← All help topics

Timeouts, tokens and context cost

How model limits work: what tokens are, why a response can be cut off or time out, how a large context affects cost and speed, and what to do about it.

1🧠 Tokens and context in plain words

Models count and bill work in «tokens» — roughly parts of words (Russian text usually takes more tokens than the same English text).

Every request includes:

  • 📝 the system instruction;
  • 🧩 the project context — description, tone of voice, audience, taboos, marketing and extra context, the text of context files;
  • 📄 the material or topic itself;
  • 🤖 and then the model's answer.

The larger the input context and the longer the answer, the more tokens — and the more expensive and slower the request.

2🪙 Which features consume tokens (full list)

Important: tokens are spent not only by the post plan. Many in-app features work via requests to AI models, and each request consumes tokens (your provider's money).

⚙️ Automatically (on schedule, when the project is active and outside quiet hours):

  1. Posts from sources — rewriting new material.
  2. Posts from the model's knowledge — from plan topics.
  3. Images for posts (billed per image, separately from text tokens).
  4. Trends — analyzing source material.
  5. Holiday posts and the holiday check.

👤 By your action in the dashboard:

  1. “Improve project” (the “About project” section).
  2. Autofilling the plan.
  3. Importing from a site and from a Telegram channel.
  4. Refining post text.
  5. Generating and updating (refining) an image.
  6. Generating posts from templates.
  7. “Analyze” a source, analyze-on-add, and “Apply recommendation to project”.
  8. “Recommend sources” (model picks them).
  9. The relevance check when adding a source — embeddings (a little; if unavailable, computed locally, without tokens).

✅ DO NOT spend tokens: the reachability check of recommended links (we just open them over the network), polling sources (tokens only when a post is created from the material), publishing to the channel.

💡 Controlling spend is up to you: per-key limits at your provider, «Generation limits» and a weekly budget, pause and quiet hours, plus the «Billing» and «Model requests» sections.

3⏱️ Request timeout

If the model doesn't reply within the allotted time, the request is aborted and the «Model requests» log shows Error: The operation was aborted due to timeout. This is not the model's answer but a guard against hangs: each request has a limited time budget.

Common causes:

  • 🐢 a slow or «reasoning» model;
  • 📦 a very large prompt;
  • 🔴 provider overload;
  • 🔗 network issues reaching the API.

4✂️ Answer length limit (cut off mid-sentence)

The model's answer has a length cap in tokens (max_tokens). If the answer doesn't fit, it is cut off — in the log you'll see the text ending mid-word.

  • ✅ For ordinary posts the cap is enough.
  • 🧩 For long structured answers (e.g. «Improve project» or website import) the service requests a higher cap.
  • 🔧 If you still see truncation — reduce context or pick a model with a larger answer limit.

5💡 How the service saves tokens (mechanisms)

Token spend is money, so the service is built to spend it economically while staying maximally effective. What's done for that:

  1. 🔁 Only new material goes to the model. When sweeping sources, already-seen links and semantic duplicates are filtered out — the same material isn't sent again; trends don't call the model at all if the material set is unchanged.
  2. ⏱️ Adaptive source polling. A source isn't polled more often than once per 30 min; if it hasn't produced on-topic posts for a while, the interval grows (6 h → 24 h); after 3 days a source isn't disabled right away but rechecked once a day for about another week, and only if nothing on-topic turns up (≈10 days) is it disabled automatically. Plus polling only during the source's active hours and a single poll of identical sources shared across projects.
  3. 🔕 Quiet hours and pause. You can set hours (e.g. night) when the scheduler doesn't run; new projects start paused and only launch once configured; paused projects aren't processed.
  4. 🧠 Context optimization. The «Improve project» button makes the context more compact without losing meaning — it goes into every request, so the savings are continuous. Economy mode and a weekly token budget further cap spend.
  5. ⛔ Limits. You can cap the max drafts per day and per week.
  6. 🖼️ Images. Pictures are compressed and resized to the network format — heavy originals aren't stored or sent in vain.
  7. 📊 Transparency. «Billing» and the «Overview» show exactly where tokens go — by day, provider, project and purpose — so you can quickly find and remove inefficient sources.

Together this yields an economical yet content-rich process.

6🐘 Large context = more cost and slower

Note: everything in the project context is sent to the model on EVERY generation. What enlarges each request:

  • 📝 long «Extra context» and «Marketing context»;
  • 📄 uploaded files (.md / .pdf, up to ~20,000 characters per file).

This raises token spend (money) and response time and brings the timeout closer. Several large files per project can noticeably increase the cost of regular generation.

7🔧 How to resolve it

  1. ✂️ Reduce context: remove unnecessary or bulky context files, trim «Extra» and «Marketing context» to the essentials.
  2. ⚡ Pick a faster or lighter model (e.g. a lite mode) in project settings.
  3. 📏 If answers come out too long — simplify the task: shorter topics and templates, a less «reasoning» model.
  4. 🔑 Make sure a working key for the selected model is set (the AI providers and API keys section) and the API is reachable.
  5. 🔁 Retry later if the provider was overloaded.

8📊 Controlling token spend

To keep spend predictable:

  • ⛔ Set «Generation limits» (drafts per 24h and per 7 days) in project settings — generation pauses when a limit is reached, and all project members get a feed notification (at most once a day). The quota frees up on a sliding window (last 24h and 7 days).
  • 🔍 Watch the «Model requests» section: it shows every real call with request and response — handy to gauge volume and spot bloated prompts.
  • ✨ Keep the context concise: prompt quality matters more than length.

Didn't find an answer? Write to us — we'll help.

Contact support