Operator console¶

Prerequisites

Stack running: make up from the repository root (see the README for first-time setup).
Console accessible at http://<your-server>:8000/console/ (or http://localhost:8000/console/ for a local stack).
A user account: either the seeded dev admin (admin@example.com / admin12345) or one you registered.

The console is a browser UI bundled inside the app container — no extra service, no extra port, no install. Once the stack is running, open:

http://<your-server>:8000/console/

(The bare / redirects there.) Everything below can also be done with raw API calls — the console is a convenience over the same API your applications use.

The first screen offers Sign in and Register. On a development stack a default admin is already seeded (admin@example.com / admin12345 — disable with ADAPTA_SEED_DEFAULT_ADMIN=0), so you can sign in immediately.

Registering creates a new organization with you as its admin — fill in an org name, email, and password. After that, you land on the Projects page. Additional team members join an existing organization via invitations (an admin issues a one-time invite token; the new user redeems it to set a password and join the team).

The top bar shows your user and a team switcher — projects are always scoped to the active team.

2. Create a project¶

Click New project. The console asks "How do you want to specialize your model?":

Give it knowledge (RAG) — for answering from your documents.
Change how it behaves (adapter) — for tone/format/skill; the adapter is imported, not trained here (the upload is converted and gated on CPU).

Pick one, name the project, and choose a base model from the catalog (e.g. a Qwen2.5 model). The base model is the foundation both paths build on.

The project card then opens to the flow for its type, with tabs: Setup, Endpoint & keys, Playground, and Usage. Tabs stay disabled until their prerequisites are met (e.g. you can't open the Playground before an endpoint exists).

3a. Knowledge (RAG) flow¶

In the Setup tab:

Upload documents — drag-and-drop PDF, DOCX, TXT, MD, HTML, or image (PNG, JPG, TIFF) files. Each file is parsed, split into chunks, embedded, and stored in this project's private vector collection in the background.
Watch the status — each file moves pending → processing → indexed (or failed with a reason). The Create endpoint button stays disabled until at least one file is indexed.
Create endpoint — once you have indexed documents, create the endpoint. RAG has no eval gate — the model's weights aren't changing, so there's nothing to verify. You go straight to keys.

What you get: a model that answers grounded in your documents, with citations — the source chunks it used are returned alongside the answer.

3b. Adapter (finetune) flow¶

In-product training was removed on 2026-10-10 (see the import guide): the console no longer trains anything, and a finetune project cannot create an endpoint until an imported adapter has passed the eval gate (import: POST /v1/projects/{id}/adapters; the adapter job on the CPU worker marks it passed or failed). What the Setup tab does today:

Documents (optional) — upload PDF/DOCX/TXT/MD/HTML/images, same as the RAG flow. They are the source the Synthesize button reads, and once an adapter is served the model can answer from them with citations. See Knowledge + behavior together.
Provide a dataset — the project's eval set, used to judge an adapter against the base model. Either:
- Upload JSONL — one instruction pair per line: {"prompt": "...", "response": "..."}, optional system. The file is validated against the eval-dataset schema; rows that don't match are reported.
- Synthesize — generate an eval set from the documents you indexed in step 1, with the project's own base model (or an optional OpenAI-compatible teacher URL). Two modes: RAG eval (questions with the expected source file — the same shape as the golden set, for measuring retrieval) and Behavior (rows shaped by the task you pick — classify or extract recommended — usable directly as an adapter eval set). Near-duplicates are dropped. It runs in the background (≈ 20–30 s for a dozen short documents on CPU); download the result as JSONL from the dataset row. A teacher URL that resolves to a link-local or cloud-metadata address (for example 169.254.169.254) is refused with a 422 unsafe_teacher_url; localhost and LAN teachers are allowed.
Reviewing and curating a dataset — once a dataset is valid, a Review button appears next to it. It opens a paginated row viewer where you can inspect each pair, drop individual rows that look wrong, and edit text fields inline. Save curated writes a new derived dataset (marked valid). This lets you clean a synthesized dataset without re-uploading.

The adapter import and its CPU eval gate run on upload; the endpoint step unlocks for a finetune project once an adapter has passed.

4. Endpoint & keys — the handoff¶

In the Endpoint & keys tab:

The endpoint card shows the endpoint slug — this is the value your application passes as the OpenAI model name.
Create a key — you get an adp_… key. The full key is shown exactly once, on creation. Copy it then; afterward only a masked prefix is visible. Keys are scoped to this one endpoint.
A copy-paste OpenAI SDK snippet is generated with this server's URL, the slug, and the key filled in — the bridge to the real API.

