Weave is a "multiversal" generative tree writing tool akin to loom
. It supports multiple generative backends such as:
- ✅
drama_llama
- llama.cpp wrapper supporting all llama.cpp models - ✅ OpenAI models
- ✅ Shim for GPT 3.5+ chat completions API, including GPT 4o.
- 🔲 Anthropic models (coming soon, using
misanthropic
)
- Download a release from the releases page and extract it.
- For macOS and Linux installation will be straightforward with an
.app
and a Debian package containing a static binary. - For Windows, a release is not yet provided but it will probably build for
openai
withcargo build --release --features="openai,gui"
. LLaMA will require a bit more work to build on Windows.
Notable features:
- Live switching of backends - Generate part of a story with OpenAI and another part with LLaMA -- all without restarting the app.
- Streaming responses - Cancel generations in progress -- both local and online.
- Live editing - Edit posts during generation. New tokens are always appended to the end.
- Advanced sampling controls - For local language models. Use any sampling methods in any order.
Coming soon:
- Multiple tabs and windows
- Per-story settings
- Anthropic models
Additionally, one goal of weave
is feature parity with loom
.
- ☑️ Read mode
- ✅ Linear story view
- ✅ Tree nav bar
- 🔲 Edit mode
- ☑️ Tree view
- ✅ Explore tree visually with mouse
- ✅ Expand and collapse nodes
- ✅ Change tree topology
- ✅ Edit nodes in place
- 🔲 Navigation
- ✅ Hotkeys
- 🔲 Bookmarks
- 🔲 Chapters
- 🔲 'Visited' state
- ☑️ Generation
- 🔲 Generate N children with various models (currently one a time).
- ✅ Modify generation settings (Complete for OpenAI and mostly for local)
- ☑️ File I/O
- ✅ Serializable application state, including stories, to JSON.
- ✅ Open/save trees as JSON files
- 🔲 Work with trees in multiple tabs
- ✅ Combine multiple trees
- This is alpha software. It may crash. It does auto-save, however you should export your work to JSON if you care about it. It is not guaranteed to be compatible with future versions, however we will try to maintain compatibility as much as possible.
- With each new generation, all tokens need to be injested again with most
backends. This is solvable with
drama_llama
(longest prefix cache) but not for the OpenAI API. So for OpenAI, it's recommended to generate larger posts. The system prompt is customizable so you can tweak the agent's instructions on verbosity. With Anthropic models, caching will be supported at user-defined breakpoints. - It is not currently possible to have a scrollable viewport so it's
recommended to collapse nodes if things get cluttered. This is because the
nodes are implemented with
egui::containers::Window
which ignore scrollable areas. This is fixable but not easily and not cleanly. Zoom does work, however (Ctrl + + and Ctrl + -).