7trade
Install

The "AI" window: language model

Читать по-русски

The bot can ask a language model — the same kind people chat with — what it thinks about a trade. This article explains how to connect a model, how it takes part in a strategy and how to set how many questions the bot asks. The strategy rules themselves are described in the article about the strategy builder.

What to know right away

  • Every question is paid. You pay your provider — the company whose model answers. The bot charges nothing for questions.
  • The model does not know the future. It reads the same candles and indicators as the bot and answers with its opinion. It is one more confirmation, not a forecast.
  • An answer takes seconds. The bot decides at a candle close, and the answer must arrive before the next one starts. On short timeframes the model often does not make it.
  • The model is off until you turn it on.

Where to open it

The "AI" item in the bot's top bar. The window has three tabs: "Prompts", "Connection" and "Mode".

It is easiest to set it up in this order: connection, prompt, mode.

Connection

Key. For the model to answer, you need a provider key. The provider issues it on its own site. The key is entered in the 7tra.de dashboard, the "Model keys" section, not in the bot window: the bot gets it by itself, through the extension. This way the key does not pass through the platform's page.

On the "Connection" tab:

Field What it sets
Provider Whose model to ask. The list shows providers you have a key for.
Model Which model of this provider. Remembered separately for each provider.
Custom API address For your own server with a compatible model. Usually empty.
Model keys in the dashboard Opens the dashboard on the site.
Refresh Fetch keys from the account right away — for example, if you have just added a key.
Check the connection Ask the model a test question. The bot shows which model answered and how fast.

You need to be signed in to your 7tra.de account: without it, keys from the dashboard do not reach the bot.

Prompts

A prompt is the text the bot sends to the model together with the data. In it you explain to the model how you trade and what to look at.

The "Prompts" tab lists your prompts. One, the active one, goes to the model. To choose it, press "Apply the prompt" on its tile. Without an active prompt the bot does not ask the model.

To create a prompt: type a name and press "Add" — the editor opens.

The prompt editor

Section What is in it
Answer protocol How the model must answer. Not editable: the bot understands the answer by this text.
System prompt Your description of how you trade. Empty — the bot puts in its own.
Data What to pass to the model: candles, indicators, patterns and more. Turn on at least one block.
Own text Added at the end, before the task.
Task The question itself. Empty — the bot builds it from the strategy.
Preview What exactly goes to the model for the open asset and roughly how long it is.

At the bottom — "Save" and "Save and apply".

Mode

Work

  • "Use the language model" — the main switch. Unticked — the bot does not ask a single question.
  • "What to do with the decision":
    • "Only show decisions" — the bot asks the model and shows its answer on the "AI" tab on the right, but does not open trades by it;
    • "Open trades" — the model's answer takes part in trading.

A strategy makes the same decisions in both modes. The only difference is whether it gets as far as a trade. So start with "Only show decisions", and when the answers suit you, switch.

Thresholds

Field What it sets
Minimum confidence, % The model states its confidence. An answer with confidence below the threshold is not accepted.
Timeout, ms How long to wait for an answer. In practice the bot waits no longer than what is left of the candle.
Simultaneous requests How many assets to ask about at one candle close. More — wider coverage and more expensive.
Do not ask more than the trades that fit Do not spend questions on signals for which there is no place for a trade anyway.
If the model did not answer "Do not enter" or "Open as if there were no filter".
Reasoning depth The deeper, the slower and more expensive the answer. For short timeframes — "Off" or "Low".
Answer limit, tokens The maximum length of the answer. A cut-off answer is lost but already paid for.

Spending ceilings

"Calls per hour" and "Calls per day" — how many questions the bot may ask. Zero — no ceiling. Providers do not report the price of a question, so the bot counts questions, not money.

When a ceiling is reached, the bot stops asking until the next hour or day.

How the model takes part in a strategy

There are two ways.

A confirmation for a normal strategy

The "Strategy rules" window has a "Model participation" section:

  • "Confirm the entry with the model" — on every strategy signal the bot asks the model, and the trade opens only if the model agrees;
  • "The model may name its own side" — off: an answer in the other direction counts as a refusal. On: the trade goes where the model said;
  • "The model decides on martingale" — the model answers whether to apply stake growth to the trade. The martingale checkboxes in the settings still take priority;
  • "Expiry" — "As in the bot settings", "Own value" or "The model chooses" within the given limits.

The model is asked only when a signal has already fired. So there are few questions.

AI strategy

A strategy without rules: the model makes the decision entirely. It is created on the "Mode" tab with the "Create an AI strategy" button. It is run, edited and deleted in the "Strategies" window, like the others.

Such a strategy asks the model much more often than a confirmation does: every time something happened on a candle — the indicators clearly leaned one way or a candlestick pattern formed. To keep spending in check, the bot asks about no more than a couple of assets per candle and pauses between questions about one asset. Still, it costs noticeably more. Set spending ceilings before running it.

In an AI strategy the side is always named by the model, and "Confirm the entry with the model" cannot be turned off: it has no other source of decisions.

The model's answers: the "AI" tab on the right

The "Bot" panel has an "AI" tab. It shows what the model answered for the current asset:

  • UP, DOWN or SKIP;
  • confidence;
  • "Why" and "Risks" — the model's explanation;
  • "What happened" — the feed of questions for the last hour.

The "Ask now" button asks a one-off question about the open asset. It does not open a trade. The same question can be asked from the "Patterns" tab — with the button on a pattern's line.

"Skip" is a normal and the most common answer. An empty feed is not a fault either: there may be no signals for hours.

What the model does not do

  • It does not trade while the lock is closed. The trading lock and "Start" apply to it too.
  • It does not see your account. The prompt gets candles, indicators and patterns, but not the balance or the account number.
  • It does not promise a win. The model's confidence is its opinion, not a probability.

Frequently asked questions

The model does not answer, the "AI" tab says "timeout". The model does not make it before the candle closes. Take a faster model, lower "Reasoning depth" or work on a longer timeframe.

The window says the provider rejected the key. The key is wrong or has expired. In this case the bot turns the model off by itself. Check the key in the dashboard, press "Refresh" on the "Connection" tab and tick "Use the language model" again.

A strategy with model confirmation is running, but there are no trades at all. Look at the "What happened" feed on the "AI" tab. The model may have answered "skip" or not made it before the end of the candle. With "If the model did not answer" = "Do not enter", a late answer means a refusal.

The strategy tile says "model off". The strategy asks for confirmation, but the model is off. No questions are asked, and trades go by the strategy's rules alone. Press the mark — the "Mode" tab opens. An AI strategy with the model off does not start at all.

What's next