Best practices

Refer to the best practices below to effectively use aiR Assist.

Prompting

Following these prompting best practices may help ensure accurate, efficient, and relevant responses from aiR Assist.

Ask clear and focused questions
To help improve results, keep questions concise and specific. Focus each query on a single topic or piece of information so it can be mapped to documents. Long, compound, or highly complex questions may take longer to process, may not map to documents, and can reduce clarity in the generated response.
Leverage keywords and synonyms
Retrieval engines like specific terms. Include likely variations of keywords (such as, bribe, gift, incentive) or entity synonyms (such as, bt or bt.us for Big Thorium). Retrieval benefits from alternative phrasing.
Stay within the saved search context
aiR Assist generates responses based on the documents included in the public saved searches used to build its indexes. Each saved search defines the specific dataset aiR Assist can draw from within the workspace. Questions should therefore relate to the content of those saved searches rather than general or external topics.
Expect some variation in repeated queries
Submitting the same or similar questions multiple times may produce slightly different answers, as aiR Assist regenerates responses dynamically. However, the core content and conclusions are generally expected to remain similar, though variation may occur.
Review citations and supporting references
Each aiR Assist response includes citations and supporting document references. Review these sources to verify accuracy and context, especially when using the results for analysis, reporting, or decision-making.
Maintain high-quality indexed data
Response quality depends on the content indexed. Ensure that the dataset includes clean, text-extractable documents and that saved searches accurately capture relevant materials. Avoid including duplicate or irrelevant documents within indexes.
Avoid overly broad or “find everything” queries
aiR Assist is optimized to find and synthesize the most relevant information, not to return exhaustive lists of all matching documents. For comprehensive discovery, you can use standard search tools in combination with aiR Assist or, depending on the use case, consider using our other aiR Suite products.
Avoid using aiR Assist for calculations
Don’t rely on aiR Assist to add up invoices, total damages, or perform complex calculations. The right data may be scattered across multiple files, and LLMs are not perfectly reliable at maths.
Break down complex reasoning or multi-step queries
Avoid questions that require multiple steps or “leaps” (such as, “show me emails from the director who signed the compliance policy”). Break these into smaller, sequential questions so aiR Assist retrieves the most relevant information and clearly understands your objective.

Examples of unsupported prompts and their supported alternatives:

Unsupported prompt wording example	Why it is unsupported	Supported alternative wording example
Find me all examples of [issue].	aiR Assist is optimized to identify and synthesize the most relevant information, rather than provide exhaustive lists of every matching document. Because it uses a RAG architecture, it retrieves a relevant subset of the corpus and therefore cannot guarantee complete coverage. See Understanding aiR Assist responses for more information.	What evidence is there of [issue]? Ask for the evidence or key examples, rather than an exhaustive list.
Find me all conversations by [actor].	aiR Assist is optimized to identify and synthesize the most relevant information, rather than provide exhaustive lists of every matching document. Because it uses a RAG architecture, it retrieves a relevant subset of the corpus and therefore cannot guarantee complete coverage. See Understanding aiR Assist responses for more information.	Did [actor] discuss [topic]? Focus on a particular topic or issue, instead of asking for all conversations.
Find me all conversations that took place between [actor] and [actor].	aiR Assist is optimized to identify and synthesize the most relevant information, rather than provide exhaustive lists of every matching document. Because it uses a RAG architecture, it retrieves a relevant subset of the corpus and therefore cannot guarantee complete coverage. See Understanding aiR Assist responses for more information.	What topics did [actor] discuss with [actor]? Focus on the substance or topic of the email exchanges, instead of asking for all conversations.
Find me emails between March 1-March 31, 2012.	aiR Assist is optimized to identify and synthesize the most relevant information, rather than provide exhaustive lists of every matching document. Because it uses a RAG architecture, it retrieves a relevant subset of the corpus and therefore cannot guarantee complete coverage. See Understanding aiR Assist responses for more information.	What was discussed about [topic] in March 2012? Tie the date to a topic so it becomes a document corpus related question.
Find me emails sent after-hours.	aiR Assist is optimized to identify and synthesize the most relevant information, rather than provide exhaustive lists of every matching document. Because it uses a RAG architecture, it retrieves a relevant subset of the corpus and therefore cannot guarantee complete coverage. See Understanding aiR Assist responses for more information.	Did [actor] discuss [topic] outside business hours? Anchor it to a topic or actor, instead of an exhaustive list.
Find me the earliest email referencing [term].	aiR Assist retrieves a set of documents and answers questions based on the most relevant results it finds, but that set may not include the earliest or latest email.	What evidence is there of [term] being discussed? Ask about the term to pinpoint the earliest instance.
Write me a prompt for aiR for Review related to this matter.	This prompt is not asking a question about specific indexed documents. It is asking aiR Assist to generate content for a task other than document search.	What are the key issues and actors in this matter? Use aiR Assist to summarize the facts, then craft the prompt yourself from that summary.
What is [document name] about?	aiR Assist cannot reference or look up documents by name, hyperlink or document ID. It only searches extracted text and metadata (emails and date), so a query related to a specific document name does not map to how it retrieves.	What does the corpus say about [subject/keyword from that document]? Query the content or key terms rather than the filename or ID.
What are the discrepancies between [doc 1] and [doc 2].	aiR Assist is optimized to identify and synthesize the most relevant information, rather than provide exhaustive lists of every matching document. Because it uses a RAG architecture, it retrieves a relevant subset of the corpus and therefore cannot guarantee complete coverage. See Understanding aiR Assist responses for more information.	What did [actor] say about [topic]? Break the main question into smaller, sequential questions that can be compared.

Indexing

To achieve the best indexing results, below are some data preparation recommendations:

Verify that documents contain extracted text to ensure they can be properly indexed and used by aiR Assist.
Verify that public saved searches are well-defined and include only relevant materials.
Exclude excessively large documents that may be skipped during indexing, in addition to duplicate and known irrelevant documents.
Ensure that the Extracted Text field is Data Grid File System enabled, as aiR Assist does not support environments where the text is stored in SQL.
ARM (Archive, Restore, and Move) is not supported for aiR Assist indexes.