Maya

Maya is a chat client for LLMs to experiment with multi-agent and multi-human conversations.

It is built in React Native, with a FastAPI backend (The Oracle), and MongoDB datastore.

---

Run

• Backend - heroku local -f Procfile.local (lsof -i :8001)
• React - npx react-native start --experimental-debugger
• iOS - npx run-ios
• Types - ./scripts/types.sh
• Web - npm run web

Deploy

• iOS - xcode release build
• Android - ??
• Web - npm run dist (push web branch)
• Backend - (push dev/main)

---

Todos

0.1.0 build:

0.1.0 test:

0.1.0 ship:

Clear prod DB + set up bot users Testflight! Test prod settings on web

After 0.1.0:

• Android ?

• Index DB

• Improve performance of refresh command

• Ideally stream doesn't delay the full message collection // loading of points of view

• Make sure context is truncated properly for uncensored local model

• Inflection.ai for third model

• Toggle streaming

• Read speed

Schema changes

User feedback

Way to request features (so I can add ideas) “Add wisdom” to integrate into our knowledge base “Add ideas” for the app Way to report bugs

Unit economics

Introduce plans / IAPs

• Delay perspective until "points of view" tapped? Only for cheaper tiers?
  • Make sure it's not requested every time someone taps "points of view"

QA:

• Font bug in body
• Handle rapid chats - ie. if you message a string of messages the server should wait, adjust and not double respond
• Add test for contexts
• Oracle summary issue (?)
• Websockets issues: 1. always reopen when app opens, 2. see why it keeps closing
• Websocket server to client sending on a multi-worker server

Personalization

• Per user context / preferences
• Reading speed
• Turn off streaming
• Privacy terms (allow ppl to keep data private? on thread basis?)

Confidence / annotations

• Confidence FE

• Confidence BE

  • Pull confidence out of perspectivedata?

• Annotation submission/retrieval

• NEED TO FIGURE OUT WHERE TO STORE CONFIDENCES IN DB (separate from perspectivedata?)

• Annotations BE

  • Dedicated collection
  • Top contributors can earn? Dont tell them, just give them gifts Don’t want ppl to contribute in order to get paid

• RAG w/ community notes

• Embeddings w/ bias

Social / virality

• Invite to chat
• Trending topics in "new chat" page?

Data / context / model improvements

• Improve context handling, to handle chat / perspective contexts differently wrt summarization. It should prefer context from the perspective or main chat, and cleverly truncate from the core
• Relationship between agents that gets summarized and added to context?
• Relationship with user that gets summarized and added to context? Ability to clear
• Thumbs up / thumbs down data saving
• Experiment with deeper trained bots

Server comms

• Background data updates
• Update "last-refresh" to paginate data loading
• Lastrefresh functionality, so it doesn't reload chats all the time
• Ensure websocket isn't created before userid exists on client?

Performance

• Metrics for server request throughput
• Reduce bundle size for web

UI improvements:

• Markdown
• First stream response w/ bounce?
• Indicators for success/failure/loading - message send, chat creation, profile update, etc?

Experimental:

• Build web demo on maya url
• Add additional or customizeable perspectives/personas
• Build web editor for chatagents (login via sms code, add python for agents, or add webhooks for people to build/host their own)
• Finetune chatbot on my Make School emails?

Group chats

• Test real time
• Figure out how to handle contacts
• Search
• Force participant set for certain protocols, allow others
• Adding other humans in chat
• Build group protocols

Tests:

• Tests for oracle
• Update tests

Backlog

• Store which model sent which message (maybe in a separate "log" db?)
• Adjust reading speed in user profile
• Toggle for streaming vs replying
• Reduce space for new lines
• Markdown
• Better data / cache / download management so you're not pulling everything every time
• Refresh authtoken each session
• Swap Mongodb with async
• Audit all async/await
• Allow custom avatars (cloudflare CDN)
• Secure storage rather than async storage
• Figure out how to deal with syncing local and server in cases they get out of sync
• Push notifications
• Invite ppl via sms
• Host on maya url rather than txtai.co
• Expire login codes
• Enable multiple dynos via redis/pubsub
  • Needed for real time group chats
• OTA updates (expo?)

