Skip to content

Writing documents that answer well

The quality of answers users receive depends directly on the quality of the documents behind them. A well-structured, up-to-date document leads to accurate, specific answers with useful citations. A poorly structured or outdated one leads to vague answers — or no answer at all.

Zahen splits documents into chunks before searching them. Clear headings help those chunks carry meaningful context on their own:

  • Use descriptive headings that name the topic (“Expenses eligible for reimbursement”, not “Section 3”).
  • Put the most important information near the top of each section.
  • Avoid long, unbroken blocks of text — break them into shorter paragraphs or lists where it makes sense.

The assistant reads your documents as written. If a policy is written in dense legal language with heavy cross-references, retrieved passages will be harder to use in a direct, helpful answer. Where possible:

  • State the rule or procedure plainly before adding qualifications.
  • Spell out abbreviations on first use.
  • Avoid pronouns that refer to earlier sections (“the above”, “as stated previously”).

Broad documents covering many unrelated topics can produce lower-quality retrieval — a chunk about one topic gets mixed with another, and the answer loses precision. Where practical, keep documents focused on one policy, process, or topic area.

This doesn’t mean every document needs to be short. A 20-page leave policy covering all leave types is fine. A single document covering leave, expenses, travel, and IT onboarding is harder to retrieve from accurately.

The assistant will cite any document with “ready” status, regardless of how old it is. Outdated content leads to wrong answers.

  • When a policy changes, update or replace the document promptly.
  • Remove superseded versions rather than leaving them alongside the new one.
  • Consider adding the effective date to a document’s title (for example, “Remote Work Policy 2026”) so it’s clear which version is current.

Zahen extracts text during processing. If a document’s content is made up of images — such as a scanned PDF — there is no text to extract and processing will fail. Before uploading:

  • Open the file and try to select and copy a line of text.
  • If you can’t, use an OCR tool to add a text layer before uploading.