Correlo

A personal, local-first connection engine. Correlo takes your documents, splits them into fragments, embeds them, and then looks for the non-obvious links between them — the ones a keyword search would never surface. Every lead is offered, never imposed: you accept it, reject it, or dig into it in conversation.

A link that rests only on a shared rare word, or on two fragments of the same document, is ruled out by an explicit filter (lib/discover/claim-rules.ts): the engine never presents as a discovery what was already findable.

The explorer is the home screen: your boards, what each one holds, and what you touched last. The interface is in French.

Requirements

• Node.js 20+ (Next.js 16).
• Embeddings: Ollama running locally with nomic-embed-text (the default), or any OpenAI-compatible endpoint via EMBED_PROVIDER=api.
• A chat model: Groq by default (GROQ_API_KEY), or Ollama locally.

Getting started

npm install
cp .env.example .env.local   # fill in GROQ_API_KEY at minimum
npm run dev

The app listens on http://localhost:3000. The SQLite database is created and migrated automatically on startup: instrumentation.ts applies everything in migrations/ and then starts the job worker. No manual setup step.

Correlo cannot be hosted. The database is a SQLite file, and ingestion, discovery and synthesis all run inside the server's own worker: it needs a writable persistent disk and a process that stays alive. A serverless host provides neither — and a deployment that did succeed would be an empty app anyway, since the corpus is local and git-ignored. That is what "local-first" means here: the corpus never leaves the disk, and the only things that go out are the ones you configure (LLM_CHAT_PROVIDER, SEARCH_PROVIDER).

Screens

Route	Purpose
/	The board explorer — the home screen. The app opens on your boards, not on a document
/canvas/<id>	The workshop for one board: drop files, review suggestions, preview (/canvas leads there)
/conversation	Conversation over the corpus; this is where connections get argued about
/sources	The documents ingested, and the state of their processing
/suggestions	The connections found, to accept or reject
/syntheses	The syntheses produced around a theme

Configuration

Everything is documented in .env.example. The settings that change behaviour:

Variable	Default	Effect
DATABASE_PATH	./data/correlo.db	Local SQLite database. This is the whole corpus: backing it up is backing up the work
LLM_CHAT_PROVIDER	groq	groq or ollama. No automatic fallback between the two
GROQ_API_KEY	—	Required when LLM_CHAT_PROVIDER=groq
EMBED_PROVIDER	ollama	ollama (local, 768 dims) or api (2048 dims)
NVIDIA_API_KEY	—	Required when EMBED_PROVIDER=api on integrate.api.nvidia.com
SEARCH_PROVIDER	none	exa turns on the agent's web search. none ⇒ the web tools do not exist
WORKSPACE_ROOT	empty	Local directory the agent may read, read-only. Empty ⇒ the tool does not exist

Changing EMBED_PROVIDER switches the embedding space. A vector only means something in the space that produced it: after a change, run npm run reembed (--dry-run to look first). Until the corpus is aligned, the analysis refuses to compare two spaces rather than produce false links.

Credentials are read from .env.local, which is never committed.

Commands

npm run dev         # development server (migrations + worker)
npm run build       # production build
npm run start       # production server
npm run lint        # ESLint
npm run typecheck   # tsc --noEmit

Long-running corpus work — the server must be running (npm run dev), because its worker is what executes the jobs; these scripts only enqueue one and follow along:

npm run discover    # a discovery run (paths and options: --help in scripts/discover.mjs)
npm run reingest    # re-ingest sources (after a chunking change)
npm run reembed     # re-embed the corpus into the active space
npm run synth       # produce syntheses

Layout

app/          routes (pages) and API
components/   interface (explorer, workshop, conversation)
lib/          the engine: ingestion, embeddings, discovery, jobs, conversation
migrations/   SQLite schema, applied on startup
scripts/      CLIs for long-running work
data/         local SQLite database (git-ignored)

License

MIT — see LICENSE.