What changed between v1 and v2 of Claira's templates
Other languages
Where a Claira template's v2 wording gives a different answer from v1, shown side by side, with how to keep a review that is under way on v1 and how to check which wording a past scan ran.
What changed between v1 and v2
Claira rewrote the templates in Quickstart so that each one states a single task, says what to answer when a document cannot be read or decided, and ends with the exact form of the answer. The templates that existed before the rewrite now show v2, and v1, their earlier wording, is still listed under the number. This article shows where the two wordings give different answers, so you can decide which one a review should run.
It covers the templates whose answers changed in a way that matters to a review already under way: Email From, Email To, Email Cc and Email Bcc, Extract Dates, Identify Document Type, Extract Titles, Extract Author, Extract Recipients, Summarize Document, Translate Content, Summarized Translation, Chronology Summary, the three relevance templates, Identify Privilege Types and the three Multi-Code templates. For the two PDF Unitization templates, Comprehensive PII (surface forms) and Special category data (GDPR Article 9), read both wordings in Quickstart. Search Term Variants, Identify Key Documents and the three templates under Images, audio and video are new and have only a v1.
If a review is under way
Quickstart always opens on the newest version. A colleague who inserts a template for the next batch of a rolling review gets v2 unless they choose otherwise, and the batch is then coded under different rules from the earlier ones. To carry on with the wording your earlier scans ran:
- In Prompt Lab or Bulk Scan, click Quickstart and find the template.
- Click the small v2 beside its title, then choose v1. The card says it is an earlier wording.
- Click Insert Template, Use with context or Load into Multi-Code. Each uses the version you are reading.
Do this every time you insert the template: Quickstart does not remember the choice. You can also skip Quickstart and reuse the prompt itself. Prompt History keeps the exact wording every scan ran, and Restore on its card puts that wording back in the prompt box.
For a new review, start from v2.
What did not change
The values you search on are the same in both versions and in every language: YESREL and NOTREL, the ISS1, ISS2 issue tags, YESPRIV and NOPRIV, the date written YYYY/MM/DD, the 1900/01/01 fallback and the 1 to 5 confidence scale. Searches, Insights and the chronology timeline that read those values work on answers from either version. The Multi-Code templates keep their field names, types and order; two of them gain one field (see below).
Where the answers differ
Each table takes one document and gives what each wording asks the AI to answer. Where v1 gave no instruction, the table says so: the answer then depended on the model and varied from one document to the next.
Email From, Email To, Email Cc and Email Bcc
| The document | With v1 | With v2 |
|---|---|---|
An email thread. The most recent message is copied to ana@example.com; an earlier message quoted below it was copied to ben@example.com. Scanned with Email Cc. | Both addresses. v1 asked for every Cc address in the document. | ana@example.com. v2 reads only the most recent message. |
| An email whose most recent message has no Cc line. | No instruction. The field could stay empty or hold a sentence. | NONE |
An address written with no angle brackets, such as Cc: ana@example.com. | Left out. v1 took only addresses inside angle brackets. | ana@example.com |
Email From and Email Bcc changed in the same three ways. Email To already read only the most recent message in v1, so only the last two rows apply to it. An unreadable document also answers NONE in v2.
Extract Dates
| The document | With v1 | With v2 |
|---|---|---|
| Minutes headed "March 2024", with no day. | No instruction for a partial date. The answer could be a completed date such as 2024/03/01. | 1900/01/01 |
| A letter dated 14 February 2025. | 2025/02/14 | 2025/02/14 |
In v2, 1900/01/01 therefore groups three kinds of document: those with no date, those dated only by a month and year, and those whose text cannot be read. v2 also names the dates it never takes, a due date or a payment date, and takes the most recent message of an email thread.
Documents with nothing to read
| The template | With v1 | With v2 |
|---|---|---|
| Identify Document Type | No instruction. | Other. A photograph or a recording with no text is still Image or Multimedia. |
| Extract Titles | A short descriptive title, written by the AI. | Undetermined |
| Extract Author, Extract Recipients | Undetermined when no author or recipient could be determined. | Undetermined, for an unreadable document too. |
| Summarize Document | A note on the limitation, then a best-effort summary. | No readable text |
| Summarized Translation | No instruction. | No readable text |
| Translate Content | No instruction. N/A when the document is already in English. | No readable text. N/A is unchanged. |
Two more changes in these templates affect documents that can be read. Extract Titles, Extract Author and Extract Recipients read the most recent message of an email thread. Translate Content translates only the passages that are in another language when a document mixes languages; v1 asked for the whole document.
Chronology Summary
| The document | With v1 | With v2 |
|---|---|---|
| A memo about a delivery, with no date on the memo or for the delivery. | A line starting with a date, YYYY/MM/DD:. v1 had no way to say a document was undated, so the date could be a guess. | A line starting Undated:, with any partial date in the sentence. |
| A document with no readable text. | No instruction. | Undated: No readable text |
The Chronology Insight lists an Undated: entry with the undated events instead of placing it on a day.
Relevance and privilege templates
Assess Relevance, Relevance with Confidence, Relevance and Issues and Identify Privilege Types keep their codes. Two things change in the answer:
| With v1 | With v2 | |
|---|---|---|
| The lines under the code | A free explanation (relevance), or the privilege types and key excerpts (privilege). | Labelled lines: BASIS: with the reason for the call, then DESCRIPTION: with one or two sentences describing the document, on every document. The privilege template adds TYPES: and EXCERPT: for a privileged document. |
| A document with no readable text, or one the AI cannot decide | No instruction. | The negative code, NOTREL or NOPRIV, with the reason on the BASIS line. |
Each template's own article gives its full v2 answer form and what to do with documents coded this way.
Multi-Code templates
| The template | With v1 | With v2 |
|---|---|---|
| Relevance and issues | Four fields: Relevance, Issues, Issue summary, Confidence. | Five: the same four, then Description. |
| Privilege and types | Three fields: Privilege, Privilege types, Privilege basis. | Four: the same three, then Description. |
| Confidence, in Relevance and issues | 1 (low) to 5 (high), with no guidance on each score. | The same scale with a meaning for each end: 5 when the document's own words settle the call, 1 when it could go either way or the text cannot be read. An undecided NOTREL scores 1. |
| Objective coding | Five fields, each with the v1 prompt of its own template. | The same five fields, each with the v2 prompt, so the differences above apply. |
When you load v2 of Relevance and issues or Privilege and types, the new Description row needs a destination field. Both templates call the row Description: if you run both on the same documents, point one of the two rows at a different field.
Checking which wording a past scan ran
Claira records that a prompt or a bulk scan came from a template. It does not record which template, or whether it was v1 or v2. What it keeps is the wording itself:
- In Bulk Scan, open the task from Bulk scan task history. Prompt version names the version the task ran, and Open in Prompt History opens it.
- Compare that wording with the template's versions in Quickstart. A v2 coding prompt ends with "Answer in exactly this form:" and its labelled lines. The v1 Email Cc prompt starts "Extract all "Cc:" email addresses".
To make this easier next time, write the template's name and version in a comment on the task or in the prompt's title. See Template versions.
Results from before the rewrite
Answers written with v1 stay as they are. Claira has no search or report that picks out the earlier answers a v2 rule would have changed: a chronology entry whose date the document did not give, for example, or an address taken from an earlier message of a thread. To check one, read the answer against its document. Earlier answers of several lines can also be stored in a different form from new ones: see Understanding Results.
Related
Need help? Contact us at support@claira.to.
Was this page helpful?
Continue reading
Need more help?
Contact our support team at support@claira.to — we are here to help.