Cortex

A self-hosted document chat app. Upload files, the server chunks and embeds them, and you can ask questions that get answered from the content you uploaded — not from the model's general knowledge.

It's built on Mastra for the agent/RAG plumbing, Ollama for local chat + embedding inference, Pinecone for the vector index, and a small Express API in front of it. The client is a plain React + Vite app.

What's in here

• server/ — Express API, auth (JWT + refresh token rotation), document ingestion, the RAG agent, and a web-search/crawl tool built on Firecrawl
• client/ — React frontend: sign in/up, document ledger, threaded chat

Each folder has its own README with setup details specific to it.

How it fits together

client (Vite, :5173) --/api proxy--> server (Express, :4000)
                                         |
                          +--------------+--------------+
                          |              |               |
                    Ollama (chat +   Pinecone        Postgres
                    embeddings)      (vectors)      (users/auth)

Documents go through parse-file -> chunk -> embed -> upsert into Pinecone. Chat requests go through the Mastra agent, which calls a vector query tool to pull relevant chunks before answering, and can also reach out to the web via Firecrawl when a question needs something the uploaded docs don't cover.

Threads and chat memory are persisted in LibSQL (server/data/mastra.db); user accounts and refresh tokens live in Postgres.

Running it locally

You'll need Node 20+, a running Postgres instance, a Pinecone account, an Ollama install (or compatible endpoint) with the models you want pulled, and optionally a Firecrawl key and a Google Safe Browsing key if you want the web tools to do anything beyond domain blocking.

npm run install:all   # installs server + client deps
npm run dev            # runs both with concurrently

Server comes up on http://localhost:4000, client on http://localhost:5173. See server/README.md for the environment variables you need to set before either of these will actually work.