Custom analyses

Custom analyses provide a flexible framework for creating and executing custom AI‑powered insights per project within Relativity. Each insight acts as an independent instruction, enabling customizable results across all your legal data intelligence use cases.

Instead of relying solely on the predefined analysis types in aiR for Review (such as Relevance and Key Documents), you can create up to five Custom analysis insights per project, each acting as independent analyses for questions, extractions, classifications, or evaluative instructions applied to every eligible document. Reviewers receive per‑document outputs, dependent on their unique Custom analyses prompt.

Custom analyses support both Text and Vision analysis models:

Custom Analysis Model	Description
Text	Analyzes the extracted text of each document. Best for emails, contracts, reports, and most standard documents. Usually the starting point for most review matters.
Vision	Analyzes the document as a native image file; not extracted text. Best for scanned materials, photos, screen shots, and handwritten documents, for example. Can read text, identify objects, locate handwriting, and detect logos on this image, for example. Surfaces information that does not exist in extracted text.

Each Custom insight analysis produces one text‑based output, which may include:

• Summaries
• Yes/No responses
• Classification labels
• Short explanations
• Identification

Each Custom insight consists of a unique title and prompt instructions. The prompt instructions tell aiR for Review how to analyze the documents and how to present the results. Examples include summarizing information, extracting items, identifying attributes, or applying classifications. These instructions form the core logic for what the aiR system model returns for each document in the Analysis Results. Each Custom insight will be listed in its own column in the Analysis Results grid and document viewer. For guidance on writing effective custom insight prompts, see Best practices for writing Custom insight prompts.

Considerations

Review the following considerations for using Custom analyses in aiR for Review:

• Unlike other analysis types in aiR for Review, Custom analysis results are not available on the Relativity Document Object until you apply them. They are initially only available in your aiR for Review project and the Viewer. To publish them to the Relativity Document Object fields, run the Apply step (see Applying the results and publishing to the Document Object).
• After the Apply process is complete, the Custom analysis results are available for export. See Exporting Custom analyses results for more information.
• Validating prompt criteria in Review Center is not supported for Custom analyses. See Prompt Criteria validation in Review Center for more information on validation.
• Insight Suggestions have no effect on how your analysis runs or its cost. They are a starting point you can use, edit, or ignore. See Using Insight Suggestions for more information.

Custom analyses workflow

The workflow for Custom analyses is similar to the other analysis types:

1. Setting up a project
2. Adding Custom insights and prompt criteria
3. Running an analysis job
4. Viewing the analysis results
5. Applying the results and publishing to the Document Object

Setting up a project

Begin by setting up the aiR for Review project as described below. For additional information on project set up, see Setting up the project.

1. Navigate to the aiR for Review Project tab.
2. Click New aiR for Review Project.
3. Fill out the following fields on the Setup Project modal:
   
   • Project Name—enter a name for the project.
   • Description—enter a useful short description of the project.
   • Data source—select the saved search that holds your document sample set. Refer to Choosing the data source for more information.
   • Analysis Type—select Custom.
   • Project Prompt Criteria—defaults to Start blank to write new prompt criteria from scratch.
   • Prompt Criteria Name—skip this field for Custom Analysis Type.
   • Custom Analysis Model—choose the analysis model desired:
     • Text—run analysis on the extracted text of documents.
     • Vision—run analysis on native image file types of JPEG, GIF, or PNG.
   • Project Use Case—choose the option that best describes the purpose of the project. If none of the options describe the project, choose Other and type your own description in the field. It will only be used for this project. Keep this description generic and do not include any confidential or personal information. The Project Use Case field is used for reporting and management purposes and does not affect how the project runs.
4. Click Create Project.

Adding Custom insights and prompt criteria

On the Custom panel, you can create and edit Custom insights and prompts. You can also delete ones that are no longer needed.

1. Add up to five Custom insights using the methods below:
• Manually enter Custom insights and prompts: Enter a unique Title (up to 30 characters) for the insight. The title displays in the Analysis Results table, Custom Projects card in the Viewer, and the published output. Next, manually enter clear and descriptive instructions in the prompt text box (up to 1000 characters) that will guide aiR in analyzing documents. See Best practices for writing Custom insight prompts for more information.
• Use preset Insight Suggestions: Click the +Add button next to a Suggested Insight to automatically add it. Afterward, if needed, edit the suggestion’s prompt to fit your matter. Editing an Insight Suggestion is the same as editing an insight you manually created. See Using Insight Suggestions for more details.
  
2. To add more Custom Insights, click + Insight to add one manually, or click + Add next to a suggested insight. Repeat the previous step for each additional insight.
3. To delete a Custom insight from the list, click the vertical ellipses next to the desired title and select Delete.
4. Click Save when finished adding Custom insights.

