Garry's Opinionated OpenClaw/Hermes Agent Brain
TypeScript
/**
* Whether a brain holds any image embeddings (`content_chunks.embedding_image`).
* The cross-modal image arm can only return rows from that column, so a
* text-only brain skips it: a question that mentions photos neither pays for
* a multimodal embed nor reports a failed vector arm while the text arm
* answers. One EXISTS probe per engine, cached for a minute; a probe error
* reads as "has images" so behavior falls back to the pre-probe path.
*/
import type { BrainEngine } from '../engine.ts';
const TTL_MS = 60_000;
const cache = new WeakMap<object, { at: number; value: boolean }>();
export async function brainHasImageVectors(engine: Pick<BrainEngine, 'executeRaw'>, now = Date.now()): Promise<boolean> {
const hit = cache.get(engine);
if (hit && now - hit.at < TTL_MS) return hit.value;
let value = true;
try {
// chunk-visibility-ok: image vectors live on image pages, which are never segmented (chooseBodyStorage)
const [row] = await engine.executeRaw<{ present: boolean }>('SELECT EXISTS (SELECT 1 FROM content_chunks WHERE embedding_image IS NOT NULL) AS present');
value = row?.present === true || (row?.present as unknown) === 't';
} catch {
value = true;
}
cache.set(engine, { at: now, value });
return value;
}
/** Test seam: forget the cached answer for an engine (after writing image chunks). */
export function forgetImageVectorPresence(engine: object): void {
cache.delete(engine);
}