Building a Kitab Corpus for AI Q&A: From Upload to Sourced Answers

Ask the Kitab in e-pondok answers only from the kitab corpus your pondok uploaded. That limit is deliberate: every answer can be traced to a document the pondok owns, not to the model’s taste. The consequence is that answer quality follows corpus quality. The good news: building a good corpus is not technical work.

Start with the books actually taught

The best corpus is not the biggest one; it is the most consulted one. Start with three to five books your pondok actually teaches: the mandatory first-year kitab, the ones teachers get asked about most, or your own teaching notes. A small, relevant corpus answers better than a warehouse nobody opens.

Accepted formats: PDFs with selectable text, .docx, .txt, or .md, up to 25 MB per file. Scanned PDFs (page images only) aren’t supported yet; the system marks them failed with a clear reason. One file per book, so each answer’s reference points at the right title.

What happens after upload

The system reads the kitab’s structure (kitab, bab, fasal), splits the text into small passages that never cross a chapter boundary, and indexes each passage by meaning. Thick kitab are processed in steps; their status (queued, processing, ready, or failed) shows on the Kitab Corpus tab. When someone asks a question, the system retrieves the most relevant passages from your pondok’s corpus only, composes the answer from those passages, and attaches reference numbers.

Three practical consequences:

• Questions can use different words: asking “the correct way to perform wudhu” still finds a passage about “tata cara bersuci”.
• Answers carry [1], [2]: each marker names the kitab, chapter, and (for PDFs) page, and opens the original passage, like a footnote.
• Not uploaded means not answered: when no passage is relevant enough, the system says the corpus doesn’t cover it instead of inventing. Those questions don’t use quota.

Keep the corpus clean

• One clean source per book: choose a PDF with selectable text or a DOCX/TXT version. If the result is messy, delete the document from the corpus and upload a better version; its index is rebuilt.
• Enter correct titles and authors when uploading, because answer references show these names. Clear chapter headings inside the file (e.g. “Bab Wudhu” or “باب الوضوء” on their own line) make references more precise.
• Grow in stages: add a book when students or teachers start asking about material the corpus does not cover yet.

An already-organized digital library speeds this up: when uploading, the file can be linked to a book already in the Library catalogue.

The controls stay with the pondok

Ask the Kitab has per-pondok settings: topics it won’t serve (say, politics), references per answer, extra instructions, and a monthly question limit to keep AI cost in check. All of it lives in the Settings tab of the Asisten menu, and only admins can change it. For ruling questions, answers always carry an adab note: a study aid, not a fatwa; decisions about personal situations go to the ustadz.

With the right corpus and the controls tuned, Ask the Kitab becomes a fast reference you can trust: answers always come from your own shelf, and they can always be traced.

Summary

The corpus is the substance of Ask the Kitab. Start from the books actually taught, upload one file per book, keep sources clean, and grow gradually. The rest, meaning-matching and sourced answer composition, is the system’s job.

Explore the feature on the Academic AI Assistant and Library pages, or start a subscription to build your pondok’s corpus.