Keysat
0ae59f3550
Introduces RECAP_MODE=multi alongside single-mode self-host: - Tenant auth + accounts (magic-link via System SMTP), per-tenant credit pool, anonymous trial minting with per-IP/-64 caps - Self-serve Pro/Max purchase: inline Lightning (BTCPay) + card (Zaprite), prepaid 30-day periods, expiry-reminder emails - Core-decoupling: relay owns cloud tier/expiry keyed by Recaps user-id - SQLite (better-sqlite3) schema for multi-mode; filesystem unchanged for single - StartOS actions/versions through 0.2.155
13 lines
1.2 KiB
TypeScript
13 lines
1.2 KiB
TypeScript
import { VersionInfo } from '@start9labs/start-sdk'
|
|
|
|
export const v_0_2_60 = VersionInfo.of({
|
|
version: '0.2.60:0',
|
|
releaseNotes: {
|
|
en_US: 'Section-boundary precision improvement on long content. The pre-analyze coalesce step (which merged adjacent transcript entries down to ~400 max so a single huge analyze prompt would fit in a local LLM\'s context window) was a holdover from before chunking existed. With the chunked-analyze path shipped in 0.2.59, each 18-minute window only sees ~216 entries at 5-second Parakeet segments — well under any reasonable model\'s context. Coalescing first was unnecessary AND was hurting precision: section starts could only land on ~18-second boundaries because that\'s how coarsely the LLM saw the transcript. Now: content >25 min skips the coalesce entirely and feeds full-granularity entries to the chunker; content ≤25 min still coalesces (safety net for tiny-context local models on the single-shot path). Net effect on long content: section boundaries can now land within ~5 seconds of the actual topic shift instead of being rounded to ~18-second buckets. No change to transcript display granularity (always full).',
|
|
},
|
|
migrations: {
|
|
up: async ({ effects }) => {},
|
|
down: async ({ effects }) => {},
|
|
},
|
|
})
|