Run a prompt on every row
Turn a prompt into a column of model output, one response per row.
Run Prompt fills a new dynamic column by running a prompt against every row of a dataset that already exists. You write the prompt once, referencing other columns as inputs, and Future AGI runs it row by row until the whole column is filled. You’ll need a dataset with the input columns your prompt will reference.
Open Run Prompt
On a dataset’s Data tab, click Run Prompt. This starts a new column and opens the panel where you build the prompt and pick a model.
Name the output column
In the Name field (placeholder “Prompt Name”), name the new column model_response for this example. It’s the first field in the panel, above the model type options, and it becomes the name of the column every row’s response lands in.
Choose a model type
Run Prompt supports four kinds of models, each shaped the same way: pick a type, then pick the specific model from that type’s list.
| Model type | Input | Output |
|---|---|---|
| LLM | The prompt you build next | Text |
| Text-to-Speech | A text column referenced in the prompt | Audio |
| Speech-to-Text | An audio column | Transcribed text |
| Image Generation | A single image prompt | An image |
LLM prompts are the message-based kind covered next. For LLM models, don’t see the model you need? Register a custom model and it joins the same list.
Build the prompt
An LLM prompt is a list of messages with roles:
- System (optional): instructions that set the model’s behavior and context
- User (required): the input message, built from your dataset’s columns
Use {{column_name}} inside a message to pull that column’s value for the current row. Take a dataset with a user_query column and a customer_context JSON column:
System
You are a support assistant that helps resolve customer tickets.User
A customer asked: {{user_query}}. Their plan is {{customer_context.plan}}. Write a helpful response.{{user_query}} pulls that row’s plain text value. {{customer_context.plan}} uses dot notation to reach the plan field inside the customer_context JSON column, without pulling in the rest of that column’s value.
Text-to-Speech, Speech-to-Text, and Image Generation prompts are simpler, since each is a single input instead of a message list:
- Text-to-Speech: in the Prompt Input box, reference the text column to speak, for example
{{script_text}}, and choose a Voice, which is required - Speech-to-Text: pick a column in the Voice Input section’s Column dropdown, which lists your audio columns; selecting one fills the message for you
- Image Generation: write the prompt describing the image directly in the Image Prompt field
Set concurrency
Set how many rows run at once, from 1 to 10. It defaults to 5.
Set generation parameters and tools
Adjust generation parameters such as temperature, top P, max tokens, presence and frequency penalty, and response format, if the defaults don’t fit your prompt, from the options button beside Select Model.
Attach tools the model can call while it runs, if your prompt needs them, in the Tool Configuration accordion above Concurrency.
Run
Click Run. Future AGI works through the dataset row by row and writes each response into the new column. Watch a row’s cell to see it complete; the column is done once every cell has filled.
What lands in the column
While a row’s call is in flight, its cell shows a loading placeholder until the response lands. If the call fails, its cell shows an error. Otherwise the cell fills with the response, and each LLM cell also records its token counts and response time.
Dive deeper
Questions & Discussion