Chat With Your PDFs, Including Nepali Documents

Most of the documents that matter in daily life are PDFs nobody wants to read: a forty-page tender notice, a lease, a syllabus, five years of past papers, a contract in Nepali that you need to understand before Friday.

Document RAG — retrieval-augmented generation — is the feature that lets you ask questions across those instead of reading them. Here is how to use it properly, and where it lets you down.

What it actually does

When you upload a document, it is split into passages and indexed. When you ask a question, the system finds the passages most likely to contain the answer and gives those to the model along with your question.

The practical consequence: the model answers from your document, not from its general training. That is the difference between “what does Nepali law generally say about X” and “what does this contract say about X”.

Nepali and Devanagari documents

This works. Current models handle Devanagari text well enough to answer questions about it, summarise it, and translate passages on request. You can upload a Nepali notice and ask questions in English, or the reverse.

The one real constraint is not language, it is format:

• A PDF with a text layer (exported from Word, or a proper digital publication) works directly.
• A PDF that is a scan or photo of a page contains no text at all — just an image. It needs OCR before anything can read it.

If an upload gives strangely empty answers, this is almost always why. Try selecting text in the PDF: if you cannot, neither can the AI.

Getting good answers

Ask specific questions. “Summarise this” produces a generic summary. “What is the notice period for termination, and which clause says so” produces something you can act on.

Ask for the source. Requesting the clause or page number turns an answer you have to trust into one you can verify in ten seconds. This is the single highest-value habit when working with documents.

Upload the relevant set, not everything. Retrieval quality drops when the index is full of unrelated material. Ten pertinent documents beat two hundred where most are noise.

Follow up rather than restart. With memory on, the conversation keeps context, so you can drill into an answer instead of re-establishing the setup each time. More on that in memory and RAG explained.

Where it fails

Being honest about this matters more than the feature list:

• Retrieval can miss. If the answer lives in a passage the search did not surface, the model answers from what it did get — and may sound confident doing it.
• Tables and complex layouts degrade. Multi-column layouts and dense tables often lose structure when extracted. Numbers pulled from a mangled table are worth double-checking.
• It does not truly read everything. For a very long document, it reads the passages retrieved, not all forty pages. Broad questions across a huge corpus are where it is weakest.

For anything consequential — legal, financial, medical — treat the output as a fast index into the document, not as a substitute for reading the part that matters.

What people use it for here

• Lok Sewa and exam prep: upload past papers and syllabus, generate practice questions, check written answers. See AI for Lok Sewa preparation.
• Legal and official notices: understand a Devanagari document quickly, then verify the clause that decides the matter.
• Coursework: ask questions across a semester of lecture PDFs.
• Business paperwork: pull dates and totals out of invoices and contracts.

Cost

No subscription on Nuvaa. Free tier to start at Rs 0, then credit top-ups from Rs 100, spent only as you use it. Document chat is not gated behind a monthly plan — it is available on any account, paid via eSewa, Khalti or Fonepay.

---

Got a PDF you have been avoiding? Upload it free and ask it something. Or see how to build an agent that does this on a schedule.