Relevance ReviewCopy Anchor Link

Use Claira to make a first-pass relevance call on every document in a review set, with a reason for each call, without leaving Nuix Discover. This workflow covers the whole review: scoping the set, agreeing the prompt with the lawyer or client on a sample, running the bulk scan, checking the uncertain and disputed documents, and reporting what was done.

Claira does not go looking for the relevant documents. It asks the same question of each document on its own: is this document relevant, yes or no, and why. Every document in the set gets a call or an error note, so nothing drops out without you seeing it.

When to use thisCopy Anchor Link

• A large first-pass review. You have thousands of documents and need a consistent relevance call on each one before people start reading.
• A deadline that human review alone cannot meet. Claira makes the first pass and your reviewers spend their time on the documents that need judgment.
• Narrowing a set before deeper work. Run relevance first, then send the relevant documents to Privilege Review or an issue review.
• A second pass after people have coded. If the documents are already coded and you want to check consistency, see QC Review.

For a handful of documents, use Single Review instead. It is free.

Before you startCopy Anchor Link

• Claira is installed and the case is active on a plan with enough tokens. A text bulk scan uses 1 token per document, or 2 with Multi-Code. Single Review, the early case assessment, Case Context, Agent Mode chat and Insights use no tokens. See Plans and Billing.
• Fields are connected. You need a field for Claira's results (a Memo field with one-to-many cardinality is recommended), a Log Field and a Prompt Log field. See Fields.
• The pleadings and a few core documents are loaded in the case. Case Context reads up to 5 documents per context.
• Attachments are indexed as their own documents. On some older cases, an index setting puts an attachment's text inside the parent email. Claira then reads the two together and you cannot tell which one drove the call. If you see this, ask your case administrator to re-index so parents and attachments are reviewed on their own.
• Someone owns the relevance decision. Name the lawyer or client contact who will approve the prompt, and book their time before you start.
• Your reviewers have their own fields. Human coding goes in separate fields from Claira's results, including a field to record whether the reviewer agrees with Claira.

The workflowCopy Anchor Link

Step 1: Scope the review setCopy Anchor Link

Decide which documents are in the review before you write anything.

1. Agree the limits with the lawyer: custodians, date range, document types.
2. Decide how duplicates are handled, for example excluding duplicates across custodians, and write down the exact criteria. The client may ask for them.
3. Add full families, so each parent and attachment in the set is reviewed.
4. Save the set in a binder or saved search in Nuix Discover and note the document count.

The count tells you how many tokens the scan will use. Scoping is also where the savings are: a date cut-off or a duplicate rule can remove a large share of the set before any tokens are spent.

Step 2: Run the early case assessmentCopy Anchor Link

The early case assessment gives you a picture of the whole case from metadata only. It never reads document content, writes nothing to Nuix Discover, and is free.

1. Open the Claira menu (the ☰ icon), hover over More at the bottom, and choose Early case assessment.
2. Click Run assessment. Reading takes about a minute for every 200,000 documents.
3. Read Volume, People, Date coverage and Suggested segments.
4. Click Download PDF and keep it with the review file.

Check your scope against it. Look for custodians you did not expect, gaps in the date range, and file types that have no text and need an Image, Audio or Video scan. Adjust the set from Step 1 if needed. See Early case assessment.

Step 3: Build the Case ContextCopy Anchor Link

The Case Context is the case brief Claira writes prompts from: the parties, the issues, the key people and dates.

1. Open the Claira menu (the ☰ icon), hover over More, and choose Case Context.
2. Click New context and choose the Litigation / eDiscovery template.
3. In Nuix Discover, select up to 5 key documents, such as the pleadings, an order and a few core emails. Click Generate from selection.
4. Claira fills in the sections and lists up to five Missing information questions. These are the things the documents could not answer. Click Answer with Agent to answer them.
5. Read every section. Record the facts in full. Leave out strategy and work product, because everyone on the case can see the context.
6. When the lawyer is satisfied, click Approve.

Only an approved context is offered to the Prompt Generator. See Setting Up Case Context.

Step 4: Write the prompt and test it on a sampleCopy Anchor Link

Write the prompt

1. Open Quickstart. In the Prompt Generator, choose your approved context under Case context, describe what you are looking for the way you would brief a colleague, and click Generate Prompt.
2. Or start from a relevance template (Assess relevance, Relevance with confidence or Relevance and issues) and click Use with context.
3. Read the prompt. It should say what makes a document relevant, who the key people are, and what to leave out.

The prompt has to stand on its own. The Case Context is not sent with each scan, so every fact Claira needs must be in the prompt text.

