Configuration

A knowledge base has a config file and a content folder.

Config file

~/.config/zemi/config.toml, or the path in ZEMI_CONFIG. Defaults are in src/config.jl. Changes made on the site go to config.settings.toml beside it, which overrides the config file key by key: remove a key there to let the config file's value apply.

[group]
name = "Reef Ecology Lab"            # site title
field = "coral reef ecology"         # given to the agents; read after "in" and "about"
description = "We study how reef communities recover after bleaching, from field surveys and models."
restarting = "Surveying the reef"    # the page shown while the service restarts (see Deployment); default: Back in a moment

[[group.focus_areas]]                # what the landscape and the arXiv watch rank papers against
name = "Bleaching and recovery"
description = "Coral bleaching, mortality and the recovery of reef communities after heat stress"

[[group.panels]]                     # the front page's panels; default: Papers, Tutor, Findings, Projects
title = "Field surveys"
text = "Each reef's surveys since 2019, with what they found."
link = "/findings"                   # optional

[watch]
categories = ["q-bio.PE", "physics.ao-ph"]   # arXiv categories to watch
days = 7                             # how far back each check looks
hours = 24                           # how often `zemi` checks (0: only when asked)
read_from = 2                        # papers rated this or higher are read in full (0-3; 4: none)
record_from = 1                      # and these only recorded

[landscape]
clusters = 5                         # default: from the library's size

The group's field and description go into every agent's prompt. [[group.questions]] replaces the tutor's questions about the reader's background.

Data and content

data_dir (default ~/.local/share/zemi) holds the database, PDFs and built site. content_dir (default content in data_dir) holds what people write. Relative paths are resolved from the config file's folder.

pages/        site pages in Markdown; index.md replaces the front page
studies/      one folder per study: study.toml and drafts
notebooks/    Slate notebooks; a paper's under papers/<paper>/
assets/       logo, front page images (hero-light.svg, hero-dark.svg), images pages use

Backups

The knowledge base snapshots itself every hour, into backups beside the data folder (the data folder's path with .backups added, ~/.local/share/zemi.backups by default). A snapshot holds the database, the config folder, the content and the records, packed, and a copy of the uploads, PDFs and figures. A file unchanged since the last snapshot is hard-linked to that snapshot's copy, so it takes no more space. The site, cloned codebases, Kaimon's index and the vectors are left out: they are rebuilt. The folder is readable only by its owner, since the config holds secrets.

[backup]
enabled = true                 # on by default
dir = "/mnt/backups/zemi"      # another disk, or a mounted share, survives losing this one
every_hours = 1
keep_all_hours = 24            # every snapshot of the last day
keep_half_daily_days = 3       # then one per 12 hours
keep_daily_days = 14           # then one a day; none older
files = true                   # false: the uploads only, not the PDFs and figures
min_free_gb = 5                # no snapshot when the disk has less free
max_gb = 0                     # past this, the oldest go first; 0 for no limit

dir must be outside the data folder. A small group's knowledge base takes a few MB a snapshot packed, and its files once: about 45 snapshots under 1 GB. Before the database is migrated to a new schema, the snapshot before-schema-<n> is taken and kept until deleted. The Status page shows the last snapshot and the backups' size, with Back up now; Deployment covers restoring.

Notebooks

Kaimon Slate notebooks, rendered through the local Slate hub, or the one on [notebooks] slate_port. ZEMI_SITE_NOTEBOOKS=0 leaves them out.

Other keys

Key	Default
ollama_url	http://localhost:11434	the embedding service
ollama_keep_alive	"24h"	how long Ollama keeps an embedding model loaded after its last use, so a search does not wait for it to load (Ollama's own default is five minutes)
qdrant_url, qdrant_key	Kaimon's Qdrant, else http://localhost:6333	the vector store, and its API key as "env:NAME" or "file:PATH"
collection_prefix	"zemi"	Qdrant collections are <prefix>_<layer>
mutool	"mutool"	MuPDF's command
contact	""	an email sent to arXiv and OpenAlex with requests
base_set	""	a file a new, empty knowledge base is seeded from
[[layers]]	five layers (see Models)	replaces the default layers
[search] layers, keyword	the layers on the strongest model, true	what a search uses unless it says otherwise
[ranking] k, weights	60, each ranking 1	how search fuses its rankings
[landscape] layer	the first paper layer searched	the layer papers are placed by
[media] max_mb	200	the largest file people and agents may attach
[media.backup]	none	a copy of every attached file, as [dataset] names a store