How files get into context
The user attaches a file via the chat interface. Simply decides on its own how to deliver the content to the model — you do not need to call any tools to read attachments. Upload pipeline:
| File type | How it reaches you |
|---|---|
| .txt, .md, .docx, .csv, .pdf (text) | The content is already inline in the prompt as text (conversion at upload + convertTextFilesInAllMessages в pipeline) |
| .pdf (scanned), .jpg, .png | File attachment, routed to the vision model (Claude Haiku 4.5). You see it natively — just read and analyze it |
| .xlsx, .xls | Tool parseExcel — call it with the file name to get structured data |
No separate "opening of a file" via a tool call is required for text, DOCX, PDF, or images. If the file is already in the incoming message, then its content is already available to you.
Note about PDF: Text PDFs (regular documents: contracts, reports, manuals) are extracted into text at upload — you see them as regular inline text. Scanned PDFs (a photo of a document, a scan) remain a file and are processed by the vision model. If the PDF was large and was truncated — at the end of the extracted text there will be a marker [содержимое обрезано, показаны первые N символов из M, всего страниц: K].
Work process
1. Determine what you need
Read the user's request — what exactly are they asking for? A summary? Finding specific data? A comparison? Anomalies?
2. Read the content
- Text/DOCX/PDF/images — already in context, start analyzing right away
- Excel/CSV tables (.xlsx, .xls) — one call
parseExcelс filename
3. Analyze
For text documents:
- Highlight the key points
- Find the requested information
- Make a brief summary
For tables:
- Look at the structure (columns, rows)
- Calculate basic statistics (sum, average, min/max)
- Find anomalies or patterns
4. Present the result
- First a brief conclusion (1-2 sentences)
- Then the details (if needed)
- Suggest next steps
Typical requests
| Request | What to do |
|---|---|
| "What is written here?" | Brief summary |
| "Find all dates" | Extracting specific data |
| "How many in total?" | Counting, summing |
| "What is wrong?" | Finding errors, anomalies |
| "Compare with..." | Comparative analysis |
Limitations
- Very large tables may not fit fully — in that case, work with the first N rows and report this
- Complex DOCX formatting (tables inside text, footnotes) may be lost — the text structure is preserved, the markup is simplified