Ask for an answer you can search and sort on. For example:

YESREL, NOTREL and the 1 to 5 confidence score come from the Relevance with confidence template. The counsel flag line is an addition: use your own wording.

If you want the call, the confidence score and the counsel flag in separate fields, set the prompt up in Multi-Code. See Objective Coding for how Multi-Code works.

Test it

1. Open 10 to 25 documents one at a time in the Document Viewer and click Scan Document. Include easy documents, hard ones and some with very little text. This is free. Fix the obvious problems first.
2. Pull a random sample from the review set with Create Random Sample in Nuix Discover. In a recent review the sample was 150 documents plus their families.
3. Put the sample in its own binder.
4. Have a reviewer code the sample, or use documents your team has already coded.
5. Run the prompt on the sample as a bulk scan (see Step 6).
6. Compare Claira's calls with your team's. For each disagreement, write down why.
7. Adjust the prompt and scan the same sample again. Change one thing at a time. Repeat until Claira's calls match yours.

To get help with an edit, click Revise with Agent under a single result, describe the change and click Suggest prompt edits.

Step 5: Get sign-off from the lawyer or clientCopy Anchor Link

Do not run the full set on a prompt nobody has approved. Send the person who owns the relevance decision three things:

• The prompt. In Prompt History (View history), open the prompt and click Download PDF.
• The Case Context. Click Download Case Context and choose PDF (.pdf) or Word (.docx). The file shows the APPROVED badge.
• Where to find the sample. Give the binder name, the field that holds Claira's call and reason, and the field where they record whether they agree.

Ask them to read the sample documents next to Claira's calls, mark each one agree or disagree, and add a note when they disagree. Agree the prompt changes together, make them, and rescan the sample.

When they approve, get it in writing. Then open the approved version in Prompt History and click the green check to mark it Final. Claira will tell you later if a scan ran any other version.

Leave time for this. The sample feedback should be back before your reviewers start.

Step 6: Run the bulk scanCopy Anchor Link

1. In Nuix Discover, select the review set from the browse list.
2. Open the Claira pane and choose Bulk Scanning mode.
3. Confirm the prompt is the Final version and the connected field is the right one.
4. Check Scan as. Auto, the default, picks a scan type for each document and charges each one at that type's rate. Choose Text to keep the scan at 1 token per document.
5. Click Scan List. In the Ready to start bulk scan dialog, read the document count, the estimated token usage against your balance, the estimated duration and the destination field.
6. Tick Write the prompt to each document's Prompt Log.
7. If Background Processing is set up on the case, tick Run as background processing task for a large set. Once Claira confirms the task is running in the background, you can close the window.
8. Click Start bulk scan.

Keep the browser window open during a foreground scan. If a scan stops early it is marked Incomplete. Click Resume Task to scan the remaining documents. Documents that already have results are not scanned or charged again. See Bulk Scan.

Step 7: Check the resultsCopy Anchor Link

Results land in the connected field in Nuix Discover, so you can search, sort and batch on them like any other field.

Check the scan first

1. Open the task in Bulk Tasks and look at the succeeded, failed and skipped counts. Why documents failed and Why documents were skipped list each reason with the number of documents it covers.
2. Failed documents carry a note in the result field naming the error. Skipped documents usually have no readable text. Scan them again with a different Scan as setting, or send them to a person.
3. Search the Log Field for the task reference, for example SC067, to pull the whole run.

Then put people on the documents that need them

Search the result field to build these sets:

• Counsel flag. Every document Claira flagged as two-sided or needing legal judgment goes to a lawyer, with Claira's reasoning.
• Low confidence. Send the lowest confidence scores, for example 1 and 2, to a reviewer.
• A sample of the relevant calls and a sample of the not-relevant calls. Agree the sample sizes with the lawyer.
• Insufficient Information. A person reads these.

Reviewers record agree or disagree with Claira's call in their own field, and add a note whenever they disagree. They do not edit Claira's field.

Then read the disagreements together. One disagreement is a document. Ten that share a reason are a gap in the prompt.

Record the result

Write down what the checks found, so the lawyer can say in a sentence how the first pass performed and show how that was measured. Keep these figures with the review file:

• Agreement on the sign-off sample. How many sample documents the lawyer or client agreed with, out of how many they read, for the version marked Final. For example, 141 of 150 (94%).
• Relevant documents found in the not-relevant sample. How many documents Claira called not relevant that a reviewer coded relevant, out of how many were sampled. For example, 3 of 400. This is the figure that shows what the first pass may have missed.
• Agreement on the relevant sample. How many relevant calls the reviewers confirmed, out of how many were sampled.
• How each sample was chosen. Its binder, its size, and how the documents were picked, for example with Create Random Sample in Nuix Discover.

