Best Practices

Refer to the tips and recommendations below to effectively use aiR Assist.

Tips on writing questions

Below are some suggestions for writing effective questions for aiR Assist:

• Ask questions in natural language, such as "Who are the key individuals in this case?" or "Which documents reference [topic]?" Avoid legal jargon and extraneous words.
• Be explicit, not implicit. Keep questions concise, focused, and explicit to the documents within the current workspace. For example, specify names/surnames of who you are investigating, specify what you are investigating, and specify exact topics you are interested in investigating. It does not support general knowledge questions.
• Ask smaller questions and not multi-part, long, or highly complex queries. This prevents aiR Assist from retrieving fragmented documents. You can build on earlier questions by asking follow-up questions. aiR Assist follows the context of the conversation and the people or events discussed.
• Use helpful terminology, keywords and their variations, and synonyms that apply to what you are investigating. For example, bribe, gift, and incentive can be variations on the same concept. Entities can be referred to in different ways so include them. For example, Big Thorium may also be called Big Thor, BT, or bt.us in documents.
• Rephrase questions until clarity is achieved. If aiR Assist does not understand your question, it will ask you to clarify it. Sometimes, it will also let you know when it cannot answer your question.
• Submitting the same question repeatedly can result in answers that vary, but they tend to be similar overall.
• Keep a handy list of preferred or often used questions or prompts for future reference.

The following are examples of question types that tend to be less effective:

• Exhaustive queries—questions that ask for "all" of something, such as all documents or all events. aiR Assist retrieves the best documents that matches the question, but may not retrieve all.
• Calculations—questions that ask to perform complicated math or to total such things as invoices or damages. aiR Assist does not exhaustively retrieve all the documents that should be calculated or make sure the calculation is correct.
• Reasoning leaps—questions that require stitching together facts spread across documents. aiR Assist will not connect facts from multiple sources. Break up the questions into smaller ones to help narrow down the relevant documents.
• Relying on metadata—questions that ask for context based on metadata. aiR Assist operates on extracted text and not the metadata behind the document. It is best to use one of the other tools within RelativityOne for metadata type searches.

Tips on setting up saved searches for indexing

Below are some suggestions for setting up the saved searches for the indexing:

• Create targeted deduped saved searches for indexes (such as a custodial index, productions index, or potentially privileged issues index). Use the available tools in RelativityOne to fine tune and clean up the documents so that you are indexing the most relevant ones. For example, dedup documents using Structured analytics (such as, email threading, textual near duplicate identification) to remove noise documents.
• When creating an index, segment the corpus by, for example, custodian or specific topic. These segments can serve as distinct indexes that may be used for targeted queries.
• Name indexes descriptively and clearly so their purpose and content can be easily identified and understood by anyone.