@juspay/neurolink/dist/types/fileReference.js

The pipe layer of an AI nervous system: one interface connecting provider neurons to your application, across three inference types — generate, stream and decide. `decide` returns typed, calibrated judgements from a non-generative model (~400ms, ~$0.00002

/**
 * File Reference Architecture Types (canonical location)
 *
 * Types for the lazy on-demand file processing system.
 * Files are registered as lightweight references with metadata and previews.
 * Full content is processed on-demand when the LLM requests it via tools.
 */
// Size tier thresholds in bytes
export const SIZE_TIER_THRESHOLDS = {
    /** < 10 KB: inline in prompt */
    TINY_MAX: 10 * 1024,
    /** 10 KB – 100 KB: full load with truncation */
    SMALL_MAX: 100 * 1024,
    /** 100 KB – 5 MB: outline + on-demand */
    MEDIUM_MAX: 5 * 1024 * 1024,
    /** 5 MB – 100 MB: streaming + chunked summarization */
    LARGE_MAX: 100 * 1024 * 1024,
    /** 100 MB – 2 GB: reference only */
    HUGE_MAX: 2 * 1024 * 1024 * 1024,
    // > 2 GB: oversized, rejected
};