# Dashboard Tour

> A guided tour of the Contextator dashboard — signing in, the account menu, top bar, project list and detail, keyboard shortcuts, footer and the Users page.

- Updated: 2026-11-04
- Source: https://contextator.com/en/docs/dashboard-tour/
- Language: en-US
- Author: Muhammet Şafak

---
The dashboard at **http://localhost:3444/** is where everything is managed. It is a plain web page —
no build step, and it polls so you can watch indexing happen. It asks you to sign in first, and what
you see afterwards depends on your account.

---

## Signing in

Opening `/` sends you to `/login` unless you already have a session in that browser. Sign in with the
account the operator gave you; on a brand-new instance there is no account yet and you are sent to
`/setup` instead, to create the first one with the code from the server log. Both are covered in
[Accounts and Permissions](/en/docs/accounts-and-permissions/).

If your password is a temporary one, the dashboard sends you to `/change-password` and nothing else
works until you replace it.

What you see once you are in follows your role: a `root` or `admin` account sees every project and the
**Users** page, a `member` account sees only the projects it was added to, and buttons it may not press
are shown disabled (Re-index and Force re-index, for one) rather than left out.

## Account menu

Top right, showing your name and a pill with your role:

| Item | What it does |
|------|--------------|
| **Change password** | Asks for your current password, then the new one twice. Saving it signs out your other sessions |
| **API tokens** | Your own tokens for the [Admin API](/en/docs/admin-api/#api-tokens) (`#/~tokens`) |
| **Users** | The account list — `root` and `admin` only. See [The Users page](#the-users-page) |
| **Audit log** | The instance-wide record of who did what (`#/~audit`) — `root` and `admin` only. See [Admin API](/en/docs/admin-api/#audit-log) |
| **Sign out** | Ends this session on the server, not just in this browser |

## Top bar

The status pills tell you whether the server is healthy before you look at anything else:

| Pill | Meaning |
|------|---------|
| **Database up / down** | The database is reachable — the embedded PostgreSQL or the external one named by `DATABASE_URL` |
| **Model ready / loading** | The embedding model is loaded. On a first start this says *loading* while it downloads — indexing and search do not work until it is ready |
| `local · Xenova/multilingual-e5-small · 384d` | The embedding provider, model and vector dimensions currently in use |
| `n open sessions` | MCP clients connected right now |
| `v1.0.0` | The running version |

## Project list (left)

Every project with its live status: `idle`, `indexing`, `queued` or `error`. Each row shows the number
of sources, the document and chunk counts and when it was last indexed. While a project indexes, the
row grows a progress bar with the current phase, files done and chunks embedded. A queued project tells
you which project it is waiting for.

- **Filter projects** — type to narrow the list.
- **New project** (or press `n`) — opens the creation dialog.

## Project detail (right)

### Header
The project name, its status, and the **MCP URL** with a copy button — this is the string you paste into
your agent's configuration. Next to it, the names of the project's sources.

Three actions:

| Button | What it does |
|--------|--------------|
| **Re-index** | Queues an incremental run: sync every source, re-embed only what changed |
| **Force re-index** | Drops every chunk of the project and rebuilds from scratch |
| **Delete** | Removes the project, its documents, chunks and files. Click twice to confirm (the second click reads **Confirm delete**). Drawn for `root` and `admin` accounts only, and refused while the project is indexing |

While a run is in progress a callout shows the phase, the progress bar, how many files were unchanged
and how many chunks were embedded. If the last run failed, a red callout shows the error verbatim.

### Statistics
Documents (and across how many sources), chunks, when it was last indexed (with what that run changed),
and the embedding model. If the project was indexed with a different model than the server now runs,
a warning appears here and search is refused until you re-index. The warning is a callout titled *Indexed with
a different embedding model* with a **Re-index now** button in it, because nothing re-indexes the project by
itself. A second callout, *Chunks are larger than the model reads*, appears when `CHUNK_MAX_TOKENS` is above
what the model represents and names the value to set before you re-index. The chunks tile reads *embedded,
HNSW cosine index*.

### Search
A search box under the statistics that asks the project the way an agent would. Optional **source**, **path
prefix** and **version** filters mirror the `search_docs` arguments, and the limit is 5, 10 or 20 hits. Each
hit shows its file, heading, the excerpt with its dimmed neighbouring text, its cosine score, and `D` and `L`
ranks — the position the vector half (`D`) and the keyword half (`L`) gave it. When the best score is under
the relevance floor, a *Below the relevance floor* callout says that `search_docs` would answer "no good
match", and the excerpts are shown anyway so you can judge whether the floor was right.

