Context Estimator Guide

Choose practical context windows for AI chats before you load a local model or paste a long document.

Quick Navigation

→ Overview → Basic Workflow → Model Profiles → Reserve → Statuses → Privacy

1. Overview

Context Estimator gives you a fast estimate of how much AI chat context a piece of text or a document may need. It is useful before choosing settings such as 4K, 8K, 16K, 32K, 64K, or 128K in tools like Ollama, LM Studio, or local browser chat.

Estimate, not exact count: Different models tokenize text differently. The default estimate is a fast local heuristic. For precise numbers, use the Exact count button — it downloads the real tokenizer of the selected profile (a few MB, only when clicked) and counts your text with it.

2. Basic Workflow

1

Paste or upload

Paste text into the input area or choose a local TXT, Markdown, DOCX, or PDF file (up to 25 MB). Scanned PDFs have no text layer — run them through the Markdown Converter (OCR engine) first.

2

Choose profile and reserve

Select the closest model profile and keep a reserve for system prompt, answer, formatting, and reasoning.

3

Read the comparison table

Use the row marked recommended: it is the smallest window with real headroom (a “Tight” window is only recommended when nothing larger fits). Every row also shows how many tokens would be left for the answer.

3. Model Profiles

Generic LLM

A safe default when you do not know which tokenizer your model uses.

Model families

Llama / Mistral, Qwen, Gemma, and GPT-style profiles adjust the estimate for common tokenization patterns.

4. Reserve

A context window must hold the source text, your instruction, system prompt, and generated answer. The reserve slider keeps part of the window free before making a recommendation.

The default 35% is a practical starting point. Increase it for detailed answers or reasoning. Lower it only when you expect a short answer. The Short answer / Standard / Detailed buttons set the slider to matching presets (20% / 35% / 50%).

5. Statuses

Too small / Tight

The input may overflow, or there may be little room left for a useful answer.

Good / Overkill

Good is the practical target. Overkill means the window is likely larger than needed.

6. Privacy and Limits

Text extraction and the heuristic estimation run locally in your browser with self-hosted libraries — no external requests. Pasted text and uploaded files are not sent to ASD123.ai for processing, and document contents are not saved in localStorage or IndexedDB. The optional Exact count downloads tokenizer files from CDN/Hugging Face when you click it; your text itself is never transmitted. Estimates include a fixed allowance of about 300 tokens for system prompt and formatting overhead. Heuristics for non-Latin scripts (Cyrillic, Arabic, Thai, ...) are coarser than for English — use Exact count when precision matters there.

PDF extraction depends on readable embedded text. Scanned, image-only, password-protected, or unusual PDFs may need OCR or manual cleanup before estimation.