Working with files.

LADLE supports 30 file attachments per chat, up to 200K tokens of context (about 300 pages of PDF, or 60K lines of code). Most people either underuse this by pasting short snippets, or overuse it by trying to fit an entire codebase in one chat. Both fail differently. Here's how to use it well.

Upload the file, don't paste its contents.

The composer accepts drag-and-drop for PDFs, images, code files, CSVs, and Excel. Uploading is better than pasting because LADLE preserves structure — a PDF stays paginated, a spreadsheet retains its rows and columns, a code file keeps its formatting. Pasted contents lose all of that and land as one blob of text, which the model can still read but which produces measurably worse answers.

Match the file size to the question.

If your question is about one specific section of a 300-page filing, don't upload the whole filing — upload the 20 pages you actually care about. The full file will work, but the model will spread its attention across the whole thing and give more general answers. Chunking a document to the specific part you care about often gives sharper responses.

Let it read before you interrogate.

The natural instinct is to upload a file and immediately ask "summarize this". A better pattern is: upload the file, then ask "read this and tell me what you see" as your first prompt. LADLE will describe the structure, the key sections, and any notable features — which orients both of you before you start asking specific questions. This adds one prompt and improves every subsequent answer.

Attach files to a Project when you'll reuse them.

If you're doing several chats over a week that all reference the same background document (a spec, a style guide, a set of reference materials), attach those files to a Project rather than to individual chats. Project files stay available across every chat in that Project without eating into the per-chat file count.

The 200K window isn't infinite.

If you're pushing the window (long documents plus long conversation), later prompts effectively can't see the earliest parts of the chat anymore — the model has to summarize its own earlier state, which loses fidelity. Signs you're hitting this: it starts saying "as we discussed earlier" without correctly recalling what you actually discussed. When this happens, start a fresh chat and re-attach the file with the sharpened question you now know how to ask.

Formats that don't work well.

Video and audio uploads aren't supported. Extremely large images (>10MB) may compress badly and lose text-readability. Password-protected PDFs are rejected. Encrypted files won't decrypt on our side — decrypt locally first.