iterthink

Your €40k AI RAG Infrastructure Can Be a Shared Folder: The Pragmatic Guide to AEC Knowledge Management

Local-First RAG Without the Enterprise Bloat

Published 2026-06-18

Simon Dilhas (Cofounder abstract & BIM Pirate)

Search for "how to build an AI knowledge base for your company," and you'll find a massive tech stack: cloud pipelines, complex vector databases, and ongoing bills of €2,000 to €10,000 per month. But for a typical architecture or engineering firm, that is pure over-engineering. Your existing SharePoint or shared network drive is already 90% of the infrastructure you need.

BIM was always supposed to be a knowledge management solution—not just a geometry problem. It promised to get the right information to the right person at the right moment. Yet today, critical institutional knowledge (SIA norms, fire safety standards, custom tender clauses) remains locked away in unsearchable project folders.

To unlock this data, companies are turning to RAG (Retrieval-Augmented Generation). But instead of buying into bloated cloud architectures, iterthink was built local-first to turn your existing file structure into an advanced AI knowledge base.

What is RAG, and Why Does It Matter for Specifications?

RAG (Retrieval-Augmented Generation) fixes the biggest flaw of standard AI: generic answers. When you ask a vanilla LLM a technical question, it pulls from its public training data. It knows nothing about your specific project constraints.

RAG automatically scans your company's internal corpus (norms, past specifications, contract templates), grabs the relevant snippets, and feeds them to the AI as context. The result? The AI answers using your institutional data and your referenced standards (like SIA 118), rather than making things up.

The "Enterprise" Bloat vs. The iterthink Way

The standard corporate approach to RAG requires setting up five to eight different web services, managing vector databases (like Pinecone or pgvector), and spending months on engineering.

iterthink cuts through this complexity by using a local-first architecture. To turn your office files into a queryable AI knowledge base, you only need to change a single configuration value:

documents_root: /Volumes/team-share/projects

By pointing iterthink to your mounted SharePoint folder, Dropbox, or local network drive, the built-in RAG system automatically indexes the files.

How the Architecture Works Without a Server

Access control: Managed entirely by your existing SharePoint or server permissions. No new IT setups required.
The intelligence (local): The RAG vector index and the SQLite database live locally on each user's machine (store.rag.sqlite3).
The shared cache: To prevent multiple users from paying the processing cost for the same document twice, enrichment results are cached as lightweight JSON files right in the shared project folder (.iterthink-cache/).

Zero Cloud Costs: Leveraging Your Rendering Hardware

Most AI guides assume you are sending every document chunk to external APIs like OpenAI or Claude, racking up heavy token bills. For processing architectural specifications, you don't need trillion-parameter general intelligence. Small, open-source local models handle structured document analysis perfectly.

Because iterthink supports local LLMs natively via Ollama, you can run models like Llama 3.1 8B or Mistral 7B directly on-device.

The AEC advantage: Most architecture offices already have powerful workstations equipped with high-end GPUs for 3D rendering and visualization. This hardware sits idle for hours—iterthink utilizes that existing VRAM to run your local AI completely for free. Your documents never leave your physical building.

The Natural AEC Roadmap: Form Follows Function

This local-first setup isn't a dead-end shortcut. It is Phase 1 of a scalable architecture where nothing you build is thrown away.

Phase	Team Scale	Infrastructure Required	Storage Backend
Phase 1: Local-First	1 – 15 users	Mounted SharePoint + Local Ollama	Local SQLite + Shared Folder Cache
Phase 2: Centralized Server	15 – 50 users	Small office server running an LLM	Shared PostgreSQL + pgvector
Phase 3: Ecosystem Integration	50+ users	Embedded directly into yourcompanyos	Full Enterprise SSO & IT Audit Logs

You only move to the next phase when real user pain drives the complexity. If your team complains that they can't cross-search un-synced projects, you move to Phase 2. If IT needs strict governance over data residency logs, you scale to Phase 3.

The Lesson for Modern BIM Practices

We have spent two decades arguing about IFC open standards and model ownership, yet our most valuable data—the lessons learned, the legal clauses, the design choices—remains buried in unsearchable text files.

RAG makes retrieving that data trivial. But don't let vendor marketing convince you that simplicity is irresponsible. A local vector index connected to a shared drive will be used by your team today. A multi-service cloud pipeline with a message queue will simply get blamed when it goes down an hour before a tender deadline.

Start simple. Let actual friction dictate your tech stack.

Explore the open-source document editor on GitHub or download iterthink to turn your shared drive into an active knowledge base today.

Key takeaways

The Over-Engineering Trap: Typical corporate RAG guides push five to eight web services and vector databases—while architecture and engineering firms already have 90% of what they need in SharePoint or a shared network drive.
The iterthink Local-First Model: Set documents_root to your mounted team share; RAG indexes locally (store.rag.sqlite3), caches enrichments in .iterthink-cache/ on the shared folder, and respects existing SharePoint permissions.
The Scalable Roadmap: Start Phase 1 (1–15 users, local Ollama + shared cache), move to Phase 2 (central server + pgvector) only when cross-project search pain appears, and Phase 3 (yourcompanyos integration) when IT governance demands it.

← All guides