Frequently asked questions (Advance Access)

aiR Assist can handle questions related to the documents in the current workspace. Questions that answer general knowledge questions are not supported. You can ask questions in natural language, such as "Who are the key people in this case?" or "Which documents mention [topic]?"

There is no set limit, but long or highly complex queries may take more time to process. For best results, keep your questions concise and focused.

aiR Assist can return up to 25 references for one question, giving you a broad and well-supported set of sources.

Yes. aiR Assist uses advanced intention capabilities to reject questions that cannot be answered or are outside its scope. It also requests clarification for vague or overly broad queries.

aiR Assist uses a Retrieval-Augmented Generation (RAG) approach to answer questions. The system first retrieves the most relevant documents for the query and then uses a large language model (LLM) to generate a summarized response.

The following best practices can help improve accuracy when working with contract, entity, or relationship extraction tasks. For additional suggestions, see Best practices (Advance Access):

• Be explicit, not implicit
  Retrieval systems work best with clearly defined prompts. Vague questions such as “Tell me about the case” are less effective than specific ones like “Describe contract terms related to indemnification.”
• Ask concise questions
  Multi-part or compound questions can fragment retrieval results. Breaking complex questions into smaller, focused prompts improves precision and clarity.
• Leverage keywords and synonyms
  Retrieval engines benefit from varied phrasing. Include alternative terms or related entities, such as “bribe,” “gift,” or “incentive,” to capture a broader range of relevant content.

When using aiR Assist, it is also important to keep the following limitations in mind:

• Lack of comprehensiveness
  Retrieval systems aim to answer the question directly rather than provide a full list of all related documents or entities.
• No metadata support
  Queries that depend on document metadata (such as, file type, custodian, or date range) may return incomplete or imprecise results, as metadata filtering is not currently supported.

aiR Assist is designed and optimized for English-language content. It has not been formally tested or validated for use with multilingual or non-English datasets. While it may process and return responses for non-English text, accuracy and completeness cannot be guaranteed, and verification of cited sources is strongly recommended.

No. At this time, you cannot share or delete questions.

No. aiR Assist does not currently include configuration options to control response variability or “thinking time.” The system prompt used by aiR Assist is designed to reduce variability, but some level of randomness is inherent to all large language models (LLM).

As a result, identical prompts may generate slightly different responses, especially when phrased implicitly or influenced by prior conversation context. The more specific and well-defined the question, the more consistent the output tends to be.

Unlike systems such as ChatGPT, which include adjustable modes (for example, auto, instant, or thinking), aiR Assist focuses on accuracy and relevance rather than adjustable response styles. aiR Assist currently uses GPT-4.1 to generate answers.

Not at this time. aiR Assist follows a Retrieval-Augmented Generation (RAG) approach, in which only the most relevant document chunks are retrieved and used to generate a response. Repeated or boilerplate text (such as, standard disclaimers, email footers, or signatures) is not explicitly filtered out. However, this type of content is typically not selected unless it is determined to be highly relevant to the user’s query.

aiR Assist relies on indexed data to generate answers leveraging retrieval-augmented generation (RAG). While it aims to offer correct responses, always verify the offered information and citations for accuracy. Using the feedback buttons in the response box can help improve the tool over time.

To leave feedback in the aiR Assist app, use the thumbs up or down icons below each answer to rate it and add comments. See Feedback controls (thumbs up/thumbs down) for more information. You may also occasionally be present with an aiR Assist Feedback dialog asking how helpful were aiR Assist's answers for completing your task. See In-product feedback survey for details.

Yes. Your conversation is private to you. At this time, each user's session and chat history remain private and separate and no one else can see it.

No. Your history persists between user sessions and remains as long as the workspace exists.

No. At this time, you cannot delete conversation history. It remains as long as the workspace exists.

Yes. Multiple users can use aiR Assist within the same workspace concurrently. However, each user's session and chat history remain private and separate.

Yes. aiR Assist maintains conversational continuity within a session, allowing it to reference the context of previous queries when formulating responses. This means that follow-up questions can build on information discussed earlier in the same conversation.

However, the continuity mechanism in aiR Assist differs from general-purpose chat systems, such as ChatGPT. Each interaction still begins with a new document retrieval step, where aiR Assist searches for the most relevant materials based on the current query and prior context. The retrieved content then informs the contextualized question, which the language model uses to generate the final response.

aiR Assist indexes do not currently have a dedicated view within RelativityOne. Indexes are managed directly through the aiR Assist interface. When switching between indexes, the system displays a status card within the chat showing the index state.

If an index was built with errors, the status card includes a “View Errors” option that lists documents that encountered issues during indexing. Documents larger than 5 MB are automatically skipped during the indexing process and are not currently displayed as errored items.

There are no known delays or edge cases that prevent aiR Assist from accessing newly indexed documents. Once the index build task is complete, all included documents become immediately available for querying. Relativity offers workflows that can identify repeated text.

aiR Assist generates its answers based solely on the extracted text from processed documents. This means that for Excel files, the content available to aiR Assist depends on how text is extracted during processing.

Depending on your processing profile settings and the engine used to extract text, certain elements (such as data from multiple tabs, formulas, or embedded objects) may or may not be included in the extracted text. To ensure the most complete and accurate results, review and optimize your processing configuration before indexing Excel files.

aiR Assist can only use the extracted text available in RelativityOne. For image-only documents (such as scanned PDFs or pictures without embedded text), text extraction must be performed in advance, either through an external process or using an OCR (Optical Character Recognition) workflow within RelativityOne.

Without extracted text, aiR Assist will not be able to index or generate answers from image-based documents.

No additional steps are required for large documents that are within the 5 MB size limit. During indexing, aiR Assist automatically divides large documents into smaller text chunks to optimize retrieval and improve response relevance. These chunks may be retrieved individually when aiR Assist processes a query, ensuring that large documents are handled efficiently without additional configuration.

No. Documents exceeding the 5 MB size limit are automatically skipped during indexing and are not currently flagged or listed in the error report. As a result, these files will not appear among errored documents in the aiR Assist interface.