Understanding Results

After Claira scans a document, the AI produces a result based on your prompt. Knowing where to find results, how to evaluate them, and what to do when they fall short is essential to getting value from AI-assisted review.

Where results appear

Claira shows results in two places:

1. The Claira response panel. Immediately after a scan, the AI's output appears in the Results area of the Claira pane. This is where you read, review, and (for single scans) edit the result before saving.
2. The connected Nuix Discover field. Once a result is saved, it is written directly into the Nuix Discover field you connected in Settings. From that point on, the result lives in your standard review workflow -- searchable, sortable, and exportable like any other field value.

For bulk scans, results flow into the connected field automatically as each document is processed. For single scans, you see the result in the Claira panel first and can edit it before it is saved.

Result quality depends on prompt quality

The single biggest factor in result quality is the prompt you write. The AI follows your instructions literally, so clear and specific prompts produce clear and specific results.

If your results are vague, inconsistent, or missing key information, the first place to look is your prompt -- not the model or the data.

Common result patterns

Depending on how you write your prompt, results typically fall into one of these patterns:

Structured classification

The AI returns a label from a defined set, such as Relevant / Not Relevant or Privileged / Not Privileged. This works best when your prompt explicitly lists the allowed values.

Free-text summaries

The AI returns a narrative summary of the document's content. Useful for case assessment, chronology building, or when reviewers need a quick overview.

Extracted data

The AI pulls specific data points from the document -- names, dates, email addresses, key phrases. This works best with prompts that specify exactly what to extract and how to format the output.

Multi-field output

Some prompts ask the AI to return several pieces of information at once (for example, a relevance determination plus a confidence score plus a short explanation). The AI can handle this when the prompt clearly defines the output structure.

How to evaluate results

The best way to evaluate results is to test your prompt against documents where you already know the correct answer.

1. Select benchmark documents. Pick 10 to 25 documents that cover the range of your dataset: clear positives, clear negatives, and edge cases. Know what the correct answer should be for each one.
2. Run your prompt and compare. Scan each benchmark document and compare the AI's output to your expected answer.
3. Look for consistency. Run the same prompt on similar documents and check whether the AI gives consistent results. If it says "Relevant" for one contract and "Not Relevant" for a nearly identical one, the prompt may need tightening.
4. Check edge cases specifically. The documents that are hardest for human reviewers are usually hardest for the AI too. Pay extra attention to ambiguous documents.

What to do when results are not right

If results are inaccurate or inconsistent, work through these steps:

1. Refine your prompt. This resolves the majority of quality issues. Be more specific about what you want, define key terms, and specify the exact output format. See the Prompting Overview for guidance.
2. Adjust one setting at a time. Change only one variable (prompt wording, output format, or scan mode), then rerun the same benchmark set so you can measure impact clearly.
3. Check the document's extracted text. If a document has no extracted text, very little text, or garbled text (common with poor-quality scans), the AI has nothing meaningful to work with. Verify that the document has usable text in Nuix Discover.
4. Simplify first, then add complexity. If a complex prompt is not working, strip it back to the most basic version, confirm that works, then add requirements one at a time.

Next steps

• Improve your prompts with the Prompting Overview and Custom Prompts guides.
• Troubleshoot scan errors at Error Codes.

Need help? Contact us at support@claira.to.