AtlasVault AI

An AI knowledge compiler for companies. AtlasVault ingests your company's documents, compiles them into a living, cited knowledge base, and uses them to generate proposals, tender responses, compliance matrices and missing-document checklists — with a source citation behind every claim.

Upload your company knowledge once. Use it forever to win contracts.

Quick start

npm install
npm run setup     # pushes the Prisma schema to Postgres and seeds the demo workspace
npm run dev       # http://localhost:3000

Demo login: demo@atlasvault.ai / demo1234 — a fully populated demo workspace (TrustCode Systems: 11 documents, compiled wiki, an analyzed government tender at 74% readiness, and a cited 8-section proposal draft awaiting approval).

Other seeded users (same password): adaeze@trustcode.ng (Admin/reviewer), tunde@trustcode.ng (Member).

AI modes

• Local Knowledge Mode (default, no key needed): deterministic and extractive — summaries, classification, requirement extraction, evidence matching and answers are assembled verbatim from your documents. Honest by construction; great for offline demos.
• Provider mode: set OPENAI_API_KEY or ANTHROPIC_API_KEY in .env. AI_PROVIDER=auto prefers OpenAI, then Anthropic. Defaults are OPENAI_MODEL=gpt-5.5 and ANTHROPIC_MODEL=claude-sonnet-5, with ANTHROPIC_OPUS_MODEL=claude-opus-4-7 documented as the Opus alternate. Generation becomes fully synthetic but stays grounded: prompts require inline citations to retrieved excerpts and an explicit "insufficient evidence" signal instead of invention.

The magic-moment demo (3 minutes)

1. Sign in → Dashboard shows knowledge health, deadlines, expiring certificates.
2. Documents → drop in a company profile / proposals / CVs / certificates; each is extracted, summarized and classified.
3. Knowledge Base → Compile knowledge base → Company Overview, Services, Past Projects, Team, Certificates pages, each citing its sources.
4. Ask AI → "Which past projects prove we can deliver a school management system?" → cited, confidence-scored answer.
5. Opportunities → drop in a tender PDF → requirements extracted, compliance matrix built, readiness scored, missing documents listed (expired certificates are flagged as disqualification risks).
6. Click Generate proposal → 8 sections drafted from your own evidence, each with citations, confidence and missing-evidence notes.
7. Request approval → approve as Admin → Export DOCX (export is blocked until approval — that's the trust model).

Architecture

Next.js 15 (App Router, TS) · Tailwind · Prisma + PostgreSQL · JWT sessions · pdf-parse + mammoth extraction · Anthropic SDK task layer with deterministic fallback · docx export. See PLAN.md for the full implementation plan, schema, page map, API surface and V2/V3 roadmap.

Platform layer

Beyond the core workflow: a marketing landing page with transparent pricing at /, a branded PDF engine (cover page, contents, headers/footers, brand color) for proposals and outputs, an Output Studio with five knowledge-grounded templates, Analytics (hours saved, win rate, evidence reuse), a feedback learning loop (thumbs-up answers become searchable workspace memory — never shared across companies), a brand kit, workspace AI/privacy toggles, an integrations directory (admin-gated and revocable), and a platform admin portal at /admin (demo owner account) with sanitized error monitoring and privacy-safe org metadata — no customer content ever.

Security model

Role-based access control (Owner/Admin/Member/Viewer) on every mutating route · workspace isolation on every query (verified: cross-workspace export → 404, cross-workspace Ask AI → "not enough evidence") · confidential documents excluded from Viewer retrieval · human approval required before export · sensitive-data warnings at export · rate limiting on auth/ask/upload/generate · branded error pages with support reference IDs; sanitized error logs (no document content, prompts or outputs) · full audit log of uploads, generations, approvals, exports and permission changes.

Supabase + Vercel setup

This app uses Supabase as infrastructure only: Postgres for Prisma and Storage for uploaded documents. Auth remains the app's custom JWT/bcrypt flow.

Create a Supabase Storage bucket named documents, then set these variables in Vercel for both Production and Preview:

• DATABASE_URL - Supabase Postgres connection string for Prisma.
• DIRECT_URL - Supabase direct Postgres connection string for Prisma schema pushes.
• AUTH_SECRET - a long random session signing secret.
• SUPABASE_URL - your Supabase project URL.
• SUPABASE_SERVICE_ROLE_KEY - server-only key used by API routes to upload files.
• SUPABASE_STORAGE_BUCKET - usually documents.
• AI_PROVIDER - optional, auto, openai, or anthropic.
• OPENAI_API_KEY - optional, enables OpenAI generation.
• OPENAI_MODEL - optional, defaults to gpt-5.5.
• ANTHROPIC_API_KEY - optional, enables Claude generation.
• ANTHROPIC_MODEL - optional, defaults to claude-sonnet-5.
• ANTHROPIC_OPUS_MODEL - optional, alternate Claude Opus setting claude-opus-4-7.

After setting DATABASE_URL, initialize the database:

npm run db:push
npm run db:seed