### What agents asked
The queries panel reads back what agents searched for, so you can see what the documentation fails to
answer. Pick the window (24 hours to 90 days) and whose searches count — agents (MCP), this dashboard, or
both. Four tabs: **Questions** (ordered by how often a question was asked and how low its best score
stayed, with how many tokens and days the asking is spread over), **Never returned** (documents no search
has returned, with how many searches ran while they were indexed), **Most returned** and **Volume** (searches
per UTC day). Nothing is labelled "unanswered": a question about something you do not document can score as
well as one you do, so the count and the score are shown and the judgement is yours. **Export questions**
downloads them as JSONL in the evaluation-harness shape.

Below the Questions table, **Reported by agents** lists what agents sent with `report_gap`: questions an
agent searched for, read what turned up and said the documentation does not answer, grouped by question and
shown as the agents' own words, as plain text. They are not ranked with the questions above them. See
[MCP Tools](/en/docs/mcp-tools/#report_gap).

Every figure describes one embedding model, one index generation and one relevance floor, named in the
selector above the table. When the floor was changed after those searches were decided, the panel flags it:
the figures were decided against the older floor, so what the project refuses from now on is not what they
show, and searches decided under the current floor are kept apart for you to pick. The foot of the panel shows
the **current relevance floor** and where it comes from — the server's `SEARCH_SCORE_FLOOR` or this project's
own value. A manager can pick another value from a list of ten in steps of 0.01 (the lowest 0.75, the
highest 0.84), **Off (0)**, or **Server default** to drop this project's own value; choosing one does
not apply it, but
prices it first — how many refused searches in the window it would have answered and how many answered
searches it would have refused, as upper bounds — and only **Set the floor to…** saves it. Managers can also
**Stop recording** or **Clear the log**; rows older than the retention period are deleted automatically.

### Eval runs
Under the project's golden questions (each a question and the file that should answer it), the **Eval runs** panel keeps the server's own answers to them. After
the index changes — a rebuild, an incremental run that wrote chunks, removed files or recorded document
changes — or after the set is edited, the server asks every golden question of the live index and records
the run. The panel draws the history of recall@5 and MRR as a chart, lists the runs with what each one lost or
skipped, and warns, naming the questions, when the newest good run lost one that was in the top five before
(and when a question was skipped for the first time because the file it expects is no longer in the index).
A run the server starts by itself shows up on an open page within seconds, without a refresh. A manager can
press **Run now**. The panel draws the figures the server recorded and scores nothing itself. The same runs
are read over the API: see [Admin API](/en/docs/admin-api/#golden-set-eval-runs).

### Document sources
One row per source: its mount name, where it comes from, its content type, document count, when it last
synced, and its own status. Each row carries:

| Action | Available for | What it does |
|--------|---------------|--------------|
| **Test** | git, Notion, Confluence, Documentation site, S3 bucket, Google Drive, GitHub issues | Checks the connection without indexing anything |
| **Sync** | all | Queues an index run (every source syncs at the start of one) |
| **Edit** / **Files** | all | Opens the source dialog; for uploads, lists and manages the stored files |
| **Delete** | all | Removes the source, its documents and its files. Click twice to confirm |

A viewer sees **View** in place of the action buttons.

**Add source** opens a dialog with one tab per kind — Local directory, Git repository, Upload files,
Obsidian vault, Notion, Confluence, Documentation site, S3 bucket, Google Drive, GitHub issues. See
[Document Sources](/en/docs/document-sources/).

### Secrets and personal data
For an `editor` and above, a section on the project page lists what the PII scan found, by document path: the rule, the
position and a masked preview, narrowed by rule. It never shows a value, and it says separately when scanning is off
(`PII_SCAN=0`), when findings are not grouped (no `SECRET_KEY`), when documents hit the per-document ceiling and when
documents were cut at indexing. The scan only reports; nothing is masked or skipped. A viewer does not see the section. See
[Configuration](/en/docs/configuration/#scanning-documents-for-secrets-and-personal-numbers).

### Members
Who reaches this project, beyond the `root` and `admin` accounts that reach every project and are
therefore not listed. Each row is a `member` account with a dropdown holding its project role —
**viewer** reads, **editor** changes sources, uploads files and re-indexes — and a **Remove** button.
**Add member** offers the active `member` accounts that are not on the project yet; if there are none
it points you at the Users page. Only `root` and `admin` can change any of it; everyone else sees the
list read-only. A note at the foot of the panel follows the project's MCP mode: while the endpoint is public it says
that membership governs the dashboard, not the MCP endpoint; once a token is required it says the
endpoint answers only to a client holding one of the project's tokens; in account required mode it says only the accounts listed here, and administrators, can reach the MCP endpoint. See [Accounts and Permissions](/en/docs/accounts-and-permissions/).

