Building a prism

You build a prism in the app. Its page has five tabs: Results, Questions, Try, Test and Integrate. This page covers Questions and Integrate, plus the actions in the prism's header. Prisms, fields and questions has the rules behind each part.

Only admins build prisms. An operator sees Results and Try, and runs the published version.

Create a prism

On the Prisms page, choose New prism, type a name such as Tenant emails, and choose Create. The prism opens on its Questions tab with one text field named Text and one empty question. Nothing is saved until you choose Create prism at the bottom of the tab, which saves version 1 and publishes it.

To start from a ready-made prism instead, choose Browse library. See The library.

Import from JSON

Import questions, at the bottom of the Questions tab, opens a page where you paste or drop a prism as JSON. You don't need to write the JSON yourself.

1. Choose Copy prompt under "Don't have JSON yet?".
2. Paste the prompt into any chat assistant. Add a few example texts and say what you want to know about them. + Add example texts puts your examples into the prompt for you, one per paragraph, after you choose Put examples in the prompt.
3. Paste the assistant's reply into the box, or drop a .json file on it.
4. Choose Import and save.

On a new prism, importing creates it from the JSON's fields, context and questions. On a saved prism, the prompt includes the current prism and asks the assistant to improve it, and importing replaces the fields, context and questions and saves a new version. If the JSON breaks a rule, the page lists each problem with where it is, and nothing is saved.

Export JSON, next to Import questions, downloads the version you are looking at in the same format.

Fields and context

The "What you send" card holds the fields: the pieces of text your tools send with each call, such as Subject and Body.

• Type a field's name. The app fills in its key, such as subject, which your developer uses. You can edit the key until the first save.
• Set a field to Text or Image. An image field takes a photo or screenshot. A prism has at most four.
• + Add field adds one, up to eight. Remove takes one away; a prism keeps at least one.
• After a save, a field's key and kind are fixed. For another key, add a new field and remove the old one.

Context is optional background every question reads, such as "We manage 40 residential buildings in Berlin." Use it for facts about your business, its terms and abbreviations.

Questions

Each question is a card. + Add a question asks "What kind of answer?" and offers three types:

• Choice picks one option from your list, such as Billing, Tech or Sales. Other and Not enough information are always possible answers, so you don't add them.
• Yes / No gives how likely the answer is yes, from 0 to 100%. The Lens turns that into Yes, No or Unsure.
• Scale picks one of 2 to 10 ordered levels, lowest first, such as Calm, Annoyed, Angry. A new Scale starts with Low, Medium and High.

Write the question as a person would ask it. The app fills in the key from its first three words; you can edit the key until the first save. After that, a question's key and type are fixed, and Change next to the type disappears. For another type, add a new question.

Done folds a card into a summary. Remove deletes the question.

Options and levels

Each option or level has a label and an optional description. The app derives the option's key from its label and keeps the key once saved, even if you edit the label later.

The description is what the model reads to tell options apart. Write one whenever two options could be confused, and say what does not belong as well as what does. + Option or + Level adds one.

Short name and instructions

Short name is an optional label of up to 40 characters, such as Routing team. It names the question in Results, Test, Home and the browser extension once its version is published. The model never reads it.

+ Add instructions opens a box for detail the model needs, such as which field to read or what counts. The model reads the question first, then the instructions, so they need not repeat it.

Copy a question

Copy from another prism opens a dialog where you pick a prism and one of its questions. Copy question adds it to this prism until you save. If the key is taken here, the copy gets _2, _3 and so on.

Save a version

The button at the bottom of the tab reads Save as version 4 when you have changes, and Saved when you don't. Every save creates the next version, and a saved version never changes. If something is missing, the save stops and the cards with problems open with the errors shown.

If someone else saved a version while you were editing, your save is refused. Reload loads their version and drops your changes.

Leaving the prism with unsaved changes asks Leave without saving? first.

Older versions

The version picker at the top of the Questions tab lists every version, marking the latest and the published one. Pick an older one to see its fields and questions, read-only. Results made with that version keep these questions. Back to latest returns to the version you edit.

