AI for Construction Documents
A spec book and a drawing set are some of the longest, most cross-referenced documents any industry produces. Getting AI to read them reliably takes more than a bigger model.
These posts cover the practical side — what it takes to run capable models on hardware you own, and how AI holds up on real engineering tasks — alongside the technical explainers on retrieval and context windows behind TeraContext.AI.
Own your model
Mid-sized models on your own hardware keep project documents in the building.
Field notes
What happens when AI is pointed at real engineering and design work.
-
AI for Documents
A Bridge's Day: Sun, Shadows, Trucks and Code Checks, Driven by Claude Code and Elmer FEM
Claude Code drove the open-source Elmer FEM solver through a hot July day, truck loads and AASHTO checks on Bridge B-676, and independent checks caught eleven plausible-looking errors.
Read the post
-
AI for Documents
Steam Shovels for Design Engineers: Driving Civil 3D, Revit, and Navisworks with Claude Code
In about 32 prompts, Claude Code drove Civil 3D, Revit, and Navisworks through a full road-and-bridge BIM project: ~$942 versus ~$32,000 of drafting.
Read the post
-
AI for Documents
The Steam Shovel Moment for Civil Engineering Drafting
Three AI agents got the same week-long Civil 3D homework: Claude Opus finished in 65 minutes, unattended; two local Qwen runs stalled.
Read the post
-
AI for Documents
Two AIs, One Subdivision, Ninety Minutes
Grok and Claude each laid out a 20-lot Civil 3D subdivision from a one-line prompt in 45 minutes. One solved it; one only looked finished.
Read the post
- What Happens When You Let an AI Build Whatever It Wants I told an AI to build whatever it wanted, then left the room. It made a fantasy-map generator — and later gave those maps a thousand years of history. Read the post
How long-document AI works
Why a 2,000-page spec book doesn't simply fit in a context window, and the retrieval techniques that make it searchable.
- The Gambler - Managing Your RAG Mastering retrieval in RAG systems using semantic search, keyword search, graph search, and re-ranking—knowing what to throw away and what to keep. Read the post
- RAG vs. GraphRAG: Choosing the Right Approach for Your Documents Understanding the differences between RAG and GraphRAG, when to use each approach, and how to combine them effectively. Read the post
- RAPTOR and Multi-Layer Summarization: Building Hierarchical Document Understanding How RAPTOR and related multi-layer summarization techniques create hierarchical document understanding for more effective AI interaction. Read the post
- The Evolution of Context Windows: Why Bigger Isn't Always Enough LLM context windows have grown dramatically, but real-world documents still exceed these limits. Here's why and what to do about it. Read the post
- Why 1M Tokens Isn't Enough: The Mathematics of Context Windows Why 1 million token context windows are less impressive than they sound: the math, the scaling challenges, and the practical limits of LLM context. Read the post
- Mamba vs Transformers: Rethinking Attention for Long-Context Processing How Mamba's state space models challenge transformer dominance for long-context workloads through linear-time complexity and selective attention mechanisms. Read the post