You can create multiple keys and revoke any of them; a revoked key stops working immediately.

5. Playground¶

The Playground tab lets you test the endpoint from the browser using a console-held key — the exact same request path a customer app uses. For RAG projects it shows the citations — each with its file, page and, for scanned pages, the OCR confidence; expand one to read the passage; for all projects it shows token usage (prompt / completion / total). It's a test tool, clearly labeled as such.

Retrieval quality. On the Endpoint tab, a RAG project with a golden set (POST /v1/projects/{id}/golden: a JSONL of {q, expected_file, lang} rows, or a synthesized rag_eval dataset) shows its latest recall@5 and MRR. A run is queued after every successful indexing (one run per burst of uploads) or on demand; it is flagged "Retrieval got worse" when recall drops by more than max(0.05, 1/questions) against the median of the previous three runs, so a single flipped question never warns. It is a warning, not a block.

6. Usage¶

The Usage tab shows daily token rollups (newest first) and totals for the endpoint, so you can see consumption over time.

7. Models catalog¶

The Models page (sidebar → Models) shows every base model the server knows about and — crucially — whether it is available (the GGUF file is present on the server). One entry per model, with:

Modality badge — text or vision; the platform default carries a Default badge.
License badge — the weights license (apache-2.0, or qwen-research), linking to the license text. Research-licensed entries are marked Evaluation only — not licensed for commercial use, are never the default, and are labelled the same way in the project picker.
Purpose (use-case line) and Best for chips — e.g. knowledge retention, code, vision.
Resource requirements — approximate RAM needed to serve, derived from the model's spec entry.
Availability status — green Ready if the GGUF is on disk; amber Not available if the model is in the catalog but the file hasn't been downloaded yet (see Downloading models).

When you select a base model in the New project dialog the same information appears as a live detail panel below the dropdown, so you can check requirements before committing. Models that are not available are listed as disabled options — they appear in the picker so you can see what's on offer, but the Create button stays disabled until you choose one that is ready.

8. Project overview¶

Every project card on the Projects page carries a stage indicator — a colored dot and label derived from the project's current state:

Stage	Meaning
Setting up	Project just created, no data yet.
Awaiting data	No files or datasets uploaded.
Indexing	Documents processing in the background.
Ready to serve	Ready for an endpoint to be created.
Live	Endpoint is active and serving.

A next-action hint below the label tells you what to do next ("Upload documents", "Upload a dataset", etc.).

Inside the project, the Overview tab gives the full picture at a glance:

Pipeline checklist — which steps are done (documents indexed, dataset uploaded, endpoint created, key generated). Each incomplete step has a call-to-action link to the right tab.
Endpoint snapshot — slug, status, model composition (base + adapter + retrieval layer), indexed-chunk count, active-key count.
7-day usage — request and token totals.

The Overview tab is always visible, even for an empty project, so you can orient yourself and find the next step without hunting through tabs.

9. Settings¶

The Settings section (sidebar → Settings) has four tabs.

Account¶

Change your password. Enter your current password to confirm, then set a new one (minimum 8 characters). The change takes effect immediately.

Team¶

Members table — all users in the team, their email addresses, roles, and join date. Anyone in the team can see this; only admins can invite people.
Invite a member — enter an email, pick a role (admin, member, or viewer), and click Send invite. A one-time token is shown. Deliver it out-of-band to the new user; they redeem it at the console's Accept invite screen to set their password and join the team.
Pending invitations — tokens that have been issued but not yet redeemed, with their expiry times.

Roles:

Role	Projects	Files / datasets	Endpoint / keys	Platform settings
admin	read + write	read + write	read + write	read + write
member	read + write	read + write	read + write	read-only
viewer	read-only	read-only	read-only	read-only

Platform settings¶

Per-team inference knobs (temperature, top-k, chunk size, RAG hit count, etc.). Each setting shows its current value, the source (override you set here, environment variable, or built-in default), and the allowed range.

Admins can set an override that replaces the env/default for this team's endpoint. Click Reset to remove the override and fall back to the env value or default. The settings page deliberately does not expose the eval-gate threshold — that is a platform invariant, not an operator knob.

Priority order: DB override > environment variable > built-in default.

System¶

Read-only status dashboard: Postgres, Redis, Chroma connectivity, disk usage, and memory usage — the same signals as GET /health/deep presented visually. The GPU detection result (GET /gpu) is also shown here; nothing in the stack requires one.

Next: Consuming the API — wiring a real application to the endpoint.