When I started building Prism AI (DudeSquad999/Prism-Ai), the obvious design was a single window with a chat sidebar on the left and a document editor on the right. Almost every AI writing wrapper uses that layout.
After testing it on real research projects, that single-pane model kept getting in the way. Exploratory research and formal LaTeX writing have completely different failure modes, context needs, and model requirements. In v3.1.0 for Windows, I committed to splitting the workspace into two focused applications: Prism Chat and Prism Studio.
Here is what actually drove that architecture, what broke along the way, and the honest trade-offs of building offline-capable research software on Windows.
1. Exploratory Context vs. Strict Compilation State
In early research (Prism Chat), you want broad project context: messy notes, article extracts, scholarly abstracts, task lists, and Markdown drafts. If a model rambles or brainstorms three angles at once, that is useful.
In LaTeX writing (Prism Studio), loose output breaks your build. A single unescaped % or hallucinated \cite{key} stops PDF compilation cold. Separating Prism Studio allowed me to treat AI suggestions as reviewable diffs against .tex and .bib files rather than auto-mutating the document.
2. Separate AI Configurations Per App
One deliberate choice in v3.1.0 is that Prism Chat and Prism Studio do not share a global model lock. Configuring Ollama or a cloud API key (Google Gemini, Anthropic, or OpenAI) in Prism Chat does not overwrite Prism Studio.
Why? Because on consumer Windows hardware, you often want a fast, conversational local model in Ollama for organizing notes, while switching either to a stronger coding/math model or a cloud provider when debugging complex TikZ figures, Beamer slides, or multi-file LaTeX builds.
3. The Reality of "Offline-Capable" with Ollama and Tectonic
A lot of local-first tools claim "100% offline" and gloss over first-run friction. In Prism AI's documentation and installer flow, I chose to be explicit about where internet access is still required:
- Ollama: Inference runs locally on your CPU/GPU, but pulling Ollama and downloading model weights requires an initial connection.
- Tectonic: Unlike a 4 GB full TeX Live install, Tectonic downloads only the
.stypackages your document actually uses on the first compile and caches them locally. After that initial cache warm-up, PDF compilation works completely offline. - Offline Demo Fallback: If Ollama is stopped or unreachable, Prism Chat falls back to a local demo response mode rather than crashing, though the UI has to make it obvious when you are seeing fallback behavior versus live inference.
4. Shipping Safe Windows Installers as a Solo Dev
Distributing unsigned Windows .exe builds (Prism-3.1.0.exe) means users rightly scrutinize how they install your app. Instead of telling people to pipe a remote URL straight into Invoke-Expression, I updated the PowerShell helper (scripts/install.ps1) to download to $env:TEMP, verify the published SHA-256 checksum, and let users inspect the script before executing. I also maintain a static browser workspace on GitHub Pages (dudesquad999.github.io/Prism-Ai/) that can connect to a local Ollama instance so people can test the interface before installing anything on Windows.
Trade-Offs Still on My Plate
Splitting the apps solves context pollution, but it introduces handoff friction when moving a finished Markdown outline from Prism Chat into a structured .tex template in Prism Studio. Tightening that project-level bridge without turning the UI into bloatware is my main focus after v3.1.0.
Repository: https://github.com/DudeSquad999/Prism-Ai
Question for you: If you use local LLMs for technical writing or research, do you prefer one unified interface for everything, or separate tools for note-taking and final document compilation?