The header shows which version callers get, such as v3 live, and a saved version that is not published yet, such as v4 saved.

Publish a version

Saving does not change what API calls run. Calls that name no version run the published version, and it moves only when you publish. Check a saved version first on the Try and Test tabs, which use the latest saved version.

Choose Publish v4 in the header. The button is greyed out while you have unsaved changes, because publishing uses the last saved version. The Publish version 4? dialog shows:

• That API calls naming no version will run version 4 instead of the current one, and that saved results keep the version that answered them.
• Test's verdict, when a test set has finished runs of both versions on the same revision. It names each question that is or leans worse, with the cases that broke and were fixed, and links to the comparison in Test. When a question got worse, the dialog puts the focus on keeping the live version. Without such a pair of runs, it says so and links to Open Test.
• The Lens columns that stop coming out, when the new version lacks the questions they read.
• The browser reviews that use another version of this prism. Reviews keep their pinned version, so publishing does not change them.

Choose Publish, or Keep version 3 to cancel.

Roll back

To go back to an earlier version, pick it in the version picker and choose Publish this version. The dialog reads Roll back to version 2? and shows the same checks. Roll back changes the calls that come after it. Saved results stay as they were.

Clone a prism

Clone prism, in the ... menu at the right of the header, makes a new prism named after this one with "(copy)" added. The copy has the same fields, context, questions and Lens, starts at version 1 and opens when it is ready. It is made from the last saved version, so the app asks first if you have unsaved changes. The Prisms page has a Clone link on each row too.

Display settings

Display settings, in the same ... menu, sets how the app reads this prism's answers. It does not change what the API returns.

• Unsure under is a percentage. A Choice or Scale answer whose top probability is under it reads as Unsure in the app. Yes / No uses the Unsure band of its Lens column instead.
• Leaning edge sets how likely a change between two versions must be before Test calls it Leaning. See Testing.

Use account default follows the account's settings instead. Admins set those on the Defaults page.

Archive a prism

On the Prisms page, choose Archive on the prism's row. The prism leaves the list and stops answering API calls. Its results stay readable by their links. Archiving cannot be undone.

Integrate

The Integrate tab is what you hand your developer, together with an API key.

Example call

The tab shows a ready curl command for this prism with its field keys, and Copy copies it. Below it is what comes back, shaped by the Lens. After the prism has answered once, the example uses the latest answered result; until then it shows empty values. The list under the Lens names the prism, its field keys, its answer columns, the version calls use and the formats on offer.

Lens

The Lens decides which columns come out of every result, in the API, in downloads and in the app. Changing it never asks the model again. It reads the published version, not a saved one you haven't published.

• + Answer column... adds a column for a question and an output, such as the top option or its probability.
• Each column has a name, an output, Move up, Move down and Remove.
• A Yes / No label column has sliders. No below sets where No ends. Unsure in between adds a band, and Yes from sets where Yes starts.
• Save Lens saves. Reset to automatic goes back to one column per question.

When you move a cutoff, the tab shows how the latest answered results would read with it, before you save. The table under "Through this Lens · latest results" shows recent results as your tools receive them. The Lens explains outputs, column names and cutoffs.

Prism name in the API

Calls name the prism by its slug, such as support-routing, set from the prism's name when you create it. Renaming the prism does not change it. Change next to it sets another one: 2 to 63 lowercase letters, digits and dashes, starting with a letter or digit. Calls to the old slug stop working at once. Calls by the prism's id are not affected.

Input storage

Input storage decides what Prismlet keeps of each text this prism answers, from Try, file runs, the API and the browser extension. The model provider still receives the text to answer it.

• Full text keeps the fields and context with each result, to read, retry and export.
• Fingerprint keeps a keyed hash instead, so a repeated text can be matched but not read.
• Nothing keeps the answers only.

Switching to a stricter setting asks first, and applies to new results. A checkbox in that dialog also removes or fingerprints the text already stored, which cannot be undone. A result without its text cannot be retried. Test sets are separate and keep the text of their cases until you delete them. See Results.