to PDF adds a step, and depending on how the PDF is made it can flatten the heading structure that helps split documents into sensible passages. Keep native Word, PowerPoint and Google files where you have them, and convert only the formats your chosen tool cannot read.

How do we find out what the system could not read?

Ask for the indexing report after the first run. It should list every file that failed, was skipped or was only partly read, with a reason. Azure AI Search records each indexer run's status with counts of items processed and items failed, and keeps up to 50 recent runs in its execution history. Review it after each major document load.

Do images and diagrams inside documents get read?

Only if the system is set up for it. Text inside embedded images needs OCR, and a diagram's meaning needs either a written description or an image analysis step that produces captions or tags. A wiring diagram or flow chart may not survive as useful text, so add a short written description beside the diagrams that matter.

Are PowerPoint decks worth indexing?

Some are. Training decks and product presentations can hold useful answers, but slides are often short phrases whose meaning sat in what the presenter said. Index decks that stand on their own, check whether your tool reads speaker notes, and leave out sales decks whose claims are not approved procedure. Test a question or two against each deck you include.

Sources and further reading

Continue planning

Plan a useful knowledge-search pilot

Describe the documents your staff use, the questions they need answered and how uncertainty should be reviewed. We can discuss preparation and a bounded pilot. Keep source documents, customer records and confidential information out of this enquiry.

Check which of your files AI search can readCall (03) 9003 0111Book a free knowledge search consultation