---

The idea

Most AEC document tools are passive file cabinets with a search box. On a real project, dozens of people work off the same documents, and two things constantly go wrong:

1. Documents contradict each other and nobody notices until it's expensive.
2. A revision quietly makes other documents stale — and no one gets told.

Generic RAG makes #1 worse: it returns one confident answer from the top-matching chunk, even when another document disagrees. Keystone does two things differently.

1. Conflict-aware RAG

Ask "what concrete strength is required?" and instead of one answer, Keystone retrieves the evidence, clusters the claims, and — if sources disagree — returns the contradiction:

⚠️ Conflict — 03 30 00 (Rev C) says 30 MPa, S-201 (Rev A) says 25 MPa.

In construction, the disagreement is the answer worth knowing.

2. Change-impact graph ("blast radius")

Every document references others (S-201 → "per Spec 03 30 00"). Keystone extracts those references into a dependency graph. Revise one document and it traces the blast radius — every drawing and RFI downstream that just became suspect — and animates it outward from the epicentre.

Both are tractable because AEC documents use rigid identifiers (CSI MasterFormat codes like 03 30 00, sheet numbers like S-201), so reference extraction is deterministic, not guesswork.

---

Stack

Layer	Tech
Frontend	Next.js 14 (App Router), React 18, TypeScript, Tailwind CSS
API	tRPC (end-to-end typesafe), Zod
Data	Prisma ORM, PostgreSQL 16 + pgvector
Auth	NextAuth (credentials + JWT), bcrypt, multi-tenant RBAC
AI	RAG over pgvector embeddings; conflict-aware retrieval; MCP server
DevOps	Docker, GitHub Actions CI, DigitalOcean App Platform

Runs with zero API keys — a deterministic local embedder backs the AI layer out of the box. Set OPENAI_API_KEY (or any OpenAI-compatible base URL) to swap in a real provider.

---

Quick start

npm install
cp .env.example .env          # defaults work for local Docker
npm run db:up                 # Postgres + pgvector in Docker
npm run db:migrate            # create schema
npm run db:seed               # demo project with a planted conflict
npm run dev                   # http://localhost:3000

Sign in with the seeded account: [email protected] / password123, then open Riverside Hospital — Phase 2.

---

Features

• Multi-tenant workspaces — organizations, members, role-based access (Owner / Architect / Engineer / Contractor / Client)
• Documents — drawings, specs, and RFIs; chunked, embedded into pgvector, cross-referenced into a graph
• Impact graph — interactive blast-radius visualization; click any document to trace what it affects
• Conflict detection — flags graph-adjacent documents that assert different values for the same unit
• Ask your project — conflict-aware Q&A with citations
• MCP server — exposes project tools (ask_project, search_documents, list_conflicts, …) to any MCP client

---

MCP

npm run mcp        # stdio server for Claude Desktop / MCP clients
npm run mcp:test   # in-process smoke test

See mcp/README.md for the Claude Desktop config.

---

Deploy

• CI — .github/workflows/ci.yml runs typecheck, lint, and build on every push.
• Docker — docker build -t keystone . produces a standalone production image.
• DigitalOcean — doctl apps create --spec .do/app.yaml provisions the web service + managed Postgres and runs prisma migrate deploy as a pre-deploy job. (Enable the vector extension on the managed database.)

---

Project structure

src/
  app/                 # routes: landing, auth, /app dashboard, /app/projects/[id]
  components/          # providers, auth shell, workspace (impact graph, ask, conflicts)
  server/
    routers/           # tRPC routers: auth, projects, documents, conflicts, ai
    services/          # ingest, conflict detection, conflict-aware ask
    lib/               # access control (RBAC)
  lib/                 # db, ai (embeddings + LLM), chunking, reference extraction, claims
prisma/                # schema + seed
mcp/                   # Model Context Protocol server
design/                # the "Lighttable" interactive prototype