### MCP access
A pill says which of the three modes this project's endpoint is in — **open**, readable by anyone who
can reach the URL; **token required**, the default for a newly created project; or **account required**,
where only an account that is a member of the project is answered, over OAuth 2.1 and re-checked on
every request. The panel flips
between them; **New token** mints one and shows its secret once, and **Revoke** cuts its client off
mid-session. Minting offers a choice of sources to bind the token to, and a token with a scope shows a badge — red when none of
its sources are left — with a note on `open` projects, where a scope is not a boundary. Each row shows a token's
name, its leading characters, and when it was last used. Buttons you are not allowed to press are not
drawn: flipping the mode needs `manager` rights over the project, minting and revoking an editor's. See
[Projects](/en/docs/projects/).

### Connect an agent
Copy-paste snippets for Claude Code, Cursor, Claude Desktop and legacy SSE clients, filled in with this
project's real URL — and with the `Authorization` header already in place when the project requires a
token. See [Connecting AI Clients](/en/docs/connecting-ai-clients/).

### Tools exposed to the agent
A reminder of the five tools every project publishes: `search_docs`, `list_topics`, `read_document`, `whats_new`
and `report_gap`, the last present while the project's query log is on.
See [MCP Tools](/en/docs/mcp-tools/).

### Index runs
The recent run history, newest first: when, incremental or force, what changed
(`12 unchanged · 3 updated · 1 removed`), and how long it took. Failed runs show their error. This
survives restarts, unlike the live progress view.

---

## Keyboard and small conveniences

- `n` opens **New project** (when no dialog is open).
- `Esc` closes a dialog.
- The selected project is in the URL hash (`#/my-project`), so you can bookmark or share a link to it.
- Destructive buttons require a second click, and the confirmation expires after a few seconds.
- Copy buttons fall back to a manual selection if the browser blocks the clipboard on a plain-HTTP LAN
  address.

## Footer

Every page of the dashboard ends with the same two-part footer. On the left, the copyright line with the
running version: *Copyright © 2026 Contextator v1.0.0 is a Tunedness production.* On the right, five
pages that open as ordinary URLs and need no account:

| Page | What it is |
|------|-----------|
| [`/about`](http://localhost:3444/about) | What Contextator is, how it works, and who builds it |
| [`/privacy`](http://localhost:3444/privacy) | What your installation stores, and the outbound calls the software can make |
| [`/cookies`](http://localhost:3444/cookies) | The one cookie there is — the session that keeps you signed in — and what it is not |
| [`/terms`](http://localhost:3444/terms) | Warranty disclaimer and what running an instance makes you responsible for |
| [`/license`](http://localhost:3444/license) | The AGPL in short, the full text at `/license.txt`, and the third-party components |

They are public on purpose: a legal notice only the operator can read is not a notice. They expose no
project data and reach neither the database nor the API.

## The Users page

Account menu → **Users** (`#/~users`), for `root` and `admin` accounts only. One row per account: its
username and display name, its role, whether it is **active**, **disabled** or has a **password change
pending**, how many projects it reaches, and when it last signed in.

| Button | What it does |
|--------|--------------|
| **New user** | Opens the creation dialog |
| **Edit** | Display name, e-mail, role, and the active switch |
| **Reset password** | Hands out a new temporary password, shown once, and signs that account out everywhere |
| **Disable** / **Enable** | A disabled account cannot sign in, and its open sessions end at once |
| **Delete** | Removes the account, its sessions and its memberships. Click twice to confirm |
| **SSO group mappings** | Maps identity-provider groups to projects and a role, when `OIDC_GROUPS_CLAIM` is used. See [Single Sign-On](/en/docs/sso/#groups-grant-project-memberships) |

**New user** asks for a username (lowercase, permanent), an optional display name and e-mail, a role —
`member` or `admin`, with `root` offered only to a root account — and a password. Leave the password
empty, or press **Generate**, and the server makes one. Either way it is shown **once**, in a dialog
with a copy button, and **Must change password at first sign-in** is ticked by default.

There is no separate button for ending someone's sessions: **Reset password**, **Disable** and
**Delete** each do it as part of what they are.

A button you may not press is greyed out and says why when you hover it — the last active root account,
your own row, or a root account when you are an admin. The full set of rules is on
[Accounts and Permissions](/en/docs/accounts-and-permissions/).
