Context

Managing what the model can see matters more than wording the prompt well. A well-worded question about a spec the model cannot see gets a well-worded generic answer.

What context is​

Everything the model can see when it answers is its context, the working memory for one conversation:

• A system prompt from the tool maker: how to behave, which tools exist, what format to use. In Nomic, this is where the procedures for a drawing review or a submittal review live.
• Your messages and the model's replies.
• Everything the model pulls in along the way: the text of your PDFs and specs, a rendered sheet, search results, a code excerpt.

Every follow-up re-reads the whole list. A good third question in a session can be shorter than a good first one because the context is already there.

Open figure

Where context comes from in Nomic​

The agent's context is scoped per session and starts from two places:

1. The project. Start a conversation inside a project and the agent sees the files and folders on the project's Files page. On Business and Enterprise that includes files synced from SharePoint, Egnyte, Autodesk Forma, Bentley, and other connected sources.
2. Session attachments. The + button on the input bar attaches files already in Nomic, new uploads, or images, for this conversation only. Files and images uploaded or pasted this way stay in the conversation's Session folder. Use Move to keep one beyond the conversation.

A third source is what the model fetches with tools (next lesson). The model can only fetch from what is in scope, so scope is the first decision.

Two consequences:

• Permissions are context. The agent sees only the files you can see. A SharePoint folder restricted to you is invisible in your session, and the agent does not know it exists. An incomplete answer is sometimes a permissions question.
• Files have to be indexed. A file from a connected source can appear in Nomic before the agent can search inside it. A file is ready to search when its Status shows Indexed, and files brought in by a sync need Index Folder (or Index Document) first. If the agent cannot find something you know is there, check that before rewording the question.

Context limits​

Every model has a context limit. As a session grows, each new prompt costs more, the model gets less reliable at picking up details from earlier in the conversation, and eventually the tool compresses or drops earlier material. Tips to manage context limits:

• One task per session. A drawing review and an unrelated RFI question do not share a conversation.
• Start fresh when the agent drifts. Repeating itself, losing track of which sheet, or answering an earlier question are signs that a new session is the cheap fix.
• Keep scope tight. A project with the whole archive attached searches harder and reads more per prompt than one scoped to this job's drawings and specs.

Models & Pricing says the same from the cost side: smaller projects, folders, and workflow context mean less input per prompt.

When the agent cannot find a file you know exists, check scope, readiness, and access before rewording the question. Try the three checks on three cases.

Open figure

Next: how the model finds context on its own.

Next: Tools