Files workflows.

Attaching files to chat is a feature that most AI products claim and few implement well. LADLE's file handling is honest about its limits, which is more useful than a system that pretends to read anything.

Rules of thumb: PDFs work well for text-heavy documents up to ~30 pages per file; spreadsheets work best when they're structured (headers on top row, one concept per column); code files work best when you attach the whole file, not a snippet.

PDF reading

**What works:** contracts, memos, papers, filings, reports. LADLE reads them page-by-page, can cite specific sections, and generally handles multi-column layouts and footnotes reasonably.

**What's harder:** scanned PDFs (OCR-dependent — quality varies), heavily-diagram-based PDFs (image reading is present but less precise than text), tables that span pages or have merged cells.

**The verification move:** after any factual answer from a PDF, ask "quote the specific sentence that supports this". If LADLE can produce the exact quote, the answer is grounded. If it hedges or paraphrases, treat the claim as uncertain and check the source manually.

**Multi-file PDF workflow:** for reading a set of related PDFs (say, the ten most recent quarterly filings from a company), attach all of them to a Project rather than to individual chats. Then in the Project, you can ask cross-document questions ("how has the language in the risk factors section changed across these ten filings").

Spreadsheet analysis

**What works:** structured CSV or XLSX with headers, up to ~10,000 rows. LADLE can compute sums, filter, group, and identify anomalies. It'll show you the specific rows behind an answer, not just the aggregate number.

**What's harder:** merged cells, header rows that aren't at the top, mixed data types in a single column, spreadsheets that rely on formulas visible only in the app (LADLE reads values, not formulas).

**The verification move:** for any aggregation, ask LADLE to show the underlying rows. If the answer is "revenue in Q3 was $2.4M", the follow-up is "list the specific rows that sum to that number". If those rows exist and sum correctly, the answer is verifiable.

**Big spreadsheets (>50K rows):** LADLE will handle them but responses slow noticeably, and context usage goes up. For very large spreadsheets, consider pre-processing: filter to the relevant subset in Excel or Python first, then attach the filtered version.

Code files

**What works:** attach whole files, not snippets. LADLE reads the imports, the shape of the code, and can answer questions about specific functions in the context of the file. Multi-file attachments (up to ~30 files or ~200K tokens combined) let LADLE reason across a small codebase.

**What's harder:** codebases that need real toolchain context (a Rust codebase with unusual macros, a Java codebase with heavy annotation-based dependency injection, etc.). LADLE reads the source, but doesn't run it — dynamic behavior is inferred, not observed.

**The productivity move:** for a bug-hunting session, attach the specific files that touch the failing code path plus the failing test. LADLE can reason about the call graph and often identifies the actual cause faster than a hand-crawl through the files.

**What to avoid:** attaching an entire large codebase (100+ files, millions of tokens). It will either hit the context window or waste tokens on files unrelated to your question. Curate.

Images

Screenshots, diagrams, whiteboard photos, chart images — LADLE can read them, and for many tasks (parsing a chart from a report, explaining a diagram) this works well. Not as reliable as text; sometimes misreads numbers on axis labels or small text. Verify anything load-bearing.

Multi-modal in the same chat

You can attach a PDF, a spreadsheet, and an image to the same chat. LADLE can reason across them: "this diagram in the PDF suggests X, does the spreadsheet data support X". This is often the highest-value use of the file feature — most real analytical questions cross document types.

The gotcha: what files persist

Files attached to a Project persist for every chat in that Project. Files attached to an individual chat only apply within that chat. If you're using a file across multiple conversations, put it in a Project so you don't re-upload it and re-consume tokens each time.