SolomindLM

AI-powered research platform for multi-source content ingestion, RAG-based chat, and automated content generation.

Features

• Content Ingestion — Upload PDFs, documents, images, and audio; scrape web pages; import YouTube, TikTok, Instagram, and X transcripts; search and add sources via Tavily
• RAG Chat — Streaming chat with citation support across all ingested sources
• Content Generation — Reports, flashcards, quizzes, mind maps, audio overviews, slide decks, and spreadsheets
• Notebooks — Organize research into notebooks with sources, chat history, and generated content

Tech Stack

Layer	Technology
Frontend	React 19, Vite, TypeScript, TailwindCSS 4, Radix UI
Backend	Convex 1.31+, TypeScript, LangChain
Auth	@convex-dev/auth
LLMs	openai/gpt-oss-120b, openai/gpt-oss-20b
Embeddings	OpenAI (text-embedding-3-small)
Reranking	ZeroEntropy (zerank-2)
OCR	Mistral
Web Search	Tavily
Content Extraction	Supadata
Audio (TTS)	Together AI

Getting Started

Prerequisites

• Bun v1.2+
• A Convex account

Installation

bun install

Configuration

Convex backend — copy .env.example to .env and fill in:

CONVEX_DEPLOYMENT=<your-convex-deployment-url>
# AI service keys: Together AI, OpenAI (embeddings), Mistral, Tavily, Supadata, ZeroEntropy

Web frontend — copy apps/web/.env.example to apps/web/.env.local and fill in:

VITE_CONVEX_URL=https://your-deployment.convex.cloud

For production, set the prod Convex URL in your hosting environment (e.g., Vercel env vars). Dev and prod deployments use different URLs.

Deploying to Vercel

Project settings

• Connect the GitHub repo and set Root Directory to apps/web (monorepo). Vercel reads apps/web/vercel.json from that root.
• Use the same Bun major version as package.json (packageManager: bun@1.2.x).

Build behavior

• Production (main / production branch): runs convex deploy then the web build, so backend and frontend stay aligned. Requires a Convex deploy key on Vercel.
• Preview (pull requests / other branches): runs only bun run build:prod — it does not push Convex code, so previews cannot overwrite production Convex. Point Preview env vars at a dev Convex deployment.

Checklist so production Vercel always deploys Convex

1. Vercel Project → Settings → Environment Variables (Production): set **CONVEX_DEPLOY_KEY** (from Convex dashboard → Project Settings → Generate deploy key; same project as your prod https://….convex.cloud). Without it, bun x convex deploy in the build fails and the build does not complete.
2. Do not change apps/web/vercel.json’s buildCommand to drop convex deploy for non-preview. CI asserts this on every run (grep on vercel.json).
3. Do not add **convex deploy** to GitHub Actions for main — it can race with Vercel and return 409 ExistingModuleHashConflict (see below).

Environment variables

Use **VITE_CONVEX_SITE_URL** as the single value for both the browser (Vite) and the Vercel edge /api/* proxy. Per Vercel routing docs, only routes[].dest supports ${VAR} expansion with an env allowlist — so the proxy reuses the same name as your client env.

Set it to your Convex HTTP site base: https://YOUR_DEPLOYMENT.convex.site (no trailing slash).

Name	Production	Preview
CONVEX_DEPLOY_KEY	Required on Vercel (deploy key)	Omit
VITE_CONVEX_SITE_URL	Required — same URL for chat/streaming and /api/* proxy	Required — dev .convex.site
VITE_CONVEX_URL	Set automatically during Production build by convex deploy --cmd	Required — dev https://….convex.cloud

If you previously used **CONVEX_SITE_ORIGIN**, delete it in Vercel and set **VITE_CONVEX_SITE_URL** to the same value (one variable instead of two).

Routing: vercel.json uses a single **routes** array: filesystem (static assets from dist) → /api/* proxied with ${VITE_CONVEX_SITE_URL} ([env in dest](https://vercel.com/docs/project-configuration/vercel-json)) → catch‑all to /index.html for the SPA. No separate rewrites block.

CLI (optional)

npm i -g vercel
vercel login
cd apps/web
vercel link
vercel env pull .env.local
vercel deploy
vercel deploy --prod

CI vs Vercel

Do not run convex deploy from both GitHub Actions and Vercel on the same push: concurrent pushes to one deployment return 409 ExistingModuleHashConflict.

• Vercel Production is the only place that should run convex deploy (see apps/web/vercel.json).
• On **main**, GitHub Actions runs typecheck plus bun run build:prod with repository variable **VITE_CONVEX_URL** set to your prod https://….convex.cloud URL (the same value convex deploy --cmd injects on Vercel). No CONVEX_DEPLOY_KEY is needed in GitHub for that job.
• Pull requests still run typecheck + Vite build with a placeholder Convex URL (no deploy).

Running Locally

# Terminal 1 — Convex backend
bun x convex dev

# Terminal 2 — Web frontend
bun run dev:web

Open http://localhost:5173.

Project Structure

SolomindLM/
├── apps/web/          # React frontend
│   └── src/
│       └── features/  # auth, chat, notebooks, sources, studio, billing
└── convex/            # Convex backend
    ├── _agents/       # LangGraph agents (chat, report, flashcard, etc.)
    ├── _services/     # External service integrations
    ├── studio/        # Content generation modules
    └── *.ts           # Schema, functions, auth config

License

Private — all rights reserved.