Collect data from documents into a table
A folder holds PDFs, scans or exports, and you need one table with fields from them. You open the folder to the agent and describe the columns. The agent reads the files, builds the table and tells you what it could not read.
What you need
- Liho Cowork for macOS, Windows or Linux, signed in with a Liho LLM account.
- A positive Liho LLM balance: each agent response is paid at the price of the selected model.
- A folder with documents. For important originals, “Read only” access is the safer choice.
Steps
Sign in and start a task
On first launch, click “Sign in with Liho LLM” and confirm in the browser; you never type a password into the app. The new task screen has a “Describe the task” field.
Download Liho CoworkOpen the folder with the documents
In the task field, click “+” → “Files and folders” → “Add folder” and choose the folder. Then choose the access level:
- “Read only”: the agent reads the files and changes nothing in the folder. The table appears in the task’s work folder.
- “Read and write”: the agent can create, change and delete files in the folder. Choose it if the table should sit next to the documents.
Single files can be added with “Attach files”, dragged into the window or pasted from the clipboard.
Describe the table
List the columns, the file format and what to do with unclear cases. The more precise the description, the fewer follow-up questions.
Example taskRead every document in the “Receipts 2026” folder. Build an xlsx table with the columns: date, seller, amount, VAT, file name. Sort by date and add a sheet with monthly totals. If a document is unreadable or an amount is ambiguous, do not guess: flag the row and list those files at the end.Check the task settings
Next to the input field you choose the model, approvals, isolation and tools. By default approvals are set to “Auto” (files and commands without asking) and isolation is off. For documents from untrusted sources:
- turn on “Process sandbox” or, on Windows, “Virtual machine”, so the agent is confined to the chosen folders (on Windows the process sandbox restricts writes only);
- choose “Auto for reads”: reading happens without asking, while changes and commands wait for your approval.
Follow the work and answer questions
The agent shows its plan and every step: which files it read and which commands it ran. If information is missing, it asks a question with suggested answers, in the chat and under “Decisions”. Messages you send while it works are queued. “Stop” interrupts the work.
Check the table
The finished file appears as a card in the chat and in the “Results” panel. You can open it, view it in the built-in viewer (CSV opens as a table) or click “Show in folder”. Compare a few rows with the originals, especially the ones the agent flagged. Each response shows the model, tokens and cost.
Repeat without extra setup
If such documents arrive regularly, create a project with instructions, for example about the column format, or schedule the task under “Scheduled”. A local schedule runs only while the app is running, and only with isolation turned on. To run tasks while your computer is off, use your own server.
Your own server and scheduled tasks
What you get
- A table in the task’s work folder or in the folder you chose.
- A list of documents the agent could not read or where the data disagrees.
- The cost of the task, shown under each agent response.
Limits
- The built-in Word, Excel, PowerPoint and PDF tools are only wired up on Windows in “Virtual machine” mode. Everywhere else the agent uses programs and libraries already installed on the computer. If they are missing, ask for the table as CSV.
- Scans and photos are not always read reliably: the result depends on the image quality and the model. Spot-check the amounts.
- By default new tasks run without isolation and with “Auto” approvals: the agent changes files and runs commands without asking. Give important folders “Read only” access.
- Every agent step is a model request. More documents mean a more expensive task; the cost is shown in the chat.