“Context window” is one of those AI-industry phrases that gets used loosely. This page defines what LADLE's context window means specifically and how to reason about it.
The definition
The context window is the maximum amount of text (including your prompt, attached files, chat history, custom instructions, and the model's response) that Claude Sonnet can consider at once. Measured in tokens; 1 token ≈ 0.75 words on average English. LADLE's window is 200,000 tokens — roughly 150,000 words, roughly 300-500 pages depending on formatting.
Rough conversions
What fits in 200K
- A full 10-K filing plus your notes plus a substantive back-and-forth about it.
- A whole small codebase (up to ~10K lines of TypeScript) plus your questions about it.
- A book chapter or a whole novella, plus discussion.
- 20-30 PDF pages of technical documentation, plus a running conversation.
- A whole research paper (typical 15-30 pages), plus 5-10 related papers, plus your analysis.
What doesn't fit
- A full-length novel (typical 100K+ words = 130K+ tokens is close to the ceiling; plus your interaction takes more).
- A large codebase (~50K+ lines).
- Multi-year archives of documents (like an entire law firm's collected briefs).
- Video transcripts of multi-hour sessions past ~4 hours of dense content.
The 'lost in the middle' failure mode
In long contexts, some models struggle to recall content from the middle of the context (the beginning and the end are recalled better than the middle). Sonnet handles this better than older models but isn't immune. If you're loading a 150K-token document and asking about content in the middle, the model may need explicit pointers (“on page 82 of the PDF, section 4.3, what does the author claim about X?”) rather than a general question.
How to use context well
- For persistent context (files you'll reference across many chats): use Projects (see /docs/projects). Files attached to a project persist across chats without you re-attaching each time.
- For one-off context: paste directly into the chat message.
- For very long documents: split into logical chunks and load one at a time; the model doesn't need everything in one shot.
- For code: attach whole files rather than pasting snippets; the model uses file boundaries as structure.
Cost implications
Every token in the context window consumes input tokens, which count against your plan's rolling 24-hour cap (see /docs/usage-limits). A 100K-token attached file on every message in a 20-message chat costs 2M tokens of input just for the file — hitting most of the Base tier's daily budget. Projects vs per-chat attachments matters here.