Spaces:
Running
v1.0 is live! Feedback, ideas, and forks welcome 👋
Hi, maker here!
MosAIc just hit v1.0, so I wanted to open the floor.
The idea: instead of a linear chat transcript, the LLM emits a small JSON overlay that reshapes the surface into panels (code, tables, diagrams, checklists) to fit your task. Keep asking and it evolves the surface (patches what's there) instead of regenerating from scratch. BYO LLM via Inference Providers, no backend.
A few things I'd genuinely love input on:
- The overlay schema as a "render target." It's deliberately tiny and the idea is that any producer (model, agent, IDE) could write to it. Does the contract feel right? (see SCHEMA.md)
- The wire format. JSON today, which is precise, but models fumble escaping/brackets and it can't stream. I weighed YAML / markdown / XML-tags / tool-calls in the ROADMAP.
Anything to share that's worked for you getting structured UI out of a model?
- Anything that feels janky.
NOTE! Known rough edge: the model infers "modify vs. start fresh" from your phrasing, so it sometimes regenerates when you wanted a small
tweak. Ideas welcome.
Issues, ideas, forks, spears, all are welcome.
Thanks for taking a look! 🎨
What a DAY!
Update! MosAIc now reads and works Google's OKF. 🧩
Since the v1.0 kickoff, the big addition: MosAIc opens a Google Open Knowledge Format bundle (the markdown + YAML-frontmatter "LLM-wiki" Google shipped this month) and renders it as a calm reading surface: one view per concept type, schema tables, cross-links that navigate in place, and a provenance strip (sourced/unsourced, source link, freshness), since OKF's frontmatter carries no trust field.
Two things past a viewer:
- Interactive. Signed in, you can point a model at the loaded bundle, "which concepts are unsourced?", "summarize how the tables join," and it reshapes, grounded only in the real concepts.
- Filter to scope. A "Filter concepts" box narrows the surface and what the model sees, so it holds up on bigger bundles.
No sign-in to try it 😀 pick "Open a sample OKF bundle" from the examples. Feedback very welcome, especially on provenance as OKF's missing layer.