ProLyt

Docs › Installing on your own server

Installing ProLyt on your own server

This guide sets up ProLyt on a server you control. App, database and AI model can all run inside your network; nothing has to leave it. The step-by-step runbook from the first real install (every phase, with the gaps it found) is docs/ONPREM_INSTALL.md in the source tree; this page is the reference.

Requirements

ItemMinimumNotes
Operating systemLinux with Docker and Docker Composethe hosted service runs Ubuntu 24.04 LTS; the installer adds Docker on a fresh box
App only (model served elsewhere)2 GB RAM, ~5 GB disk4 GB RAM is comfortable
App + a local 8B model (Ollama)12 GB RAM, ~15 GB disk8 GB works only with the model's context limited to 8,192 tokens (measured)
App + a local 14B model (Ollama)16 GB RAM, ~20 GB diskthe model alone takes ~10 GB at Q4/Q5; a GPU is optional but much faster
Building the image on the boxmore than 4 GB free RAM during the buildor build on another machine and load the image
NetworkHTTPS to your usersno outbound traffic is needed once installed

The real memory ceiling is file ingest, not idle use: an Excel upload peaks at about 26× the file size in RAM, a CSV at about 4×. Set MAX_UPLOAD_MB to match the box.

1. Get the software

ProLyt is installed from its source tree; there is no separate release package. On a box with internet access:

``` git clone <the repository you were given> /opt/prolyt && cd /opt/prolyt ./setup.sh prod-onprem-multi # several people sign in; one workspace, unmetered

or: ./setup.sh prod-onprem # a single implicit user, no sign-in

```

setup.sh copies .env.example to .env, seeds the target's settings and generates ADMIN_TOKEN and CREDENTIALS_KEY. For an air-gapped box, build the image on a connected machine and move it with docker save / docker load (see section 7). The scripted installer scripts/onprem_first_install.sh runs the same steps in phases and writes a report of what the box did.

2. Configure .env

SettingMeaning
DOMAIN, PUBLIC_URLthe address users open, e.g. prolyt.example.com
TLS_CONFIGempty = public domain, automatic Let's Encrypt; tls internal = the bundled web server's own CA (export its root certificate once for your fleet); tls /certs/cert.pem /certs/key.pem = your company PKI (files in deploy/certs/)
MULTI_TENANT1 = accounts and sign-in (the multi-user target seeds this, with DEPLOYMENT_POSTURE=internal and DEFAULT_PLAN=selfhosted, i.e. unmetered)
DEFAULT_MODELthe model used when a question names none
LLM_RUNTIMEollama (the bundled Ollama container, default), custom + CUSTOM_RUNTIME_URL for your own OpenAI-compatible server — see *Running your own inference*
EMBEDDING_MODEtable and document search; the on-prem targets seed local (a small model baked into the image, CPU only, no GPU needed)
CREDENTIALS_KEYencrypts stored database credentials, provider keys and chat transcripts. Back it up separately: without it the stored transcripts and credentials cannot be read, and a backup never contains it.
ADMIN_TOKENprotects the admin routes used to create workspaces
MAX_UPLOAD_MBupload ceiling (see the memory note above)

The Ollama container's context length is set on the ollama service in the compose file (OLLAMA_CONTEXT_LENGTH); CONTEXT_WINDOW_DECLARED in .env tells the app what you set so it can show it on the inference panel.

3. Start and verify

`` docker compose -f docker-compose.prod-onprem-multi.yml up -d --build python scripts/smoke_onprem.py --compose docker-compose.prod-onprem-multi.yml python scripts/deploy_verify.py --compose docker-compose.prod-onprem-multi.yml --service app --max-age 600 ``

smoke_onprem.py uploads a table, asks a question, pins a dashboard tile, then destroys and recreates the container and checks everything survived. Note that docker compose restart does not re-read .env; only up -d does.

4. Create the first workspace and owner

On the multi-user target:

`` ADMIN_TOKEN=<from .env> ./scripts/create_tenant.sh "Acme Corp" ``

This prints the workspace's API key and its invite code. Colleagues sign up at https://<your domain>/login with the invite code and join as members. Promote the first owner with the admin route POST /admin/tenants/<id>/owner (the owner can then promote others on the Workspace page). On the single-user target there is no sign-in.

5. Connect data and a model

6. Backups

`` scripts/backup.sh --dest /var/backups/prolyt --compose docker-compose.prod-onprem-multi.yml ``

The script stops the app briefly (both storage engines keep a write-ahead log; a copy taken while it is open is empty, not old), archives the two data volumes and a copy of .env without CREDENTIALS_KEY, and verifies the archive. Keep the key in a password manager. For a nightly job install scripts/analytiq-backup-cron.sh (it runs the backup and the verifier, and fails loudly if either fails); --keep N prunes old sets. Restoring is a procedure, not a script: docs/BACKUP.md. Rehearse it once on a spare machine.

7. Air-gapped installs

8. Updating

  1. Tag the running image so you can roll back: docker tag analytiq:latest analytiq:rollback-<date>.
  2. Update the source tree (git pull, or load the new image), keep your .env.
  3. docker compose -f <your compose file> up -d --build, then scripts/deploy_verify.py (it asserts the container was actually recreated).
  4. To roll back: docker tag analytiq:rollback-<date> analytiq:latest and up -d --force-recreate app (without --build).

Settings files are additive: a newer version reads older data; rolling back past a version that added a setting may need that setting removed first — the release notes say when.