####################################################

Done

• Diff screen sizes
• Give feedback (report bug / request feature) - include analytics
• Refresh needs to load the other data (eg. perspective)
• Header for perspectives page
• Don't request new perspective every time
• Update related with latest messages
• Cancel button to stop stream (also stop stream on server? kill web socket and restart it?)
• Update messaging in discussion (maya + alternate between two others)
• Hide search + related when keyboard is open?
• Test updated user data model
• Make search boxes bigger on web
• Strip newline / whitespace from end of message
• Save context on perspectives, retrieve context from subthreads
• Update and test summarization code to handle perspective contexts
• Analytics
  • Set user - setUserId('user@amplitude.com');
  • Buttons/success for each step of onboarding
  • New chat
  • Share
  • Message send
  • Tap on points of view
  • Tap on search result
  • Tap on related
  • Tap on sidebar
  • Tap on chat
  • Tap on profile
  • Web view analytics
• Hash phone numbers in DB
• Save phone => hash in auth DB
• Welcome / disclaimer
  • We store messages with an anonymous userid, encrypted at rest
  • We will use anonymized data to improve responses
  • Questions: ashu@desaidata.com
• Favicon
• Sharing stuff for website
• Create concept of "thread" that is linked to chatid + messageid but is it's own chat
  • Use this to update messages
  • Use protocol to differentiate
• Make search results smaller
• Perspectives FE
• Perspectives BE
• Chatlist refresh, doing so in drawer screws with animation
• New chat doesn't work yet due to the way sidebar is rendered
• Sidebar is reverse chronological
• Profile needs to be fixed
• Build message refresh
• Build websockets
  • Helper function for websocket to create the right json
  • Enable websocket to route different commands
• Set chat topic
• Build out discussion protocol, to allow other devs to implement a different framework
• Build out discussion on server w/ multiple LLMs
• Hook up frontend
• Frontend chat and
• LLM streaming
• Speeding up streaming
• Test oracle, upload to dev server
• Run on actual phone
• Create new message, select protocol, select AIs
• 1:1 format list maps bot ID to protocol
• Calculate slug for each chatid
• Change search and results for everything
• One LLM needs to be uncensored
• Share link + web
  • Save hash in chat on creation
• Force update
• Don't stream the secondary responses - why so slow?
• Share functionality w/ web based viewer of chatid
• User prompt
• Ask clarifying questions to user
• Draft has a bug where it won't reset back to empty when creating a new message

---

Thought dump

• Explore "perspectives" modal
  • Offers new perspectives
  • Similar text interface as chat

Perspective analogy domain mapping pattern matching Consider this from another angle Globalist thinking 3D UX?

• Modal view to add annotations

• Exa bot

• Eval bot

• Community notes bot

---

Concept

Goals:

• Higher bandwidth information transfer
• Explore more perspectives // holistic perspectives
• Dead simple getting started UX (peel back layers of onion)
• Inherently social / sharable

Non goals:

• Maximize profit
• Non-text modalities (video, actions)

What does high bandwidth info transfer mean?

Explore information / conversations more like how our brains work. Less linear, more recursive https://a9.io/glue-comic/ Additional LLM perspectives Human annotations (ala community notes) to expand training data set Dig into threads, bring context back to macro discussion Allow threads to persist as their own things

---

Run, debug, deploy

Here's how I run things:

• Cursor IDE (VSCode w/ ChatGPT built in)
• Create VS workspace containing both "maya" and "the-oracle" repositories
• Hide everything non-essential in workspace settings, pull from finder

Extensions

• Prettier (typescript formatting)
• ESLint (typescript linting)
• Ruff (python linting)

Set up several terminals in primary workspace

1. the-oracle - split terminal (backend)

• server heroku local -f Procfile.local
• tests/types pytest, pytest -c populate.ini, ./scripts/types.sh

2. maya - split terminal (frontend)

• react npx react-native start --experimental-debugger
• ios npx run-ios, npx react-native run-ios --device toshphone, npx react-native run-android

3. the-oracle/maya split terminal (version control)

• git - the oracle
• git - maya

---

Updating types

Types update automagically from python into typescript!

./scripts/types.sh

---

Maya backend architecture

Maya chat operates on a websocket for the following functions:

• Sending/receiving messages
• Creating/updating chats

Additionally it has the following API endpoints:

• Websocket (manages message sending/receiving and creating/updating chats)
• Refresh (updates the chat list and recent messages)
• Update profile (updates a user's username/avatar)
• Auth / verify (sms based authentication)

Data types

ChatInfo Chats follow a discussion protocol "maya" is the base protocol "perspectives" is a protocol to offer more perspectives "annotations" is a protocol for annotations

---

Installation

Install/update xcode and command line tools

Install node, watchman (via homebrew), and cocoapods (view rubygems, ruby version 2.7.x)

brew install node
brew install watchman
sudo gem install cocoapods

See React Native - Environment Setup instructions till "Creating a new application" step, if you have trouble.

---

Running locally

1. Local mongodb (check command to run in background)

2. Local LLM (use OpenChat ollama)

ollama run openchat

3. Offline chatbot (Deepseek coder via Ollama and VSCode Continue plugin)

ollama run deepseek-coder:33b