Skip to main content
Sofie can do more useful work when the right information fits into the current chat context. The context window is the working space Sofie uses for the next response. It can include your recent messages, selected context, attachments, source results, artifact excerpts, and the instructions needed to complete the task. As a chat grows, that working space can fill up. Sofie helps by showing a context indicator and, in long conversations, summarizing older parts of the chat so you can keep working.
Context management is about relevance. A smaller set of clear sources usually works better than a large pile of loosely related material.

​
What a context window is

Every AI model has a limit on how much information it can consider at one time. That limit is the context window. Think of it as the material Sofie can carry into the next answer:
  • Your current message.
  • Recent messages in the conversation.
  • Attached file content or file summaries.
  • Selected CoDraft, CoSheet, CoMeeting, Workspace, Orchestration, or plan context.
  • Results from Workspace search, Web search, or Deep research.
  • Relevant tool results, artifact edits, and cited source snippets.
  • System instructions that tell Sofie how to behave.
The context window is not the same as your full Sofie account history. Sofie does not automatically place every file, meeting, document, spreadsheet, or past chat into every answer.

​
Why context matters

Context affects what Sofie can reason from in the current response. For life sciences work, context quality is especially visible in source-heavy tasks such as deviation investigation, CAPA development, URS review, validation protocol drafting, batch record review, and regulatory research. Sofie can help organize and analyze source material, but you still need to review the sources and conclusions before using the output in controlled work.

​
Tokens in plain language

Sofie measures context in tokens. A token is a small piece of text, such as part of a word, a full word, punctuation, a number, or formatting. You do not need to count tokens yourself. Use tokens as a rough signal:
  • Short prompts use few tokens.
  • Long pasted text uses more tokens.
  • Attachments, transcript excerpts, spreadsheet ranges, search results, and artifact content can add a lot of tokens.
  • Some models have larger context windows than others.
When a task needs many sources, use a Workspace and ask Sofie to search for the relevant parts instead of pasting everything into chat.

​
Use the context indicator

After Sofie responds, the chat input can show a small circular context indicator near the top-right of the input area. Hover over it to open the Context Window card. The card shows: Use the indicator as a warning light, not as an exact quality score. A low percentage does not guarantee that Sofie has the right sources. A high percentage does not always mean the task will fail, but it means you should narrow the source set, summarize, or start a cleaner chat.

​
What raises context usage

Context usage usually grows when you:
  • Keep one conversation going across many different tasks.
  • Attach large PDFs, spreadsheets, presentations, images, or transcripts.
  • Add several Workspaces, CoDrafts, CoSheets, CoMeetings, or Orchestrations at once.
  • Use Deep research or broad Web search queries that return many sources.
  • Ask Sofie to compare many files without narrowing the question.
  • Paste long source text directly into chat.
  • Ask Sofie to create or revise large artifacts repeatedly in the same thread.

​
What does not automatically fit

Sofie can find and use information from selected sources, but the full contents of every connected source are not always active at the same time. Do not assume the next answer includes:
  • Every file in a Workspace.
  • Every row in a large CoSheet.
  • Every line of a long CoMeeting transcript.
  • Every message from a long chat in exact wording.
  • Every result from a broad Web search or Deep research run.
  • Every previous version of a CoDraft.
Ask Sofie what it is using when the source set matters:

​
How summarization and compaction work

When a conversation becomes long, Sofie may summarize older messages so the conversation can continue. You may see a divider labeled Messages summarized. Expand it to review the summary and the number of messages summarized. This process is sometimes called compaction. It keeps the conversation usable by replacing older detail with a shorter summary of what happened earlier. Compaction is useful because it can preserve:
  • The main task and objective.
  • Decisions made in the chat.
  • Sources and artifacts discussed.
  • Open questions.
  • Assumptions.
  • Follow-up steps.
  • Important constraints you gave Sofie.
Compaction can also lose detail. Exact wording, minor source distinctions, intermediate reasoning, or stale branches of the conversation may no longer be present in full.
Treat a summarized section as a working summary, not as a record copy. Reopen source files, artifacts, or original citations when exact wording matters.

​
Summarization is not Memory

Sofie has several continuity tools. They do different jobs. If you want Sofie to remember a durable preference, use Memory or a Directive. If you want a team source set to remain available for future work, use a Workspace. If you want a long conversation to remain readable, use conversation summarization and pinned chats.

​
When to keep going and when to start fresh

Use this table when the context indicator is high, Sofie seems to miss details, or the chat has become broad.

​
Manage context during source-heavy work

1

Define the task

Tell Sofie the exact job: review, compare, extract, summarize, draft, analyze, or build.
2

Choose the smallest useful source set

Add only the Workspace, artifact, attachment, meeting, sheet, or search mode needed for the current step.
3

Check context chips

Review the chips above the input. Remove anything that should not influence the answer.
4

Ask for source visibility

Tell Sofie to list sources, cite key claims, separate assumptions, and flag missing evidence.
5

Watch the context indicator

If usage is rising, summarize, narrow the sources, or move durable work into an artifact.
6

Create a handoff before switching tasks

Ask Sofie for a continuation summary before starting a new chat or changing the source set.

​
Prompts for better context control

Use these prompts when a chat is long, source-heavy, or moving into a new phase.

​
Examples for life sciences work

​
Deviation investigation

Start with a focused source set:
If the chat gets long:

​
Validation protocol drafting

Keep source review separate from drafting:
Then move durable output into CoDraft:

​
Batch record review

Use the context indicator to avoid overloading one chat with too much source material:
For broad batch packages, use a Workspace and work in passes:

​
What to do when Sofie hits the limit

If Sofie shows Context Limit Exceeded, the active context is too large for the next response. Recover with this flow:
1

Start a new chat

Click Start New Chat from the error card or open a new chat from the sidebar.
2

Bring forward the minimum context

Paste a short continuation summary, add the key Workspace or artifact, and attach only the source files needed for the next step.
3

State the next action

Ask for one output, such as a source table, CoDraft section, CoSheet analysis, or Orchestration design.
4

Ask Sofie to flag gaps

Tell Sofie to list anything it cannot know from the new context.
Useful restart prompt:

​
Keep context clean over time

Good habits:
  • Start a new chat when the topic, Workspace, source set, or output changes.
  • Rename and pin chats that contain useful work trails.
  • Move durable outputs into CoDraft, CoSheet, CoMeeting, Workspace, or Orchestrate.
  • Use Workspace search for project knowledge instead of pasting long source text.
  • Ask for continuation summaries before long chats become hard to follow.
  • Use Plan Mode for multi-step work so you can review the path before Sofie executes.
  • Use citations and source tables when output needs review.
  • Remove context chips that are no longer relevant.
The best context habit is to make each chat answer one clear question from one clear source set. When the job changes, start a cleaner thread.