Dorio
Back to Case StudiesCase Study

From thousands of buried documents to answers employees can trust

We built an AI-powered knowledge platform where teams ask questions in plain language and get answers grounded in their company's own documents — cited, confidence-scored, and scoped to exactly what each person is allowed to see.

Book a Discovery Call

Inside the platform

The AI knowledge platform: assistant answering a question with citations, inside the full admin workspace

The AI Assistant answering from company documents — with sources, page references, and a confidence score

Cited

Answers link back to the source passages they came from

0–100

Confidence score attached to every AI-generated answer

PDF & DOCX

Parsed, chunked, and indexed automatically in the background

Per-role

Permissions decide which documents each person can ask about

System Context

Frontend

Next.js + React (TypeScript)

Backend

FastAPI (Python)

AI

Google Gemini + RAG over a vector index

Type

Multi-tenant SaaS knowledge platform

The knowledge existed. Finding it was the problem.

A growing organization had accumulated thousands of internal documents — policies, SOPs, contracts, manuals, onboarding guides — spread across folder trees that had stopped making sense years ago. The information employees needed was almost always written down somewhere. Nobody could find it.

So people stopped looking. They asked colleagues instead, and the same handful of experienced staff spent time every week answering questions that were already documented. Onboarding took longer than it needed to. And sensitive material sat unshared, because there was no safe way to control who could see it.

This wasn't a knowledge problem — it was a retrieval problem. The organization didn't need more documentation. It needed a way for every employee to ask a question and get the answer their documents already contained.

File-name search couldn't see inside PDFs. Folder conventions decayed with every reorg. And a generic chatbot was off the table: answers about policy and contracts have to be verifiably correct, not plausible.

Business Challenges

What the organization was up against

Six compounding problems, one root cause: knowledge that couldn't be reached by the people who needed it.

Years of documents scattered across folders

Policies, SOPs, contracts, and manuals lived in nested folder trees that only a few long-tenured people could navigate.

The same questions asked again and again

Employees interrupted managers and senior staff for answers that already existed in writing — nobody knew where.

Slow onboarding for new staff

New hires spent their first weeks hunting for process documents instead of doing the job they were hired for.

Knowledge trapped inside PDFs

The answer was on page 47 of a handbook. Search stopped at the file name.

No permission-based knowledge access

Sharing HR or finance material meant sharing it with everyone — so it stayed locked away instead.

No way to trust an answer

Even when something turned up, there was no way to tell whether it was the current version — or the right document at all.

Our Solution

A knowledge platform, not a chatbot

We designed and built an AI-powered knowledge platform where employees ask questions the way they'd ask a colleague — and get answers assembled from the company's own documents. Answers cite the documents they came from, down to the page and quoted passage, and carry a confidence score so people know how much weight to put on them.

Just as importantly, the platform respects the organization's boundaries. Access control is applied before retrieval, so the AI can only read what the asker is allowed to read. Spaces keep HR answering from HR documents. And when the documents don't contain an answer, the assistant says exactly that.

The result is a single, trusted place where company knowledge is uploaded once, organized automatically, and drawn on every day — instead of a folder graveyard nobody opens.

Core Capabilities

Everything an internal knowledge platform needs

Not a demo with a chat window — a complete product covering search, organization, governance, and administration.

AI Document Search

Search that reads inside your documents, not just file names — from a chat page or the global command bar.

Natural Language Questions

Employees ask in plain language — no query syntax, no tags, no training.

Citation-Backed Answers

Answers name their source documents, down to the page and section, with the supporting passage quoted.

Automatic Document Organization

Bulk uploads come back as a proposed folder structure, ready for an admin to approve — with optional instructions to steer it.

Role-Based Access

Granular permissions grouped by area, with default Admin and User roles plus unlimited custom roles.

Spaces for Department Isolation

HR, finance, or legal content lives in its own space, visible only to the roles and members you assign.

PDF & DOCX Processing

Drag-and-drop uploads, including very large files — split into chunks and processed in the background.

Conversation History

Past conversations are saved and can be reopened anytime — an answer found once stays found.

Suggested Questions

Admin-curated question starters plus AI-generated follow-ups guide people to what's worth asking.

Feedback Loop

Users rate answers; flagged ones land in an admin review queue, and corrections are reused for repeat questions.

Admin Dashboard

Documents, folders, users, and recent activity at a glance — with per-space analytics for deeper insight.

AI Token Usage

Consumption is metered against the company's plan, visible throughout the dashboard, with self-service top-ups.

How It Works

From document to trusted answer

One pipeline handles everything between an uploaded file and a cited answer — no manual filing, tagging, or curation anywhere in it.

Upload documents

PDF and DOCX, even very large files

AI processes content

Parsed and split in the background

Knowledge index

Embeddings stored in a vector index

Employee asks

Plain-language question

Content retrieved

Only from permitted documents

Grounded answer

Generated from sources only

Sources & confidence

Linked sources and a 0–100 score

Product Tour

The platform, screen by screen

Six views of the system in use — the assistant employees live in, and the controls administrators run it with.

AI Assistant

Ask in plain language. Get an answer you can verify.

