Lesson 7.1: Providers and models #

Which companies (or local servers) answer the assistant's requests, and which models may write drafts, are decisions of the administrator. This lesson reads the Models and providers page and explains the rules behind it: providers, keys, the probe and the write gate.

At a glance #

Learning objectives #

After this lesson, learners can:

• read EI: Models (Models and providers): providers, models, key presence, probe and test results;
• explain region, retention and "terms checked" of a provider;
• explain the write gate: only models with native tool calling that passed the probe (or are trusted) write drafts;
• run a Test prompt and read its result;
• decide which models editors may choose (expai/model_use, the allowed models).

Trainer's notes #

How to present it #

Sources: expai: For administrators and the providers, probe and local models pages under EI: Help (extension/expai/doc/providers.md, extension/expai/doc/probe.md, extension/expai/doc/local-models.md).

1. Providers (4 minutes). The extension speaks to Anthropic, to any OpenAI-compatible service and to a local open-weight model through Ollama. Each provider has a region (us, eu, local), a retention and the date its terms were checked (TermsCheckedAt; the page shows "not checked" until set). The shipped values are a starting point, not a legal statement.
2. The page (4 minutes). Setup > EI: Models: one section per provider with its models (Model, Id, Tools, Probe, Test, Drafts, Cost), the local model server (actions such as Start the server, Stop the server, Pull the model, Remove the model, Refresh the status, sent to a root-owned helper with Send to the helper), Load model, Test prompt, and MCP tools. Keys are never shown; only whether one is present.
3. The write gate (5 minutes). A model may create and change drafts only with native tool calling and either WriteGate=trusted or a live probe with a pass rate of at least 90 percent and every mandatory task passed (among them a prompt injection task). Otherwise it reads only: editors see "reads only". The probe is run from the command line (extension/expai/bin/php/expaiprobe.php).
4. Who may use which model (3 minutes). The site default model is always available; other providers' models need expai/model_use (limitation Provider), except providers listed as open. Users may choose among allowed models where the site lets them.
5. Fallback (2 minutes). When the chosen model cannot answer, the site default answers and the editor sees one line saying so.

Common questions #

Which provider should we use? That depends on your data protection rules, region, retention and budget; the documentation names no preferred one. Read the providers' terms and record the date.

Is a local model good enough? For short tasks; on an ordinary server a local model takes minutes per step and may only read (it must pass the probe to write).

Where are keys stored? In settings/override/expai.ini.append.php or an untracked key location, never committed; lesson 7.2.

Pitfalls #

• Opening the gate by hand for a model that failed the probe's injection task. Do not.
• Leaving "terms not checked". Read the terms and set the date.
• Testing with real content. Test prompt sends a request; use a neutral prompt.

Hands-on exercise #

Use the training installation with an administrator account (or follow on the projector).

1. Open Setup > EI: Models. Write down each provider with its region and retention, and whether its terms were checked.
2. For each model, read Tools and Probe: which may write drafts, which read only, and why?
3. Find the key status of each provider (present or missing). Confirm that no key is visible on the page.
4. Run Test prompt on the cheapest model with a neutral prompt (as the page offers) and read the result: time to the first token, tokens per second, total.
5. Open the result's link in EI: Log (action test).
6. On the dashboard page Exponential Intelligence, open Details and read what an administrator with expai/settings sees about the default model.
7. Write down which models you would allow editors to choose, and why.

Expected result #

• The learner's notes list providers with region, retention and terms state, models with their write gate state, and a test prompt result.
• The learner can explain why a model reads only.
• The trainer checks two learners' model choices and reasons.

Quiz #

1. What three things are recorded per provider for the data policy and the summary?
2. When may a model write drafts?
3. What do editors see next to a model that may not write? A. Nothing B. "reads only" C. "blocked" D. Its key
4. Which policy lets an editor choose models of a provider outside the open list?
5. What happens when the chosen model cannot answer?

Further reading #

• expai: For administrators
• The providers, probe, local models and provider support pages under EI: Help.

Related pages #

• Module 7: Administration
• Exponential Intelligence: course overview
• Trainers: start here
• Free documentation: this page is free to copy, translate and improve under the GNU FDL 1.3.