Guide · Beginners
Preparing your documents for accurate answers
Formats, structure, duplicates, corpus priorities — the best practices that make a real difference to answer quality.
A document assistant is only as good as its documents. Here is what really matters, given how EasyRag works.
Understand what indexing does
EasyRag extracts the text from each document, splits it into chunks of about 400 words (with a slight overlap), adds a sentence of context about the document to each chunk, and then indexes it for both semantic search and full-text search.
As a result, each chunk should make sense more or less on its own. Well-structured documents produce better chunks.
Choose the right formats
| Situation | Recommendation |
|---|---|
| Legacy Word document (.doc) | Save it as .docx: the .doc format is not supported |
| Legacy spreadsheet (.xls) or OpenDocument | Convert it to .xlsx or CSV |
| Scanned PDF | It requires OCR (Mistral OCR key); otherwise, use the original text version |
| Highly visual presentation | Add the key text to the speaker notes or provide a companion text document |
| Training video | It can be transcribed (transcription key required); written material remains more reliable |
Structure the content
- Clear headings. Section headings are detected and help place each chunk in context.
- One idea per paragraph. Avoid blocks of text spanning several pages with no subheadings.
- Definitions close to where they are used. An acronym defined on page 1 and used on page 40 will be poorly understood in the chunk from page 40.
- Simple tables. Deeply nested tables lose their structure during extraction.
Avoid contradictions
Two versions of the same procedure in the same corpus will produce inconsistent answers.
- Remove outdated versions, or
- move them to a separate, lower-priority corpus: in case of contradiction, the source from the higher-priority corpus (priority 1 = highest) wins.
Organize your corpora
One corpus per domain or per level of authority makes it easier to handle:
- prioritization (official texts before explanatory documents);
- reuse of the same corpus by several assistants;
- permissions for viewing and downloading sources, which are set per corpus.
Check before publishing
- Run the corpus test search on your ten most frequent questions.
- Test the assistant on those same questions and check every citation.
- After publishing, monitoring (Pro license) reveals unanswered questions: that is your list of documents to write or complete.