Skip to documentation
eli.aidocs

Get started

Quickstart

Go from a fresh clone to editing documents and asking grounded, cited questions — locally, on a laptop. In local mode there is no login to configure: a local user and default workspace are created for you.

You need Node 26+ and Postgres 17+ (18 preferred) with the pgvector and pg_trgm extensions. Pick whichever Postgres is easiest for you below — everything else is the same.

  1. Get a Postgres with pgvector + pg_trgm

    Choose one. All three give you a standard Postgres reachable over `DATABASE_URL`.

    Postgres 18 + pgvector (non-superuser app role)bash
    docker compose up -d db

  2. Configure and install

    Copy the example env, set your database URL and app secret, then install and migrate.

    Codebash
    cp .env.example .env       # edit DATABASE_URL / APP_SECRET
yarn install
yarn db:migrate

  3. Run the web app and the worker

    The web process serves the UI and API; the worker runs background jobs (extraction, embeddings, scheduled runs). Start them in two terminals.

    Codebash
    yarn dev                   # web app on http://localhost:3000
yarn worker                # background jobs (separate terminal)

    AUTH_MODE=local (the default) auto-signs-in a local user with a default workspace — zero friction. http://localhost:3000 serves the marketing landing page; click Open app (or go straight to http://localhost:3000/open) and you land in your workspace — /open routes to your last-used workspace, falling back to the default.

  4. Seed the demo knowledge base (optional)

    To see the graph, chat, and agents working against real content with no API cost, seed the bundled demo. It wires a local Ollama provider and loads a densely wikilinked consulting KB. With ollama serve running:

    Codebash
    ollama pull llama3.3          # agent slot — reliable structured tool-calling
ollama pull qwen3-coder:30b   # chat / extraction
ollama pull nomic-embed-text  # embeddings (vector retrieval)

yarn tsx scripts/seed-demo.ts # loads the KB and points the workspace at Ollama

    The seed builds a “Meridian Advisory ↔ Northwind Trading Co” knowledge base: it ingests (chunk → FTS → embeddings), builds the entity graph (wikilink resolution + deterministic co-occurrence relations), and points the workspace at Ollama.

  5. Ask your first grounded question

    In the app:

    • Search the graph under Entities and explore it under Graph.
    • Ask a grounded question in Chat — answers stream with clickable citations back to the source chunk.
    • Create an agent under Agents that traverses the graph with tools, then inspect its full trace under Runs and any approval-gated tool calls under Approvals.

Agents need a reliable tool-caller

Agents require reliable structured tool-calling through the OpenAI-compatible endpoint. llama3.3 and gpt-oss parse cleanly via Ollama; some models (e.g. qwen3-coder) occasionally leak text-format tool calls under multi-tool prompts — set the agent slot to a known-good tool-caller in Settings.

Running on a shared or remote host?

Set AUTH_MODE=multi for email + password auth. A production host with a public APP_URL refuses to boot in local mode. See Deployment for the compose topology, TLS, and backups.