Size the samples with the lawyer, in proportion to the matter: a larger set, a contested review or a method the other side may question calls for larger samples. If you size them with a statistical sample calculator, record the confidence level and margin of error you chose. Claira does not set a pass mark: the lawyer decides whether the figures are good enough, and what to do if they are not.

Step 8: Rerun what was overturnedCopy Anchor Link

When reviewers have overturned enough calls to show a pattern, fix the prompt and rerun those documents.

1. Search for the documents coded as a disagreement and group them by the reviewer's reason.
2. Revise the prompt to deal with each reason. Keep every fact that was already in it.
3. Test the revised prompt with Scan Document on a few of the overturned documents.
4. If the change alters what counts as relevant, go back to the lawyer for approval. Mark the new version Final.
5. Run a bulk scan on the overturned documents only. Decide with the lawyer whether the change also affects other documents of the same kind, and rerun those too.
6. Check the rerun the same way as Step 7.

A rerun is charged like any other scan. Claira adds values to a field and never replaces what is already there, so a one-to-many field would hold both calls side by side. Consider writing the second pass to its own connected field to keep the two passes apart.

Step 9: ReportCopy Anchor Link

Keep a record that shows what was reviewed, with which prompt, and what people checked.

• Bulk Task History Report. Click the download icon at the top of the Bulk Tasks tab. The PDF lists every bulk task on the case with its prompt, settings, counts, tokens used and comments.
• The prompt and its history. In Prompt History, click Download PDF and choose Full version history.
• The approved Case Context and the early case assessment PDF.
• The numbers. Documents in scope, scanned, failed or skipped, relevant, not relevant, flagged for counsel, sampled, overturned and rerun, with the figures from Record the result in Step 7.
• A summary of what the relevant documents show. In Insights mode, choose A bulk task under Use AI results from, type your request and click Generate Insight. An Insight reads a bulk task of up to 25,000 documents and can be downloaded as PDF, Word or Excel. See Insights.

Record decisions as you go with Add Comment on the task. Comments appear in the report.

Tips for better resultsCopy Anchor Link

• Use your team's words. If your reviewers say "harmful", the prompt and the tags should say "harmful" too, not a synonym.
• Say what relevance is measured against. Tell Claira whether to assess each document against the litigation as a whole or against one narrow issue. The two give different answers.
• Separate people who took part from people who are only mentioned. Ask Claira to list who sent, received, attended or organized, apart from who is merely named.
• Keep human judgment in human fields. Calls such as sensitivity, for a document that helps the case but is embarrassing, are coded by people in their own field.
• Do not simplify the prompt. When you edit a generated prompt, keep the case facts. Details that are cut from the prompt are details Claira does not have.

Common mistakesCopy Anchor Link

• Skipping the sample. A prompt that reads well can still miss what the lawyer means by relevant. The sample is where you find out.
• Changing the prompt after sign-off. Any edit creates a new version. Get it approved and mark it Final before you scan.
• Starting the scan without checking the plan and token balance. The Ready to start bulk scan dialog shows the estimated usage against your balance. Read it before you start.
• Reviewing parents and attachments as one document. Check the indexing before you start, not after the results look odd.
• Mixing Claira's results with other coding. Keep fields filled by search or bulk coding apart from fields filled by Claira. When you update them, add values instead of overwriting, and apply family-level coding to the whole family.
• Treating the not-relevant pile as finished. Sample it. That is where a missed document hides.
• Changing several things between tests. You will not know which change helped.

Frequently asked questionsCopy Anchor Link

Does Claira decide what is relevant? No. Claira gives a recommendation and a reason for every document. Your team checks the results and the lawyer makes the final call.

How big should the sample be? Use 10 to 25 documents to get the prompt working. Before sign-off, use a larger random sample from the review set. A recent review used 150 documents plus their families.

What does it cost? A text bulk scan uses 1 token per document, or 2 with Multi-Code. Failed and skipped documents are refunded when the task ends, except plan tokens when your plan renews before then. See What a bulk scan refunds. Testing with Single Review is free.

Why did the same document get a different answer the second time? The wording of a reason can change between runs. The decision should not. If a relevance call flips on the same document, treat it as a borderline document or a sign the prompt needs tightening. See Why you might not get the same answer twice.

Does Claira look anything up online? No. Claira never browses the internet. It reads the document and the prompt.

Can the Agent run the review for me? The Agent can write the prompt and start a bulk scan once you approve it. It cannot run the early case assessment or run case searches for you, and the sample test and the sign-off stay with your team. See Agent Mode.

Need help? Contact support@claira.to. We can build and test a relevance prompt with you on a live case.