Custom insight information can be edited or deleted prior to running the analysis.

Using Insight Suggestions

Insight Suggestions are preset prompt templates that appear at the top of the prompt configuration panel. They help you get started quickly by providing automatic insights and related instructions that you can add to the panel.

The suggestions displayed correspond the analysis model you selected when setting up the project (Text or Vision):

• Text projects show text-based suggestions. For example, summarizing a document or extracting key facts, people, dates, and amounts.
• Vision projects show vision-based suggestions. For example, summarizing an image, identifying key objects, or extracting and locating visible text.

These Insight Suggestions for Text analysis and Vision (images) analysis cover some common requested needs.

Running an analysis job

After setting up all the custom insights, use the steps below to run the analysis job. For additional information, see Running the analysis.

1. Click Analyze [X] documents.
2. Review the confirmation summary modal showing the total number of documents to be analyzed.
3. Click Start Analysis.

Custom insight information cannot be edited or deleted after running the analysis.

Each Custom insight runs independently across eligible documents that meet the following criteria:

• Match the document requirements as stated in Job capacity, size limitations, and speed.
• Do not have prior results with the same prompt.

Documents that cannot be analyzed will display with errors. Refer to Handling document errors in Custom analyses for more information.

Results are written to Fixed Length Text fields and will have a maximum of 3,000 characters per output.

See How are aiR Units handled when reapplying and republishing the same prompt criteria for additional information surrounding aiR units.

Viewing the analysis results

Once the analyses complete, each custom insight displays as individual columns in the Analysis Results table and in a dedicated Custom Projects card within the Viewer. For additional information, see Viewing and filtering analysis results. Iterate on the results as needed by refining instructions, rerunning samples, and reviewing output until you are confident to run it on larger document sets.

Custom insight column results populate as they become available rather than the order in which they were added to the Custom panel, unlike other aiR for Review analysis types. This allows results to appear live as each field completes, without waiting for all fields to finish.

To open the document in the Viewer, click on the Control Number link. See Viewer documentation for more information on using the Viewer.

If an error displays for an insight, see Handling document errors in Custom analyses for guidance.

If you modify Custom insights, change the data source, or create a new project set, only the updated insights will re-run on the next analyses.

For more information on analyzing and filtering results, see Analyzing aiR for Review results.

To export Custom analyses results, you must first apply the results, which saves the data to the Relativity Document Object. See Applying the results and publishing to the Document Object and Exporting Custom analyses results for details on each process.

Applying the results and publishing to the Document Object

After confirming that the results meet your expectations, apply the prompt criteria to the document population. This also publishes the results to the Relativity Document Object, where they can be used across the review workflow for filtering, sorting, and display alongside other document fields in the Document List.

Since the iteration process occurs inside the project, you may rename, add, delete, or change insight fields numerous times as you are iterating on the output. The Apply process is separate because applying the final results on the sample or full data set ensures the final insight fields are created correctly in the workspace Document Object.

Running Apply

To apply prompt criteria for a project set:

1. Within the desired project set, click the project set + sign. For more information on project sets, see Using project sets.
2. Select I want to apply my prompt criteria to a document population.
3. Click Create Apply Set.
4. Click Analyze [#] documents.
5. Review the confirmation summary, including the estimated number of documents and insights to be processed, and add notification email addresses, if needed.
6. Click Start Analysis.

To export Custom analyses results after applying the prompt criteria, see Exporting Custom analyses results for details.

How the Apply process works

When you run Apply on a project set, each insight result is written to the Document Object automatically. For every insight in your project, a field is created and populated using the naming convention aiRCustom::[Insight Name] ([Project Name]).

The Apply Set will generate new results for documents that meet the following criteria:

• Match the document requirements as stated in Job capacity, size limitations, and speed.
• Do not have prior results with the same prompt.

Existing results on the latest prompt version will not be re-written or re-charged.

After publishing, the results are immediately filterable like any other document field.

• One record per document: All of a document’s Custom analyses results are written to a single shared object on that document.
• Free-form results: Each insight publishes its own text output under its own field, reflecting the prompt you wrote.
• Errored documents are skipped: A document that errored on a given insight is not written for that field.

Re-applying a project

You can re-run and re-apply a project as your prompts evolve.

• Updates in place: Re-applying a project updates that project’s results on each document rather than creating duplicates.
• Other projects are preserved: Re-applying one Custom analyses project does not affect the published results of any other project on the same document.