The assistant answers only from documents the asker can access. Answers carry their citations — filename, page, and the quoted passage — alongside a confidence score and suggested follow-up questions.

AI Assistant chat showing a question answered with source citations, page references, and a confidence score
Dashboard

The state of your knowledge base at a glance

Administrators land on a live overview: total documents, folders, and users; the latest uploads with their folders; and the newest accounts with their roles — no digging through menus.

Admin dashboard with document, folder, and user counts plus recent documents and recent users tables
Document Library

Every document organized, nothing lost

An expandable folder tree keeps the library navigable as it grows. Unsorted files wait in a built-in Unfiled area, and moving a document is a single dropdown — the index updates itself.

Document library with an expandable folder tree, document counts, and per-document move controls
AI Folder Organization

One click turns a pile of uploads into a filing system

After a bulk upload, the AI reviews every file and proposes a professional folder structure — guided by instructions like "separate contracts by year." Admins review the plan, adjust it, and apply it. Hours of manual sorting collapse into minutes.

AI folder organization screen showing a suggested folder structure ready to review and apply
Spaces

Department knowledge stays with the department

An HR space is visible only to HR. Each space carries its own members, role grants, and analytics — questions asked, answer rate, average confidence — so teams can see how their knowledge is actually used.

Space detail page with overview statistics, documents, members, and permission tabs
Users & Roles

Fine-grained control over who sees what

Admins manage every account from one table — create users, assign roles, block or remove access. Custom roles bundle exact permissions, grouped by area, and changes apply immediately to everyone holding the role.

User management table with role assignment plus a role permissions panel
Security & Trust

Designed so nobody sees an answer they shouldn't

An internal knowledge platform is only usable if it's safe to put real documents in. Access control isn't a feature here — it's the architecture.

Company-level isolation

Every account, document, and query is scoped to its own company, enforced on every request — AI retrieval included.

Role-based permissions

Fine-grained permission codes, grouped by area and enforced by the backend on every request — not hidden buttons in the UI.

Department Spaces

Spaces are private by default. A space with no assigned roles or members is visible to nobody but its owner and admins.

Source-only answers

Retrieval is filtered to the documents the asker is allowed to see before the AI ever reads a word. Access control happens first.

Sensitive values masked

Personal identifiers, banking details, salary figures, and medical information are automatically masked in generated answers.

No guessed answers

Quoted citations are verified against the actual document text, and when the documents don't contain an answer, the assistant says so instead of improvising.

Engineering

Nine systems behind every answer

Wiring a text box to an AI model takes an afternoon. Making the answers trustworthy, safe, and billable takes real systems — nine of them, working together:

  • Authentication — JWT-based sessions with per-request account, company, and plan checks
  • Permissions — a three-axis access model: platform admins, company admins, and fine-grained per-role permission codes
  • Document processing — chunked uploads reassembled and parsed by background workers, so a 500-page manual doesn't block the UI
  • Embedding pipeline — content split with overlap, embedded in batches, and retried automatically on rate limits
  • Retrieval — permission-filtered vector search: the allowlist is computed before the query ever touches the index
  • Citation verification — quoted passages are checked against the actual chunk text; a quote that can't be verified is dropped, not shown
  • Feedback loop — flagged answers enter a review queue; corrections outrank the model when the question comes back
  • Usage metering — token budgets checked before every AI call and real usage deducted after it, tied to Stripe billing
  • Multi-tenancy — company-scoped data access on every query, with billing state changed only by verified payment webhooks
Next.jsFastAPIGoogle GeminiVector DatabaseRAG ArchitectureJWT AuthenticationStripe BillingMulti-Tenant SaaSRole-Based Access ControlBackground Processing
Outcome

Less time searching. More time acting.

The platform changes the default behavior around company knowledge: instead of hunting through folders or interrupting a colleague, people ask — and get an answer they can check.

  • Employees ask questions in natural language and get sourced answers in seconds
  • New staff learn the ropes from the knowledge base instead of from senior people's calendars
  • Administrators keep complete control over access — by company, space, role, and individual
  • Knowledge becomes searchable instead of buried, and stays organized as it grows
Technology

The stack behind the platform

Proven, boring-in-the-best-way infrastructure around the AI core — chosen for reliability and maintainability, not novelty.

Frontend

Next.js App Router, React, TypeScript, Tailwind CSS

Backend

FastAPI (Python), SQLAlchemy

Database

MySQL + Chroma vector store

AI

Google Gemini — embeddings & generation, LangGraph pipeline

Authentication

JWT with role-based access control

Payments

Stripe — subscriptions & token top-ups

Document Pipeline

Chunked uploads, background ingestion workers

Infrastructure

Docker, webhook-driven provisioning

If your team's answers are locked in documents nobody can find, this is the system that gets them out — end to end, from document ingestion to access control to citations people can check. We'll walk you through how it would work on your documents.

Book a Discovery Call

Interested in building something similar?

Let's discuss your workflows — and where an AI system like this would remove the most friction.

Not ready for a call? Tell us what's broken →

  • Fixed price before we start
  • You speak directly with the builder
  • If it's not worth doing, we'll tell you

No pitch. Just a clear answer.