Delta: LLM conversation branching

Delta is a local-first, app/UX for branching conversations with LLMs. It allow you to easily rewind and branch conversations by respond to the model differently to take the dialogue in another direction.

I've found the approach useful for idea exploration, brainstorming and trying out different models for a use case. The branching paradigm also makes it easier to manage past context and avoid copying and pasting while iterating on an LLM prompt or application.

Setup

Instal dependencies

npm install

Run project

npm start

Build

To build a distributable version of the app as a zip or dmg, run

npm run make

The outputs can be found in the out folder.

Since this app's code isn't signed yet because it's in early development, you will need to run the following for macOS to allow you to open it.

xattr -cr /path/to/Delta.app

Migrations

Data for the application is stored in a sqlite3 database in app.getPath("userData"). Specifically

ls "/Users/$(whoami)/Library/Application Support/delta/delta_data/"

Database migration files can be found in src/main/migrations. Migrations are automatically applied on app start by ensureTables in src/main/db.ts.

To create a new migration file, run

npm run knex:migrate:make <migration_name>

then edit the created file to modify the database schema.

Since the application automatically runs all migrations on start, a convenience script is provided to downgrade the migration version in case you need to iterate on it during development

npm run knex:migrate:down

You can then modify the migration file and start the app again to auto-apply or manually run

npm run knex:migrate:latest

Features

• Cloud model API support through Vercel ai SDK
  • Currently supports Google, Anthropic, OpenAI and Ollama (locally)
• Realtime response streaming
• Full-text search support across conversations through sqlite3
• Semantic search support across conversations with sqlite-vec and nomic-embed-text through ollama
• Resizable and hidable chat and canvas views
• Easy response regeneration and copying
• Drag-and-drop image support (for models that support them)
• Conversation export in Obsidian Canvas format
• Chat export in JSON format
• Hotkeys to toggle sidebar, chat, canvas, chat input and input deselection

Roadmap / help wanted

• Add proper versioning
• Add Github action to automate and host builds
• Test on Linux, Windows
• Support OpenAI-compatible API server, configurable through settings page
• Add support for more models and provides
  • Improve model discovery, not relying on hardcoded strings wherever possible
  • Import model selector to enable search, provider hiding
• Remedy remaining type issues in the code
• Re-evaluate knex as a migration framework. Something with better Typescript support might be preferable.
• Refactor config/secret store to use ~/.config/delta/config.json rather than app.getPath("userData")
• Add hotkey documentation/configuration
• Improve build process to minimize inclusion of unneeded code and dependencies, to reduce npm start time (see forge.config.ts)
• Fix spellcheck support in chat input
• Add UI for full conversation database download/export