All Startup Ideas

Corporate Knowledge Base AI Assistant

RAG assistant over Notion, Confluence, Slack, and Google Drive. Cited answers for internal teams—priced below Glean for mid-market companies.

Internal knowledge is scattered across Notion workspaces, Confluence spaces, Google Drive folders, and three years of Slack threads. New hires ask the same questions in #general. Tenured employees become human search engines. When someone asks "What's our SOC 2 scope?" the correct answer exists—in a PDF, a Notion page from 2023, and a Slack thread where legal clarified an exception. Nobody has time to stitch it together. Notion AI and Confluence AI help inside their silo. They do not answer across systems with unified permissions. Enterprise buyers with budget buy Glean or similar six-figure deployments. Mid-market teams (80–500 employees) are priced out but still drown in duplicate documentation and stale wikis.

Built for Developers, Solo Founders.

Suggested stack: Next.js 14 + Supabase pgvector, OpenAI embeddings + Cohere rerank, Inngest + OAuth vault, Claude Sonnet. Weekend scope: about 12 hours.

The Problem

Internal knowledge is scattered across Notion workspaces, Confluence spaces, Google Drive folders, and three years of Slack threads. New hires ask the same questions in #general.…

The Solution

A web app that connects to Notion, Confluence, Slack (public channels), and Google Drive via OAuth. Nightly (or hourly) ingestion chunks documents, stores embeddings in pgvector…

Market Research

Notion reports 100M+ users; Confluence remains default in Atlassian-heavy engineering orgs—buyers already store truth in these systems.

Competitive Landscape

Glean — Enterprise work AI + search across SaaS apps. Best-in-class for F500; long sales cycle, opaque pricing. Typically six figures/year; quote-only

Business Model

Growth ($499/mo) — Up to 100 employees, 3 connectors, 5K queries/mo

Recommended Tech Stack

Next.js 14 + Supabase pgvector — chunks(id, source, external_id, text, embedding, acl_json). sync_jobs table per connector. Match user ACL at query time.

AI Prompts to Build This

Copy these build prompts into Claude, Cursor, or your AI coding tool. Create a free account to unlock the full research behind them.

  1. 1. Ingestion + Chunk Pipeline

    Build a Notion → Supabase ingestion worker (Inngest cron).

Steps: list all pages user token can access → fetch blocks → flatten to markdown → chunk 800 tokens with 100 overlap → embed with text-embedding-3-large → store acl_json { notion_page_id, workspace_id }.

Dedupe by content hash. Log sync_errors. Admin UI: last sync time, page count, failed pages.

  2. 2. Permission-Aware Retrieval API

    Implement POST /api/ask { question, user_id }.

1) Load user's allowed source IDs from connector ACL tables.
2) Vector search: WHERE source_id IN (...) ORDER BY embedding <=> query LIMIT 30.
3) Rerank to top 8.
4) Claude: answer with citations; if insufficient evidence, say "I don't know" and log to unanswered_questions.

Never return chunks user cannot access—filter in SQL, not post-hoc.

  3. 3. Slack Bot MVP

    Add Slack Events API bot: @mention in allowed channels triggers same /api/ask pipeline. Reply in thread with answer + up to 3 citation links. Rate limit 10 questions/user/day on Starter plan. Store thread_ts for feedback thumbs up/down.