Find the right evidence, then read it precisely

A purchase number, a customer name, an unusual phrase or a product code gives a strong starting point. Keyword search finds those written clues in the text being searched. Try a distinctive term first, then inspect the result's account, sender and subject to establish whether it belongs to the right conversation.

Choose the text layer that contains the clue. Ordinary message search and search over existing extracted body or attachment text serve different purposes. The extracted layer can reveal wording inside a supported document, while its coverage explains which parts were available. A snippet locates a possible match; opening the relevant passage supplies the surrounding meaning.

Sometimes you remember the idea rather than the phrase. One supplier writes about a delivery delay; another calls it a revised dispatch window. Semantic similarity uses embeddings, numerical representations of meaning, to bring related material together. A relevant message can become a starting point for discovering other mail about the same kind of issue.

Similar meaning is a way to select candidates for reading. It can help explore a theme or find comparable exchanges, while exact terms remain useful for identifiers and quoted language. Organization-managed account embeddings provide a shared preparation resource under current source access. Their existence does not make every user's classification judgment the same.

Filters express the parts of a question that should be exact. Narrow by authorized account, sender, recipient, subject, date range, mailbox or label, attachment presence, or relevant processing state. A search for a contract number within the supplier's mail is more focused than the same term across every account.

Interpret the selected field as well as the words. Mail dates refer to the message's date, while an ingestion observation describes collection. A topic filter selects related working context without granting access to its members. For research spanning uncertain dates, begin with a broader scope and narrow after inspecting candidates so useful undated evidence stays in view.

Paging lets the assistant survey a collection in manageable groups. Continue using the supplied continuation information until the intended traversal is finished, keeping the same scope. Some searches inspect a bounded batch and may return no matches while still offering another page. That next page is useful work to continue.

Then spend reading effort where it matters. Review candidate metadata, open a prepared summary for orientation, and request the body passage or attachment needed to resolve the question. Keep synchronization and extraction coverage with the results. This approach scales the research by narrowing each step, rather than assuming the assistant needs the whole archive in one conversation.

A summary helps answer, 'What is this about, and should I open it?' An original body helps answer, 'What exactly was said?' An attachment may answer a third question that the body merely points toward. Move between these forms according to the decision: a broad triage can begin with summaries, while a contractual commitment calls for the relevant original.

Reading an existing summary uses a prepared result and does not by itself start a new analysis. Summary reading, body reading and attachment inspection remain distinct evidence. A partial body read can be useful without representing full coverage. Ask the assistant to state which form it used when that difference affects confidence in the conclusion.

An assistant's reading record belongs to its issued identity. It is separate from the provider's seen or unread flag, and from another assistant's progress. Reading evidence distinguishes summary and body coverage; work status describes what the assistant is doing with that information. The result is a useful separation between 'has read' and 'has finished the task.'

Retrieving content and acknowledging presented coverage are separate steps. A list or search result does not establish body reading. The same verified assistant can reuse qualifying evidence for the exact unchanged canonical body when selecting its initial unread work, while its notes and work remain scoped to the relevant context. Another assistant still needs its own evidence.

A useful citation identifies the source occurrence and the representation supporting a claim. Include enough context for an authorized reader to locate it: the relevant account or source, message identity, and the passage or file used. Distinguish a direct statement in an original from a summary's interpretation or an external report quoting another source.

Build the conclusion from the evidence actually retrieved. If one message states a price and another supplies the date, retain both references. If attachment coverage is incomplete, describe the conclusion's scope in the same place. Source references make review efficient: the next reader can open the evidence behind a statement without repeating the whole search.

Initial unread work asks whether the assistant has qualifying reading evidence. Unseen comments ask whether discussion has been acknowledged. Reconsideration asks the assistant to review evidence again for an explicit reason. These are useful separate views: a completed item can still receive a meaningful request to recheck a changed assumption.

A reconsideration request carries a reason and the required evidence. The receiving assistant reads the applicable material and decides how to resume; another participant does not overwrite its work state. Follow-up can be driven by configured event subscriptions, webhooks and authorized client behavior. A comment supplies context, while the connected workflow determines when an assistant continues.