Skip to guide
OmensDOCUMENTATION
OMENS / USE OMENS

Ask with evidence

Use your own OpenAI API key to explore conditional cross-market scenarios.

Connect your key

  1. Open Omens AI, or the predictions step in Setup & help.
  2. Select the OpenAI option. Enter your API key and a model available to your account.
  3. Select Check model access. This checks access without generating a prediction; it does not verify billing balance or structured-output compatibility.

Review before you send

  1. Write a question about an asset, entity, location or event and choose a horizon.
  2. Select up to 30 fresh context records. Check their sources and missing evidence.
  3. Review the exact evidence snapshot.
  4. Explicitly send the request. Generation uses your API account.
Example research question

How could the weather and market observations in these records affect energy markets over the selected horizon? Identify missing evidence and competing explanations.

Editing the question, model, horizon or selection requires a new review. Feed refreshes do not silently replace an already reviewed snapshot.

Understand the result

Scenarios include a proposed mechanism, affected markets, evidence references, counterevidence and uncertainty. Treat them as hypotheses to investigate. Omens does not claim calibrated forecast probabilities, guaranteed outcomes or automated trading.

The result retains its original question, model and evidence snapshot. Exact-context results may be reused in tab memory for up to five minutes, bounded by evidence expiry. Clear results or reload to remove them.

Keys and request boundaries

Your OpenAI key stays in tab memory. For an explicit check or generation request, it passes through the Omens server to fixed OpenAI endpoints. It is not saved in context or browser storage.

The adapter sends the reviewed question and evidence, uses structured output with no tools, and requests store: false. Provider data policies still apply; this setting is not a promise of zero retention.

The app’s current limits are 30 evidence records, 100,000 context bytes and 3,000 output tokens. These are resource limits, not a fixed price. Requests that time out are not retried automatically.