When it fires
In a sync of the knowledge base, for each chunk that is embedded, after its vector came back from the AI provider and before it is sent to the vector database. Unchanged chunks are skipped by a sync and do not pass it, and the metadata is not part of what tells a chunk changed: after adding or changing a callback, run a forced sync (Force re-sync in the Sync tab, or wp kimaai vectors sync --force) so every chunk gets the new fields.
Add fields, do not remove or rename KimaAI’s: the knowledge search reads post_id, title, post_url, anchor, heading and chunk_text, and filters by lang, post_type, access and the tax_ fields. Use strings, numbers, booleans or lists of strings, which every vector database stores. A changed access does not widen who sees a chunk: every hit is checked against the post’s current rules before it is used.
Parameters
| Name | Type | Description |
|---|---|---|
meta | array | The metadata KimaAI stores with the chunk (vector_id, post_id, post_type, mime_type, post_url, slug, title, lang, access, heading, chunk_index, model, dim, token_count, chunk_text, anchor and a tax_<taxonomy> list per scope taxonomy). |
post | WP_Post | The post the chunk belongs to. |
chunk | array | The chunk, with text, chunk_index, heading_path, anchor, token_count, start_char and end_char. |
clean | array | The post’s prepared text, with lang, text, title and html. |
Returns
array: The metadata to store.
Example
add_filter(
'kimaai_embedding_vector_metadata',
function (array $meta, WP_Post $post): array {
// Who a document is written for, from a custom field: the knowledge search can filter by it.
$audience = (string) get_post_meta($post->ID, 'audience', true);
$meta['audience'] = $audience !== '' ? $audience : 'everyone';
return $meta;
},
10,
2
);
Related
kimaai_kb_search_meta_filter: filters the knowledge search by these fields.kimaai_embedding_chunk_text: changes the text a chunk is embedded with.