An AI assistant can only help with company work when it can reach the information that explains that work. Connecting an IMAP account solves access to the mailbox; preparing its contents solves a different problem. A long mail archive contains repeated signatures, quoted conversations, attachments, uncertain claims and many messages that deserve different treatment. Sending the whole archive into every conversation is an expensive way to rediscover the same context.
SyncPier keeps the information archive and its preparation work together. Extraction makes content usable, classification organizes it, importance judgments help choose attention, summaries reduce repeated reading, and embeddings help discover related material. Each product answers a different question. A tag is not a summary, a similarity score is not a permission, and a concise account of a message is not its original evidence.
What this chapter explains
- Keep originals and derived results linked, with explicit coverage and source identity.
- Use organization-managed embeddings while preserving each person's spam and classification judgments.
- Choose models, permitted content and budgets for each function, then retrieve only the evidence a task needs.
On this page
Each prepared layer points back to its source. An assistant chooses the level of detail its task needs.
Start with a preserved source
Preparation begins with a known mail occurrence in a known account. The original, its source identity and the representation supplied by the provider remain the reference point for later work. A readable projection can make an HTML body easier to search, but it must not silently replace the original or pretend to be an exact provider file when it is a converted artifact.
The same principle applies to SMS and external reports. An automation report may have an exact JSON original and a readable presentation. Matching text does not merge distinct sources into one message. Keeping those distinctions lets a person open the evidence behind a result, understand who supplied it, and recover from a decoding or indexing failure without treating a generated description as the sole surviving record.
Make content readable and show what was missed
Extraction converts supported content into text that search and analysis can use. HTML conversion, document extraction and image text recognition have different coverage and resource needs. A mail body can be available while an attachment is unsupported, password protected or still waiting for processing. Those are useful facts to expose because the assistant's next step depends on them.
Attachment names, image counts and extracted attachment content also mean different things. A summary that lists a spreadsheet's name has not inspected its cells. Whether attachment content enters analysis is a setting and depends on supported handling. The assistant should be able to explain its input scope instead of presenting an apparently complete answer assembled from a body alone.
Organize mail without making everyone's judgment identical
Categories, tags and directories turn a large archive into useful working groups. Discovery can suggest a vocabulary from the mail itself; assignment applies that vocabulary to particular messages. Fixed classification rules govern the applied result, while model analysis can provide another view. A suggestion should remain distinguishable from a committed organizational choice.
Spam and classification judgments belong to the user because people can reasonably treat the same message differently. A supplier announcement may be relevant to procurement and distracting to engineering. Personal classification does not grant access, change another person's work, or automatically authorize an action at the mail provider. Source folders and SyncPier organization remain concepts whose relationship is controlled explicitly.
Judge attention for a stated purpose
Importance asks whether a message deserves attention under a particular policy. It is more useful when its question and reason are clear: an upcoming response deadline, a service interruption, or a request that blocks another person. A model's judgment is a recorded assessment, not a universal fact about the mail. Its input, policy and relevant source revision matter when someone later challenges it.
Screening for suspicious instructions, phishing or spam is another preparation concern. Such signals help an assistant treat incoming text as evidence rather than authority to change its own rules. They do not promise perfect detection. Choosing what to read next should combine the recorded judgment with the task, available evidence and the person's own decisions, rather than equating every model score with an order.
Use summaries as a route into the archive
A body summary gives the assistant a compact starting point: the request, key change, unresolved question and relevant reply context. It can help decide which originals deserve closer reading before spending the conversation's context budget. Summary input must be understood, especially for quoted replies, forwarded messages and attachments. A short description cannot stand in for evidence it never received.
A mail-body summary can be reused by the same user's assistants and workspaces when each access still has current account rights. That reuse avoids repeating preparation without sharing comments, private work state or another user's analysis. Two different messages with identical bodies are not automatically interchangeable: their reply context, permitted audience or preparation policy may differ. Any broader computation reuse needs an explicit operational choice.
Find related material through meaning
An embedding is a numerical representation used to compare the meaning of texts. It can help find a relevant message when the user's words differ from the sender's vocabulary, or shortlist related material for classification and topic suggestions. It is an aid to discovery; the assistant still needs to inspect the permitted evidence and explain why it supports an answer.
For organization-managed accounts, embeddings are organization-owned preparation work, so every user need not embed the same mailbox separately. That shared infrastructure stays separate from personal spam and classification decisions. Long input is represented in pieces rather than silently omitted, and coverage remains important. Similarity never grants original access, proves that two mail occurrences are one, or turns a suggested topic into an established conversation.
Control model work and preserve its provenance
Summarization, embeddings, importance judgments and verification are separate functions. The operating configuration chooses an appropriate model and settings for each rather than making a model brand part of the information's identity. Input and output limits use provider tokens, including the space needed for instructions and reasoning where applicable. Long input must follow that function's explicit coverage policy, not disappear behind an arbitrary byte limit.
Model processing can send permitted content to the selected provider. Consent, account settings, private context and budget therefore belong in the decision before work begins. Results should retain the function, input coverage, model or executor attribution, policy revision and usage provenance needed to explain them. Measured provider usage and an agent's self-reported usage are different evidence. A failed, deferred or incomplete stage should remain visible so the next assistant can continue from facts instead of guessing that preparation succeeded.
Questions that come up in practice
Does a summary mean the assistant read the original mail?
No. Summary reading and original-body coverage are distinct. A summary can guide selection, but a claim about exact wording, an attachment or a disputed detail should be checked against the relevant authorized original content.
Can the organization share embeddings without sharing my personal judgments?
Yes. Organization-managed embedding preparation and a user's spam or classification decisions have different ownership. Sharing the retrieval infrastructure does not merge personal notes, work status or discussion audiences.
Should every message be sent to the most capable model?
The configuration should follow the function, permitted data and budget. Local extraction, fixed rules, compact summaries and selective original retrieval can reduce repeated model work. No single model choice removes the need to record scope, uncertainty and provenance.
Use preparation to make the archive easier to inspect, then let the assistant retrieve the specific summary, original passage or attachment that the question actually requires.