How are aiR Units handled when reapplying and republishing the same prompt criteria

When reapplying and republishing results for the same prompt criteria, no aiR Units are consumed as long as: (1) analysis results already exist for the documents, and (2) the prompt criteria have not changed at all since the initial analysis run. If the exact same prompt is reused across versions or even across different projects in the same workspace, the system recognizes it and displays the prior analysis result without consuming additional aiR Units. However, be aware that the prompt criteria must be an exact match, even an accidental extra space would count as a different prompt and, therefore, would consume aiR Units. For more information on aiR Units, see aiR Units in RelativityOne standard workspaces and Community article aiR Units, Thresholds, and Included Products (valid login credentials required).

Exporting Custom analyses results

Exporting Custom Analyses results directly from the Analysis Results tab is not currently supported. To export Custom Analyses results, you must publish your results to Relativity Document Object using the Apply workflow (see Applying the results and publishing to the Document Object). Following publishing, you may export your results from the Document List.

Handling document errors in Custom analyses

Custom analysis errors are reported at the insight level for each document, rather than for the document as a whole. Because each insight runs independently, one insight can succeed on a document while another fails. Each insight column shows its own success or error status.

Non-responses

Sometimes, the aiR System model analyzes a document but intentionally does not return an answer. This is expected behavior, not an error. It helps ensure high-quality results when the model cannot provide a conclusive response.

A non-response can occur when:

• a document is blank
• the text or image is too faint or illegible
• the prompt instructions do not apply
• the model cannot find relevant content

In these cases, the non-response message may appear in one of the following ways:

• No relevant content found.
• This document appears to be empty and could not be analyzed.
• This document could not be read or interpreted.
• A blank (null) cell appears.

Error reference

The following table provides some of the errors you may encounter, along with information on how to resolve them:

Error	What it means	Likely resolution
Failed to parse completion	The aiR system model produced output that couldn’t be read in the expected format.	Retry the document. If it persists, simplify or clarify the insight’s prompt, or contact Relativity Support.
No content to analyze	The document had no extractable text for a text analysis.	Confirm the document has extracted text. For image-based documents, use a Vision project instead.
Content exceeds limits	The document’s content is too large to analyze in a single pass.	Reduce the size of the document set or split large documents. See Job capacity, size limitations, and speed for additional information.
File provides insufficient context	The document has too little content for the aiR system model to act on.	Verify it is the intended document. Near-empty files may return a limited result or none.
File is not supported	The file type provided as the data source is not supported.	Convert the file to an acceptable file format. Make sure image files are native images (JPEG, GIF, PNG) for Vision analysis. See Job capacity, size limitations, and speed for additional information.
Uncategorized error occurred	An unexpected error occurred.	Retry the document. Also, make sure image files are native images for Vision analysis. If it persists, contact Relativity Support.

Common use cases

Custom analyses is most effective for per-document tasks that you can express as a clear instruction. Customers have explored use cases such as:

• Summarizing documents and surfacing key facts, people, and dates
• Extracting and structuring entities — names, dates, and monetary amounts
• Classifying or categorizing documents
• Describing and summarizing non-English text
• Vision model tasks such as, describing images, identifying objects, and extracting visible or handwritten text

Best practices

To get the best results:

• Be clear and specific—clear, well-structured prompts consistently outperform vague ones. State exactly what to return and in what form.
• Define the output format—if you want results returned in a specific structure (for example, a label, a yes/no, or a short list), define it. Vague asks tend to return more than you want.
• Don’t over-constrain—piling on many conditions and caveats can sometimes reduce quality. If results degrade, simplify and iterate.
• Use one insight per question—each insight runs independently; keeping each to a single, focused task improves consistency.
• Provide good inputs for Vision analysis—higher-resolution, well-oriented images produce markedly better and more reliable results.

For additional guidance, see Best practices for writing Custom insight prompts.

Known limitations

Custom Analysis is powerful but, like any AI feature, has limits. Knowing these helps you set expectations and review results appropriately.

• Output format can vary—the aiR system model may not follow the requested structure or conditional logic perfectly across every document (for example, occasionally varying formatting or dropping a requested element). Defining the output format clearly reduces this.
• Image quality drives reliability—low-resolution, very small, or partially obscured images are the most common source of mis-identification in Vision analysis. Text in unusual orientations or scripts may also be missed.
• Handwriting and dense extraction—handwriting transcription may occasionally complete or guess characters. Output is limited to about approximately one page of text. Review handwritten and hard-to-read extractions.
• Run-to-run variation—results can vary slightly between job runs on the same document, particularly for subjective or estimative tasks. Perform quality assurance accordingly.
• Multiple values in one document—when a document contains several candidates (for example, multiple dates), the aiR system model may pick one without flagging the ambiguity.
• Mental math—the aiR system model is not a calculator and may produce unreliable results for math-related tasks, such as adding figures or converting units.
• Configuration limits—maximum of five insights per project, prompt and output lengths are capped, and very large documents may exceed processing limits.