Custom AI Development for Relativity Assisted Review

Relativity Assisted Review (RAR) provides a powerful foundation for Technology-Assisted Review (TAR), but its native models are generalized for broad legal language. A custom AI integration allows you to build models trained on your specific matter types—such as M&A due diligence documents, patent litigation technical files, or internal investigation communications—and plug their predictions directly into the RAR workflow. This is achieved by using the Relativity REST API and Event Handlers to create custom objects (like a CustomModelScore object), populate fields with model outputs, and then surface those scores in the Active Learning workspace or as a priority ranking in the document list. The integration acts as a parallel scoring engine, enriching RAR's native prioritization with domain-specific signals.

The implementation typically follows a batch-and-sync pattern: documents are exported from a saved search via the API, processed by your custom model (hosted on your infrastructure or a cloud service like Azure ML), and the results are written back as choice fields or decimal scores. For real-time workflows, an Event Handler can be configured to trigger model inference when a document's Extracted Text field is updated, allowing for immediate tagging. Governance is critical; all model predictions should be stored with audit trail metadata (model version, inference timestamp, confidence score) and integrated into Relativity's RBAC so only authorized reviewers can see or act on the AI-generated tags. A phased rollout starts with a pilot workspace, where model predictions are visible as non-decisive fields, allowing reviewers to provide feedback that retrains the model in a continuous loop.

This approach moves beyond a black-box AI service. It creates a governed, auditable system where custom intelligence becomes a repeatable asset. The result is not just faster review, but more consistent and accurate issue coding for niche document types, turning a general-purpose TAR tool into a specialized expert for your recurring legal challenges.

A production integration connects a custom model to RAR via Relativity's REST API and custom objects. The core pattern involves creating a dedicated Custom Object (e.g., Custom Model Prediction) that stores model scores, confidence levels, and metadata, linked to the base Document object. An external AI service, hosted on your infrastructure or a managed cloud, processes batches of documents retrieved via the API. Predictions are written back to the custom object, where they can be surfaced as RAR Fields in the viewer, used to power Saved Searches for prioritization, or drive Batch Sets for reviewer assignment. This architecture keeps the custom model's runtime outside of Relativity's core, ensuring scalability and avoiding performance impact on the review workspace.

Rollout follows a phased, matter-specific approach. Start with a pilot on a closed Batch Set of documents (e.g., 10k technical emails for an IP case). Use the RAR Continuous Active Learning (CAL) framework to compare the custom model's predictions against human reviewer coding, measuring precision/recall for your target issue (e.g., "Relevant to Patent Invalidity"). Integrate the model's top-confidence predictions as Suggestions in the review interface, allowing reviewers to accept/reject them, which feeds back into the model's training loop. Governance is critical: maintain an audit log of all API calls, document IDs processed, and predictions made, and establish a human-in-the-loop checkpoint for any model-driven decisions that affect privilege or final production sets.

This approach is valuable when native RAR classifiers or conceptual analytics lack domain-specific precision. For example, a model trained on SEC filing language can better identify "Forward-Looking Statements" in financial investigations, while a model fine-tuned on clinical trial protocols can spot "Patient Eligibility Criteria" in healthcare litigation. The result is a hybrid review system where RAR manages the workflow and sampling, and your custom model acts as a high-precision specialist, reducing the number of documents reviewers must examine manually for nuanced, case-specific issues. For related architectural patterns, see our guides on AI Integration with Relativity APIs and Scripts and AI for Predictive Coding and TAR in E-Discovery.

Governance starts with the model's training data and intended scope. For a custom model, you must define its legal domain (e.g., antitrust communications, employment policy violations, specific contract clauses) and document the curated training set used, ensuring it's representative and free of privileged material. In Relativity, this model should be deployed as a supplement to RAR, not a replacement, with its predictions surfaced as custom fields or objects. Access controls via Relativity's native security groups are critical to restrict who can train, run, or view the model's outputs, while detailed audit logs track every prediction batch and user interaction for defensibility.

A phased rollout mitigates risk and builds confidence. A typical implementation follows this pattern:

• Phase 1: Silent Validation. The model runs in the background on a closed matter, comparing its predictions to a human-reviewed control set. Metrics like precision, recall, and reviewer agreement are logged without influencing the active workflow.
• Phase 2: Assisted Prioritization. The model's outputs are used to pre-populate a review queue or suggest coding, but all decisions require explicit reviewer confirmation. This phase focuses on UX, measuring time savings and user feedback.
• Phase 3: Conditional Automation. For high-confidence predictions on well-defined issue types, the model can apply tags automatically, but with a defined exception workflow routing low-confidence items for human review. This phase often integrates with Relativity workspace templates and event handlers to automate the setup for new matters.

Security is paramount. The model itself, whether hosted on-premises or via a secure cloud API, must be accessed through a service account with minimal permissions. All data in transit between Relativity and the inference endpoint must be encrypted. For highly sensitive matters, an air-gapped deployment may be necessary. Finally, continuous monitoring for model drift is essential; as case types evolve, the model's performance should be periodically re-validated against new gold-standard sets to ensure its predictions remain reliable and defensible.