AI document assistant
Upload a PDF, Word file or slide deck and ask it anything. Every answer comes back with the page or slide it came from, so nobody has to take the AI’s word for it.
Long documents → RAG assistant → Cited answers
Live snapshot- Client
- In-house R&D
- India
- Industry
- Legal, HR & operations teams
- Year
- 2024
- Timeline
- 4 weeks
- Services
Results
- Document types: PDF, Word and PowerPoint
- 3
- Of answers cite a page or slide
- 100%
- Open model that can be hosted privately
- 7B
[01]The challenge
Policies, contracts and decks run to dozens of pages, and the answer someone needs is usually one sentence in the middle. Search finds keywords, not answers.
General-purpose AI chat makes it worse: it answers confidently without saying where the answer came from a non-starter for HR, legal and compliance work.
[02]Our approach
- 01
Each file type gets its own parser, so page numbers in PDFs, tables in Word files and speaker notes in slides survive the trip into the index.
- 02
Text is split into overlapping chunks, embedded with an open sentence-transformer model and stored in a FAISS index, so retrieval runs locally and fast.
- 03
The model may only answer from the passages retrieved, and must cite each one it uses. When nothing relevant comes back, it says so instead of guessing.
[03]The solution
DocuMind is a Python service with a chat interface: upload a PDF, DOCX or PPTX, ask a question in plain language, and get a short answer with numbered citations to the exact page or slide.
It runs on open models, all-mpnet-base-v2 for embeddings and Mistral 7B for answers so it can be hosted privately for documents that shouldn’t leave the company.
[04]What we built
- 01PDF, DOCX and PPTX parsing that keeps page and slide numbers
- 02Overlapping chunks for answers that span paragraphs
- 03Local FAISS vector search with mpnet embeddings
- 04Answers grounded only in retrieved passages
- 05Numbered citations on every answer
- 06Open models that can run on private infrastructure
[05]Live snapshot
An interactive recreation of the product’s core flow — try it.
documind · ask
what is the notice period after probation
- HR_POLICY
- employee-handbook.pdf · p.12, p.14
- answer + 2 citations
Confidence96%
1.2swas Ctrl+F
Interactive · sample documents
[06]Stack
- Python
- RAG
- LLMs
Your project next?
Tell us about the problem you’re facing. We’ll show you how we’d approach it — and what success could look like.
Booking new projects for Q4 2026
We reply to every enquiry within one business day.