mirror of
https://github.com/mauriceboe/TREK.git
synced 2026-10-23 15:36:33 +00:00
Maurice
4ceade9d8b
Cloud models got the first 4,000 characters, so totals and lodging sections further down never reached them, without a word. They now get 24,000 with the start and the end kept and a warning when the middle is left out; a server whose context is too small is asked once more with 8,000. A total the model repeated on every flight of one booking stays on the first flight only, when the document prints it once as a total, and the prompt says so too.
53 lines
2.3 KiB
TypeScript
53 lines
2.3 KiB
TypeScript
/**
|
|
* How much extracted text a cloud model is sent. A confirmation keeps its total
|
|
* and payment block near the end, and 4,000 characters cut both off a long email
|
|
* without a word (#2652). 24,000 characters are about 6,000 input tokens, far
|
|
* under the context of any current cloud model, and cost roughly one or two cents
|
|
* per import on a mid-tier model, a fraction of a cent on a small one.
|
|
*/
|
|
export const CLOUD_TEXT_WINDOW = 24_000;
|
|
|
|
/**
|
|
* A local model reads on the CPU, where every character costs prompt-eval time. A
|
|
* flight itinerary lists its legs throughout a long document, so it keeps a
|
|
* generous window; a single booking has the essentials up top (an 11-page rental
|
|
* voucher took ~200s at 16k, its booking data sat in the first ~2k).
|
|
*/
|
|
export const LOCAL_FLIGHT_TEXT_WINDOW = 16_000;
|
|
export const LOCAL_TEXT_WINDOW = 6_000;
|
|
|
|
/**
|
|
* The second try for a server that refused the cloud window for its context
|
|
* length: a self-hosted vLLM or llama.cpp with 8k tokens takes this together
|
|
* with the prompt, the schema and the answer (#2652).
|
|
*/
|
|
export const COMPACT_TEXT_WINDOW = 8_000;
|
|
|
|
/** Share of a window kept from the start; the rest comes from the end. */
|
|
const HEAD_SHARE = 0.75;
|
|
|
|
/** How far a cut may move to land on a line break instead of inside a line. */
|
|
const LINE_SNAP = 200;
|
|
|
|
/**
|
|
* The text cut to `window` characters, keeping the start and the end. The legs
|
|
* and names sit up top, the total and the payment summary at the bottom, so the
|
|
* middle is what goes. The cuts land on line breaks where one is close, so an
|
|
* amount is never split in half, and a marker says how much is missing.
|
|
* `omitted` is 0 when the text fit.
|
|
*/
|
|
export function fitTextWindow(text: string, window: number): { text: string; omitted: number } {
|
|
if (text.length <= window) return { text, omitted: 0 };
|
|
let headEnd = Math.floor(window * HEAD_SHARE);
|
|
let tailStart = text.length - (window - headEnd);
|
|
const lineBefore = text.lastIndexOf('\n', headEnd);
|
|
if (lineBefore > 0 && lineBefore > headEnd - LINE_SNAP) headEnd = lineBefore;
|
|
const lineAfter = text.indexOf('\n', tailStart - 1);
|
|
if (lineAfter !== -1 && lineAfter < tailStart + LINE_SNAP) tailStart = lineAfter + 1;
|
|
const omitted = tailStart - headEnd;
|
|
return {
|
|
text: `${text.slice(0, headEnd)}\n[... ${omitted} characters omitted ...]\n${text.slice(tailStart)}`,
|
|
omitted,
|
|
};
|
|
}
|