Ionosphere.tv
perf: don't cache timestampsNs in concordance — compute on demand
The concordance cache held millions of nanosecond timestamps (one per
word occurrence across all transcripts). These were only used by the
getTimecodes endpoint. Now timestamps are computed on demand from the
transcript when requested, keeping the concordance cache lean.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
+36
-19
+36
-19
apps/ionosphere-appview/src/routes.ts
+36
-19
apps/ionosphere-appview/src/routes.ts
···
5
5
import { readFileSync, existsSync, writeFileSync, mkdirSync } from "node:fs";
6
6
import path from "node:path";
7
7
import {
8
+
decode,
8
9
decodeToDocument,
9
10
} from "@ionosphere/format/transcript-encoding";
11
+
import { lemmatize } from "./lemmatize.js";
10
12
11
13
export function createRoutes(db: Database.Database): Hono {
12
14
const app = new Hono();
···
534
536
app.get("/xrpc/tv.ionosphere.getConcordance", (c) => {
535
537
// Serve from cache if available
536
538
if (indexCache) {
537
-
return c.json({ entries: stripTimestamps(indexCache.entries) });
539
+
return c.json({ entries: indexCache.entries });
538
540
}
539
541
540
542
console.log("[index] Building concordance (this takes a couple minutes the first time)...");
···
577
579
}));
578
580
579
581
const entries = buildConcordance(transcripts, concepts);
580
-
indexCache = { entries, builtAt: Date.now() };
581
-
console.log(`[index] Concordance built: ${entries.length} entries in ${((Date.now() - start) / 1000).toFixed(1)}s`);
582
+
// Strip timestampsNs from cache to save memory — they're computed on demand by getTimecodes
583
+
const stripped = stripTimestamps(entries);
584
+
indexCache = { entries: stripped, builtAt: Date.now() };
585
+
console.log(`[index] Concordance built: ${stripped.length} entries in ${((Date.now() - start) / 1000).toFixed(1)}s`);
582
586
583
-
return c.json({ entries: stripTimestamps(entries) });
587
+
return c.json({ entries: stripped });
584
588
});
585
589
586
590
// On-demand timecodes for a specific term + talk
591
+
// Computed from the transcript rather than cached — keeps memory low
587
592
app.get("/xrpc/tv.ionosphere.getTimecodes", (c) => {
588
593
const term = c.req.query("term");
589
594
const rkey = c.req.query("rkey");
590
595
if (!term || !rkey) return c.json({ error: "missing term or rkey" }, 400);
591
-
if (!indexCache) return c.json({ timestamps: [] });
596
+
597
+
const row = db.prepare(
598
+
`SELECT tr.text, tr.start_ms, tr.timings
599
+
FROM transcripts tr
600
+
JOIN talks t ON tr.talk_uri = t.uri
601
+
WHERE t.rkey = ?`
602
+
).get(rkey) as any;
603
+
604
+
if (!row) return c.json({ timestamps: [] });
605
+
606
+
const decoded = decode({ text: row.text, startMs: row.start_ms, timings: JSON.parse(row.timings) });
607
+
const words = row.text.split(/\s+/).filter((w: string) => w.length > 0);
608
+
const termLower = term.toLowerCase();
609
+
const termWords = termLower.split(" ");
610
+
const timestamps: number[] = [];
592
611
593
-
// Search entries and subentries for the matching term + talk
594
-
for (const entry of indexCache.entries) {
595
-
if (entry.term.toLowerCase() === term.toLowerCase()) {
596
-
for (const talk of entry.talks) {
597
-
if (talk.rkey === rkey && talk.timestampsNs) {
598
-
return c.json({ timestamps: talk.timestampsNs });
599
-
}
612
+
for (let i = 0; i < words.length; i++) {
613
+
const wordLower = words[i].toLowerCase().replace(/[^a-z0-9'-]/g, "").replace(/^['-]+/, "").replace(/['-]+$/, "");
614
+
// Match single words or bigrams
615
+
if (termWords.length === 1) {
616
+
if (wordLower === termLower || lemmatize(wordLower) === termLower) {
617
+
if (i < decoded.words.length) timestamps.push(Math.round(decoded.words[i].start * 1e9));
600
618
}
601
-
for (const sub of entry.subentries || []) {
602
-
for (const talk of sub.talks) {
603
-
if (talk.rkey === rkey && talk.timestampsNs) {
604
-
return c.json({ timestamps: talk.timestampsNs });
605
-
}
606
-
}
619
+
} else if (termWords.length === 2 && i < words.length - 1) {
620
+
const nextLower = words[i + 1].toLowerCase().replace(/[^a-z0-9'-]/g, "").replace(/^['-]+/, "").replace(/['-]+$/, "");
621
+
if (wordLower === termWords[0] && nextLower === termWords[1]) {
622
+
if (i < decoded.words.length) timestamps.push(Math.round(decoded.words[i].start * 1e9));
607
623
}
608
624
}
609
625
}
610
-
return c.json({ timestamps: [] });
626
+
627
+
return c.json({ timestamps: timestamps.sort((a, b) => a - b) });
611
628
});
612
629
613
630
// Invalidate all caches (call after data changes)