Dynamic Columns

About

A dynamic column is generated automatically by the platform. Instead of entering data yourself, you configure a method (like running an LLM prompt or an evaluation) and the platform computes a value for every row.

For example, starting with two static columns:

user_query	expected_answer	model_response	is_correct
What is the capital of France?	Paris	Paris	true
Who wrote Hamlet?	Shakespeare	William Shakespeare	true

Here model_response is a dynamic column created by running a prompt against each user_query. And is_correct is another dynamic column created by running an evaluation that compares model_response to expected_answer.

Dynamic columns can be regenerated at any time. If you change the prompt or switch models, you can re-run the column and the values update across all rows.

---

When to use

• Get model outputs: Run an LLM on every row and store the responses for comparison or evaluation
• Score outputs: Run evaluations and store the results (pass/fail, scores, explanations) alongside your data
• Extract structured data: Pull entities, JSON keys, or classifications out of unstructured text columns
• Enrich with external data: Call APIs or vector databases to add context to each row
• Transform data: Apply custom Python logic to compute derived values

---

Supported Methods

Method	What it does
Run Prompt	Run an LLM prompt that can reference other columns as variables. Learn more
Vector Retrieval	Connect to a vector database and retrieve the top-k chunks for a query
Entity Extraction	Extract named entities (people, organizations, locations) from text columns using a model
JSON Key Extraction	Parse a JSON column and extract specific keys or nested values
Custom Code Execution	Write and run Python code for transformations or complex operations
Text Classification	Assign categories or labels to text using a model
API Calls	Call an external API endpoint for every row and store the response
Conditional Logic	Apply different actions based on conditions (if/else branching across rows)

---

How It Works

1. Choose a dynamic column method from the list above
2. Configure the method (select a model, write a prompt, define the logic)
3. Map input columns (e.g. use user_query as the input to your prompt)
4. Run the column. The platform processes all rows in parallel and fills in the values.
5. View the results in your dataset. Re-run anytime to refresh.

---

Next Steps

• Static Columns: Columns with fixed data you provide directly
• Run Prompt in Dataset: The most common dynamic column method
• Experiments: Compare dynamic column results across different configurations