Lesson 5.1: AI chains and the human checkpoint #

An AI chain writes in steps instead of one answer: a model writes, another reviews the text against a score and a checklist, the first revises until the review passes, a third prepares title and metadata and checks the links, and then an editor decides before anything else happens. This lesson runs one chain per learner and practices the checkpoint.

At a glance #

Learning objectives #

After this lesson, learners can:

• start a chain from the Intelligence tab (Write with a chain) or from Start writing (Use chain, Items, Start the chain);
• follow a run's steps, scores, issues and cost on its run page;
• decide at the checkpoint: Approve and continue, Ask for changes (with a note) or Stop the run;
• explain that a chain's result is one draft, never a published page.

Trainer's notes #

How to present it #

Sources: the Chains page and the chains documentation under EI: Help (extension/expai/doc/chains.md).

1. The idea (3 minutes). Read the Chains page's text: A chain writes in steps: a model writes, another reviews and the text is revised until it passes, a third prepares the title and summary, then an editor approves before anything else happens. Every step is a queued job with its cost; the result is a draft.

2. The shipped article chain (4 minutes). Draw it:

   write -> review --fail--> revise -> review ... (at most 3 reviews)
                |pass
                v
          SEO and QA -> checkpoint (editor) -> translate (hand-off) -> done

   The draft is written at the checkpoint, below the chosen location, as the user who started the run.

3. Starting (3 minutes). In the Intelligence tab, the section "Write with a chain": choose the chain, the number of items, the brief, Start the chain. On the dashboard's Start writing: Use chain, Items, Start the chain. Chains need the job queue; steps run in the background.

4. The checkpoint (5 minutes). The run page shows state, cost, tokens, time, the draft, every step with model, verdict, score, issues and checklist, the metadata and the link check. Three buttons: Approve and continue (the draft is read back, so your edits in the edit form are kept), Ask for changes (a note is required; the run revises and reviews again and returns), Stop the run (queued jobs are canceled; the draft stays a draft).

5. Escalation (2 minutes). After three failed reviews the run goes on and the checkpoint shows the reasons.

Common questions #

Is a chain better than a single request? For longer texts with a quality bar, the review step catches problems; it costs more requests.

Who may see my run? You, holders of expai/log, and users of the chain who may edit the run's item.

What does the translate step do with a new draft? It is skipped for an unpublished draft; translations follow after publishing through translation sync.

Pitfalls #

• No job queue. Without a running queue the run never starts its steps. Check before the class.
• Approving unread. The checkpoint exists for reading; open the draft first.
• Many items. Each item is one run with all its steps; start with one.

Hands-on exercise #

Use the training installation, a role with expai/chain, and your course folder.

1. Open your course folder's Intelligence tab and find "Write with a chain" (or use Start writing on the dashboard page with Use chain).
2. Choose the article chain, Items: 1, and the brief: An article for beginners on composting in a small garden, about 200 words, practical steps, no brand names.
3. Press Start the chain. The run page opens.
4. Follow the steps as they finish (refresh or wait). Write down the review score and any issues.
5. When the run waits at the checkpoint, open the draft from the run page and read it.
6. Press Ask for changes with the note Add one sentence on what not to compost. Wait until the run is back at the checkpoint and read the draft again.
7. Press Approve and continue.
8. Open EI: Chains and find your run in the list; read its final state and cost.
9. Publish the draft from My drafts, or discard it.

Expected result #

• One chain run per learner reached the checkpoint, went through one round of changes, and was approved.
• The draft below the learner's folder contains the requested sentence; it is published or discarded by the learner.
• The trainer checks by opening one learner's run page on the projector.

Quiz #

1. Name the steps of the shipped article chain in order.
2. When is the chain's draft written?
3. What does Ask for changes require? A. Nothing B. A note C. A new brief D. A new chain
4. What happens to the draft when you press Stop the run?
5. What does a chain need on the server to run its steps?

Further reading #

• The chains and job queue pages under EI: Help; expqueue.

Related pages #

• Module 5: Chains and the log
• Exponential Intelligence: course overview
• Trainers: start here
• Free documentation: this page is free to copy, translate and improve under the GNU FDL 1.3.