Skip to main content

Indexing Your Content

The bot can only answer from what it has been given. This page covers the three sources it draws on, how content is kept current, and how to keep things out of it.

Source 1: Your published content

In Flarepipe Chat > Settings, tick the post types the bot should read. Posts and pages are selected by default; any public post type can be added. Only published content is indexed; drafts, private posts and pending items are ignored.

Each item is run through the same filters your theme uses, stripped of markup, and split into chunks of roughly 1,500 characters. Every chunk is stored with the URL of the page it came from, which is how the bot is able to link a visitor to the source of an answer.

Excluding individual items

Every post and page gets an Exclude this content from the chatbot checkbox in the sidebar of the editor. Use it for anything that would be misleading out of context: pricing pages that change often, landing pages full of promotional phrasing, legal text you do not want paraphrased, old announcements that are still published but no longer true.

Old content is the usual culprit

When a bot gives a confidently wrong answer, it is far more often because it found a genuine but outdated page than because the model invented something. Before launch, look at what you have published that is no longer accurate, and exclude it.

Source 2: Chatbot-only knowledge

The plugin adds a Chat Knowledge post type. Entries there are indexed but never rendered on your site. This is where the answers live that nobody wrote a page about:

  • Questions support gets weekly that the site does not address.
  • Policy details: refunds, turnaround times, what you do and do not cover.
  • The specific wording you want used for a sensitive or easily-misread topic.
  • Corrections: if the bot keeps getting something subtly wrong, write the right answer here.

Writing these as a plain question followed by its answer works well, because it closely matches the shape of what a visitor will type.

Source 3: Site facts

The settings screen has a Site facts box. Unlike the other two sources, this text is injected directly into the model's instructions on every single request rather than being retrieved when it seems relevant.

Use it for the small number of things that must always be true and must never be missed: what your business is, what you sell, your opening hours, the one product name people get wrong. Keep it short. Everything you put here is sent with every message, and a long block dilutes the instructions around it.

Anything that is merely useful belongs in a knowledge entry, not here.

Keeping in sync

  • On save. Editing a post queues a re-index moments later, so corrections take effect almost immediately.
  • On delete or unpublish. The item's vectors are removed straight away, so retired content stops being cited.
  • Daily. A scheduled sweep re-indexes anything modified since the last run, catching changes made outside the normal editor flow.
  • On demand. Sync content now in settings forces a full pass.

Chunk identifiers are derived from the post they came from, so when a post gets shorter the now-orphaned chunks are deleted rather than left behind to be quoted months later.

Sync depends on WP-Cron

The initial bulk index and the daily sweep both run through WP-Cron in small batches to avoid PHP timeouts. On a quiet site with no traffic, WP-Cron may not fire. If content is not appearing, that is the first thing to check.

Rebuilding from scratch

Rebuild index in settings deletes the vector index, recreates it, and re-indexes everything. It is destructive and asks you to confirm. You need it in two situations: if you change the embedding model (the index and the model are welded together by dimension, so a change invalidates everything already stored), or if the index has ended up in a state you no longer trust.

How much can you index?

Vectorize's free allocation is generous but finite, and at 768 dimensions it works out to a few thousand chunks, comfortably enough for most sites, and roughly a few thousand pages worth of typical content. See limits and costs for what happens as you approach it.