Build the index

The index lets the assistant find the right paragraph by meaning, even when the visitor uses different words from your document. You build it by clicking Build the index at the top of Answer Desk > Knowledge.

What is the index?

Each document is split into passages (short pieces of text; older versions called them chunks). The index provider turns each passage into a list of numbers that describes its meaning (these are called embeddings or vectors). When a visitor asks something, the question is turned into numbers the same way and compared with your passages. Everything is stored in your own WordPress database: no extra service is needed.

Build it

1. Go to Answer Desk > Knowledge.
2. The Search index card at the top shows the count of passages indexed and how many are still to index, with an estimated cost in dollars. A label next to its title says Complete, To build, or Words only when no index provider is set.
3. Click Build the index. It runs in batches with a progress bar; you can watch it without the page freezing.
4. When it finishes, the box says The index is complete. Every document is searchable by meaning.

The cost is usually a few cents for a whole site. The estimate is only an order of magnitude; the actual bill comes from your provider.

When to build it again

• After you upload or import new documents.
• After you edit an imported page: its text is refreshed automatically, but its new passages need indexing.
• After you change the index provider or index model (see below).

You do not have to remember this. When documents or imported pages change after the index was built, Answer Desk marks the index as out of date: a notice with a Build the index link appears on the Answer Desk screens, and once on the Plugins list (Answer Desk: some documents changed after the last index build), where you can close it until the index changes again, and on Start here the Build the index step loses its tick. Both go away when the index is complete again. The plugin does not rebuild the index by itself, because each build is a paid request to your provider.

Who builds the index

On Answer Desk > AI provider, the Who builds the index card has two settings:

• Index provider: by default The same one that answers, when it can. ChatGPT, Gemini and Mistral can. OpenRouter cannot, so choose Voyage AI, ChatGPT, Gemini or Mistral here (with its key saved in the API keys card). Claude cannot either, but it does not need an index: leave None: Claude reads the documents itself (see below).
• Index model: the recommended model is listed first. Leave it unless you know why you want another.

Claude without an index

With Claude answering and no index provider, there is nothing to build. The Search index card on Knowledge shows Read by Claude, and the Build the index step on Start here counts as done. On AI provider, the How Claude reads your documents setting chooses how Claude reads them:

• Automatic (the default): whole when they are short, by search words when they are long. Short means up to about 33,000 words in all, roughly 80 pages of text.
• Always whole: Claude reads every document with every question. It is the most accurate. Anthropic keeps the documents in its cache for five minutes, and reading them again from there costs a tenth of the normal price, so a busy chat stays cheap. The first question after a quiet spell pays the full price again.
• Always by search words: a small, fast Claude model turns the question into search words, synonyms and translations included, and Claude reads only the passages that contain them. It is the cheapest: each question costs about the same as with an index.

Under the setting you see how many words your documents add up to and how Claude is reading them right now. In Try it on the Overview, a note under each answer says whether Claude read all your documents or which words it searched for.

An index from Voyage AI, ChatGPT, Gemini or Mistral is still worth it for very large archives, where searching by meaning is more precise than searching by words.

Without an index

With Claude, see the section above. With the other providers, if no index provider is set, the assistant still works by matching words instead of meaning, and the Knowledge screen says so. It will miss questions asked in different words, so you should build the index.

Very large archives

The search runs inside WordPress. On a very large index (many thousands of passages), the server may not be able to search all of it in time. When this happens the plugin says so in the Try it results instead of slowing down silently. Remove documents the assistant does not need, or split very long ones.