Choose the task and conversation unit
Specify whether retrieval targets a message, a thread, an attachment, or a business decision supported by several messages. Subject lines alone are unreliable conversation boundaries. Preserve parent-message references and timestamps where suitable so the model can follow a discussion without confusing separate exchanges with similar titles.
Control quoted duplication
Replies often repeat earlier content and signatures. Decide whether the dataset contains original messages, rendered threads, or cleaned text with a traceable transformation history. Repeated quotations can inflate apparent evidence counts and place the same information in both training and evaluation unless the grouping policy accounts for them.
Define answerability over time
A later reply may reverse a decision or clarify an ambiguous request. State whether answers should reflect the latest message or the information available at a given date. Include examples that require the model to distinguish a proposal from an approved decision, with supporting evidence for the expected interpretation.
Review sensitive information before sampling
Business correspondence can contain personal details, credentials, confidential attachments, and third-party material. Specify exclusions and review requirements for the actual project before collecting examples. Fiund can discuss a scoped source request based on the task, message structure, documentation needs, and acceptable handling of attachments and identifying fields. A useful question might ask what a team agreed to do before a particular deadline. Review whether the supporting message records a suggestion, confirmation, or cancellation. Preserve that distinction in the expected answer, and make sure a later reply is not used when the question explicitly concerns an earlier point in the thread.
Review checklist
- Define message, thread, and attachment units
- Track quoted-text transformations
- Specify decision time and answer evidence
- Set exclusions before sample selection