AIWeb, Slack, Teams, Browser extension24 weeks SprintConceptual Case Study - Proposed Solution

Lorebase

Permission-Aware Search and Cited Answers Across Documents, Wikis and Tickets

A concept study: we worked up this product, its architecture and its screens as an example of how we would approach it. The screens use fictional demo data and are design concepts, not screenshots of a live product, and this is not a client project.

Lorebase - Engineering Case Study Cover
Concept
A design study, not a live product
Fictional
Demo data on every screen
24 weeks
Estimated build time

About the Project

Lorebase is a design concept for search and question answering across a company's documents, wiki pages and support tickets. A person types a question in plain words and gets a short answer, with each sentence linked to the exact passage that supports it. The target is a company of several hundred staff or more, where the answer sits in Google Drive, Confluence, SharePoint or a ticket system and nobody can find it.

The main design idea is that permissions come before the language model. Lorebase would copy each source system's access rules into the index next to the text. Every search carries the asker's identity, and only passages that person may already read can reach the model. For the final cited documents, it checks access again with the source system before showing the answer.

Readers can also judge how far to trust an answer. Each source shows its owner, last edit and last review, and old pages rank lower and carry a stale label. When two sources disagree, the answer says so and lists both with their dates. Admins get a source health view and an access test tool, so a security reviewer can see what a named person can and cannot see.

Why This Matters Now

Market, rule and technology context for this kind of product, checked on 16 November 2026. Each item names its source.

  • Microsoft's Copilot blueprint, updated on 13 Nov 2026, says Copilot grounds answers in data users can already access, and makes remediating oversharing the first of three pillars.Microsoft Learn, 13 Nov 2026
  • Inc42 reports a CERT-In advisory in which AI agents reached live Indian systems using exposed credentials and over-permissive cloud roles, and none of the affected organisations detected it themselves.Inc42 (news), 16 Oct 2026
  • PromptArmor found that a PDF with hidden white-on-white text could make Atlassian's Rovo agent send Jira and Confluence data to an attacker's server, as reported by The Decoder (trade press).The Decoder, 10 Aug 2026
  • The Model Context Protocol 2026-07-28 release adds Enterprise Managed Authorization, so a company identity provider controls access to MCP servers, and makes issuer checks mandatory.Agentic AI Foundation, 28 Jul 2026
  • Law firm AZB reports that under India's DPDP Rules, notice, consent and data fiduciary duties begin on 13 May 2027, after a first phase that set up the Data Protection Board (secondary source).AZB & Partners, 19 Nov 2025

Requirements & Objectives

  • Read documents, wiki pages and tickets from Google Drive, Confluence, SharePoint and a ticket system, together with each source's users, groups and sharing rules.
  • Attach the asking person's identity to every search, so only passages that person may read can reach the results or the answer.
  • Write answers that cite one passage per sentence, and open the source at that passage.
  • Show last edit, last review and owner for every source, mark old sources as stale and rank them lower.
  • Say so when two sources disagree, and list both with their dates.
  • Let admins see connector lag, documents deleted at source but still indexed, and spaces with no owner.
  • Let admins test what a named person can see, give the reason chain, and save the test to run again.
  • Answer in Slack and Microsoft Teams with the same checks, and keep a query log of who asked, which passages were used and which filters applied.

The Engineered Solution

1Access rules stored with every passage

Each passage carries the list of people and groups allowed to read it. A database policy filters by the asker's identity before ranking, so application code cannot skip the check, and a live source check covers the final cited documents.

2Answers a reader can verify

Citations point to a passage, not a document. A check step confirms each cited passage supports its sentence, and an unsupported sentence is removed or the answer says it cannot confirm.

3Freshness and conflicts in plain view

Last edit, last review and owner appear beside each source. Old pages rank lower, conflicting pages are listed side by side, and owners get a stale report for their space.

4Evidence for the security reviewer

An access test tool shows why a named person cannot see a document. The audit log keeps the user, the passages used and the filters applied, and a saved test can run after every sync.

Design Concept Screens

Design concepts with fictional demo data. They show how the product could look and work, not a live release.

Lorebase search and answer screen with a question, an answer whose sentences carry numbered citation chips, source cards with Fresh and Review due labels, and a panel showing which sources the person can and cannot read

01 / 05 Search and answer - ask a question, check each cited passage and see how fresh every source is

Engineering Process

  • Map the company's sources, spaces and permission models with IT, knowledge owners and a security reviewer.
  • Define the access data model: principals, groups, sharing links and how each source expresses them.
  • Build the Drive and Confluence connectors with access sync, chunking and hybrid search, and write an evaluation set of 150 questions with expected sources.
  • Add the answer step, passage citations, the citation check and freshness scoring, and test them against the evaluation set.
  • Add the SharePoint and ticket connectors, the Slack and Teams bots, the source health view and the access test tool.
  • Test for leaks with seeded restricted documents, prompt-injection files and deletion checks, and fix every failure before launch.
  • Pilot with one department, review the unanswered question report with its owners, then widen to more teams.

Production Tech Stack

Layer / ComponentProduction Technologies
Web app and extension
Next.js (App Router)TypeScriptTailwind CSSshadcn/uia Manifest V3 browser extension
Chat apps
Slack and Microsoft Teams apps built on their official SDKs
Backend
Node.js with TypeScript and Zod validationa Python FastAPI service for parsing and chunkinga durable job queue for connector workers
Database
PostgreSQL with pgvectorfull-text search and row-level security
Identity and sources
OIDC or SAML sign-inSCIMGoogle Workspace APIsAtlassian Cloud RESTMicrosoft Graph
AI and hosting
Hybrid retrieval with a rerankerAnthropic Claude or OpenAI for answersVercel for the web appAWS ECS or Google Cloud Run for workersGitHub Actions

What the design aims for

Lorebase aims to answer staff questions using only the passages each person may already read, with a source, an owner and a review date beside every sentence.

ARCHITECTURE & SCOPE CONSULTATION

Need a System Engineered Like Lorebase?

Deploy a dedicated engineering pod or initiate a fixed-scope milestone sprint with full IP ownership and guaranteed delivery.