2024 ROADMAP

[samlhuillier]

The vision for Reor is for it to become the self-organising knowledge base that runs LLM workflows and is backed by a knowledge graph.

What does that mean?

1. Reor provides a core set of API functions like search and filesystem functions which LLMs will use as tools.
2. Users can then define an LLM workflow which could combine any number of steps, inputs and tools.
3. That LLM workflow can run anywhere in the app: in chat, within the writing assistant, as a cron job that runs daily or something that summarises a webpage every time a link is pasted, etc.

Ultimately, every note will be a node in a knowledge graph with a set of properties and relationships to other notes. Really powerful things can then be done like automatically linking notes and easily querying notes via the natural language->LLM->SQL flow.

Key upcoming tasks:

• Chat UI Improvements #344
• Migrate LLM api calling to renderer process #346
• Writing assistant improvements #350
• Tool calling in chat #347
• Allow for custom system & user prompts that can be persisted and reused

Other important tasks:

• Keyboard shortcut #234
  • Modal pop up for chat and search
• Multi line issue in markdown
• Hybrid search
• Improved embedding scheme/retrieval
  • And temporal reranking for notes. Or potentially just leverage the LLM for this by feeding it the dates notes were edited
• Google Keep like view of all notes
• Instant create of a note, rather than having to give a note a name before it opens
• Allow back and forth buttons to navigate on chat histories as well as notes
• AI generated UIs
• Tsne/umap knowledge graph
• Branching features on chat. Maybe even a visualisation on branchings.
• Drag and drop files to chat
• Improve prompt library to allow users to see what's possible
• Allow follow ups in writing assistant
• Pass full vector db item properties to a chat for citations and title and date time analysis
• Migrate chat store away from electron store to json files within the user vault directory
• Migrate UI components to shadcn
• @ syntax for bringing files into chat context #467
• Paging in search tool for model to be able to look deeper into results
• RAG to index backlinks
• LLM to auto generate backlinks
• Improve recommendation/support for attaching models (perhaps even recommended llm for your system)
• Allow users to provide some context about their work and themselves to the model
• Better filesystem (new note gets created in the right folder)
• Chat within a particular folder
• Search tool (for llm) that uses the file name rather than content for search
• Chat in sidebar (not just as a main full screen page)
• Chat and editor split screen
• Image support in editor #477
• Google search tool
• Integrate voice mode models
• Fix backlinking
• Cuda & non-cuda binaries for windows and ubuntu
• Improve chat retry and stop generation controls
• Tags
• Update video in readme
• Update welcome to Reor note
• Support multiple file types and pdfs
• Support other model providers (like azure)
• Improve reliability around large files (files get automatically embedded even if they are very big)
• Migrate from transformers.js to ollama for embeddings
• Improve electron store interface
• Rename file should not overwrite content if it already exists
• Show api name under model dropdown
• Fix content security bug
• Allow for editing of custom llm apis
• Update app automatically
• Customise dialog implementation for shadcn so it works seabed from context
• minimize and close buttons on linux
• date filters in search tool should allow both date created and date modified
• If search tool is available, force search tool rather than 'initial knowledge base search'
• Fix/investigate bug with duplicate search entries for each note
• Show search tool query arguments
• To allow easy prompt reuse, we can just proactively store prompts for users that they often use and just predict them
• Fix object has been destroyed bug
• Allow cancel generation

Of course, the above is not set in stone. We are very keen for contributors to propose ideas to work on. Check out our contributors guide to get started 😄