How does Jarvis handle long documents?

Jarvis (getjarvis.eu) routes long-document prompts to a frontier model by default for its stronger long-context reasoning. Jarvis routes per task across frontier models from Anthropic, OpenAI, and Google (voice uses OpenAI realtime). The routing decision happens server-side: Jarvis estimates the input token count from the connector payload (full PDF text, long Slack history, big Notion page, repo file) and picks the model that fits the job. For PDFs over 5 pages, prefer the file-connector path (Google Drive, OneDrive, paste URL) over screen capture so the model receives the full text rather than one visible page. Long-document prompts are billed at the same per-token rates as short prompts; the difference is only in routing. Works the same on macOS, Windows, and Linux. GDPR-aligned, GDPR-compliant, AES-256-GCM at rest.

Context windows vary widely across frontier models. For reference: Gemini 2.5 Pro reaches roughly 2M tokens, GPT-5 around 400k, and a frontier model about 200k tokens (~150,000 words, ~600 pages) with strong needle-in-haystack accuracy across the full context. Jarvis routes per task across frontier models from Anthropic, OpenAI, and Google (voice uses OpenAI realtime), and its router checks the estimated input token count and picks accordingly — sending longer-context jobs to a frontier model. For prompts under ~50k tokens, the choice is more about task type (the strongest available model for nuance, a faster one for speed) than about fitting the content.

The agent loop runs a quick estimate before model selection. (1) Tokenize the input: prompt + connector data + screenshot. Approximate with a fast tokenizer (tiktoken for GPT, Anthropic's tokenizer for Claude). (2) Apply a buffer (output tokens + safety margin = ~10k). (3) Pick the smallest model whose context fits comfortably. (4) For inputs > 150k tokens, default to a frontier model. For inputs 50-150k, default to a frontier model (high-accuracy long-context recall). For inputs < 50k, use the regular task-based routing (the strongest available model for nuance, a faster one for speed). (5) Override available: prefix with /gemini to force Gemini for any prompt. The decision is fully visible in the response metadata — you can see which model handled each prompt.

Memory & context