Reading documents

The agent can read supported files from the conversation without adding them to the knowledge base.

Main section: Tools.

Reading documents from a dialog

The agent can read an attached document and use its contents in a reply. For example, a user attaches a price list and asks: “Find the delivery cost in this file.” No separate setting is required, and the file does not need to be added to the knowledge base.

Supported files

Supported formats are TXT, Markdown, PDFs with a text layer, DOCX, CSV, and XLSX, up to 5 MiB (5,242,880 bytes); the project's file limit also applies. Only attachments from the current dialog are accessible. For multiple files, identify the required one by name and explain what to find, compare, or calculate.

Files are not read automatically or added to the knowledge base. Forwarding a prepared attachment does not require reading it. Document reading does not recognize scans without a text layer; attached images use the separate image recognition capability.

Reading limits

Long documents are read in parts: up to 8,000 characters by default, at most 12,000 per read and 48,000 in total per run, including reads across multiple files. For larger amounts, specify the relevant sections or split the task into several requests. The agent must state what it read and what remains rather than claim to have reviewed the entire document. No separate paid scan recognition starts, but the extracted text counts toward model token usage.