Playground

Test custom prompts against any AI model with real-time response formatting, token tracking, cost metrics, and persistent history.

Overview

The Playground is a free-form prompt testing environment. Unlike the schema-driven enrichment workflow, it lets you send any system prompt and user prompt to a model and inspect the raw response. Use it to experiment with prompt engineering, test model capabilities, or run one-off queries.

Models
Any
Languages
40
History
Persistent
Cost Tracking
Per call

Interface Layout

The Playground uses a split-pane layout. All inputs are persisted in local storage across sessions.

  1. 1The three inputs are tabs of one pane
  2. 2Source documents ride the same call
  3. 3Each run stamps its own time, tokens and cost
  4. 4History, under the answer it belongs to
Prompts on the left, everything the model gave back on the right: because the three inputs share one tabbed pane, the prompt keeps the full height of the window instead of being squeezed into a third of it.

System Prompt

Set the model's behavior and persona. This is sent as the system message and persists between executions so you can iterate on the user prompt without re-entering context.

User Prompt

The main prompt sent to the model. This is where you write your query, instruction, or test case.

Output Schema

Paste an example of the JSON you want back, and it is compiled into a real structured-output contract sent to the provider — the same machinery enrichment uses. This is how you test whether a model can hold a given output shape before building a schema around it.

Response

Displays the model's response with auto-detected formatting. JSON responses get syntax highlighting in the Monaco editor; plain text renders as-is. Copy to clipboard with one click.

History

Every execution is saved to your account, not merely to this browser. Filter by model, read the prompt previews, and click an entry to bring the whole call back. Any record in History can also be reloaded here — its prompts and its attachments — so you can iterate on the wording of a real run without paying for the whole pipeline again.

Configuration

The sidebar configures the call:

OptionDescription
ModelSelect a single AI model from any configured provider. The virtualized dropdown shows pricing and provider info.
LanguageChoose from 40 supported languages. Affects the language instruction in the prompt sent to the model.
AttachmentsAttach documents to the call. The model list narrows to models that can actually read what you attached, so an incompatible pairing is impossible rather than merely discouraged.

The fourth thing you configure is not in the sidebar: the Output Schema tab of the prompt pane. Paste a sample of the JSON you want back and the call carries a structured-output contract inferred from it — the field names, and the types the values imply.

  1. 1The third tab of the prompt pane
  2. 2A sample of the answer, not a JSON Schema
  3. 3Nesting and arrays carry into the contract
Here the sample is a hint: every key becomes a required field, and a null value marks one optional. Choose a model that cannot be held to a response schema and an amber note says so — the schema is then ignored rather than silently half-applied.

Execution & Metrics

After each execution, the response panel displays detailed metrics:

Processing time— Total round-trip time in milliseconds, including network latency and model inference.
Input tokens— Number of tokens in your system + user prompt as counted by the model.
Output tokens— Number of tokens in the model's response.
Cost— Estimated cost in USD based on the model's per-token pricing.
  1. 1Tokens in, then out — as the provider counted them
  2. 2What this one call cost
  3. 3Copy takes the answer alone, without the metrics
The row belongs to the answer under it and is replaced by the next run, which is what makes the Playground the cheap way to settle a model choice: same prompt, two models, three numbers to compare.

Execution History

Every execution is saved to your organization as a prompt record — the same records the History page lists — and the panel under the response reads back the organization's runs:

Model filter— Narrow the list to one model. The picker appears once your history holds more than one.
Prompt preview— Each entry shows the first 50 characters of the user prompt, the timestamp, and the model name.
Success indicator— A green tick or a red cross — failed calls are kept, so a refusal or a timeout stays visible beside the runs that worked.
Restore session— Click an entry to bring back its system prompt, user prompt, output schema, model and language — together with the answer it produced and that call's metrics, without spending anything.
Clear history— Empties the panel in one click by moving every Playground record of your organization to the trash, so they also leave the History page.
  1. 1Filter the list down to one model
  2. 2Clear the whole panel
  3. 3A failed call is kept, and is still replayable
Clearing erases nothing permanently: the entries are records on your account, and clearing moves them to the trash the History page restores from.

Common Use Cases

Prompt Engineering

Iterate on system prompts and instructions to refine model behavior before building them into enrichment schemas.

Model Comparison

Run the same prompt against different models to compare output quality, speed, and cost before selecting models for enrichment.

Quick Queries

Run one-off knowledge extraction queries without needing to set up a full schema and enrichment pipeline.

Next Steps

Enrichment Strategies

Move from free-form prompts to structured enrichment.

Schema Generation

Generate schemas from sample data for structured output.

Models & Pricing

Configure providers and compare model pricing.