Data

The Data page of a chatbot names it and decides what it knows. Everything the chatbot knows reaches the model in its instruction, the system message sent with every visitor message. You build that instruction either with the wizard or by hand from variables.

The General card holds the chatbot’s Name and Delete chatbot, which removes the chatbot, its settings and any files it synchronized.

The wizard

Use the wizard is on for a new chatbot. The wizard lists the site’s post types; switch on each one the chatbot should know about and choose what to Include of its posts: Title, Content, and Post meta where the post type has meta fields. Current user gives the chatbot the signed-in visitor’s display name, email and roles, so it can greet them by name.

System instruction is your own text, placed before what the wizard generates. Use it for the chatbot’s role and tone:

You are a helpful assistant for our website. Be friendly and concise.
If you don't know the answer, say so.

Generated instruction shows the complete instruction the chatbot will use, read-only. The wizard builds it from variables: one query variable per post type, named query_ and the post type, and a currentUser variable. Each post type’s template gives every post its public URL and its type beside what you include, so the chatbot can link the posts it recommends. Switching the wizard off keeps those variables and the instruction, so you can start from them by hand.

Variables

With the wizard off, Variables lists the chatbot’s variables and Instruction holds its instruction. A variable is a named value that replaces {name} in the instruction when a visitor sends a message. Select Add variable, choose its Source, and give it a name in Variable:

You are an assistant for the page {pageTitle}.
The visitor's name is {visitorName}.

The sources are:

Plugins can add their own sources; see Custom sources.

For the post, taxonomy, term and user sources, enter the ID to read, or switch on Use the current value from the page to read the post, term or visitor of the page the chatbot is on. Meta key picks one value out of the result with dot notation; for example, data.display_name on the User source gives the user’s display name. Without a meta key, the whole result is sent as JSON; for the User source that is the display name, email and roles, and a key such as data.user_login reads any other field. Below a variable with an ID or a query, Preview shows the value it currently resolves to.

Query variables

A query variable lists posts. Post types shows the post types it reads, which the wizard sets; a query you add yourself reads posts. Search value limits it to posts that match a search, and Match all terms requires a post to match every term the query names rather than any. Raw query arguments replaces all of that with your own WP_Query arguments as JSON:

{ "post_type": "product", "posts_per_page": 20, "orderby": "date" }

Raw arguments run as written, so set posts_per_page: without it, WordPress returns five posts.

Template is rendered once for every post, one per line. Write post fields in curly brackets, such as {ID}, {post_title}, {post_excerpt} or {post_content}, and reach into the post’s meta with a dot, as in {post_meta.price}. {permalink} is the post’s public URL; {guid} is the GUID WordPress stored when the post was created, which is not always a working address:

Product: {post_title}
URL: {permalink}
Price: {post_meta.price}

Without raw arguments a query has no limit, so a large post type makes a long instruction, and every visitor message pays for it.

Synchronizing to OpenAI file search

For large queries, Enable file synchronization uploads the rendered posts to OpenAI’s file search instead of putting them in the instruction. The model then looks up what it needs for each question. It needs an OpenAI model on the chatbot’s Model page; with any other model the switch is disabled and says so. A synchronization switched on before the model changed can still be switched off.

Synchronization interval sets how often the file is refreshed: every minute, every 10 minutes, hourly, daily, weekly or monthly. Saving the chatbot uploads the file at once and schedules the refreshes, which run as the sleek_ai_chatbot_file_sync action in Action Scheduler. wp sleek ai chatbot sync uploads again on demand, and wp sleek ai chatbot sync-status shows whether each file is ready.

Until a file is ready, visitors’ messages are refused with “Synchronized chatbot data is not available yet. Sync the data and try again.” rather than sending the whole query inline, which could cost far more than expected. The editor’s preview uses the inline data instead, so you can try the chatbot while the file uploads.

Retrieval priority

With file synchronization, OpenAI’s file search decides which content a reply reads. Retrieval priority on the Data page lets you weigh that choice instead, so that, for example, your documentation pages win over blog posts that mention the same thing. It needs an OpenAI model and at least one synchronized query; without them the switch is disabled and says what it needs, though a priority switched on before can still be switched off.

With it on, every synchronized post is uploaded as a file of its own, with its title, URL and type, into one search index for the chatbot. For each visitor message SleekAI searches that index itself, with the visitor’s two previous questions as context, and ranks what it finds: relevance first, raised or lowered by each post’s priority. The best eight posts reach the model as reference data, each with its title and URL, so the chatbot links the posts it names.

Each synchronized query gets a priority, Low, Normal or High, or an exact weight from 0.5 to 1.5 beside it; Normal is 1. Below the queries, each post type the queries read can take its own priority, which takes precedence over its query’s; Use query priority leaves it to the query. A priority moves a post past others that are about as relevant, never past a much better match, and never brings in content that does not match at all.

Saving the chatbot rebuilds the index; unchanged posts are not uploaded again. Until the first index is ready, visitors’ messages are refused as with file synchronization, and a failed rebuild keeps the previous index. Turning the priority off keeps the index, which the chatbot then searches through OpenAI’s file search; moving the chatbot to another provider removes it, and moving back builds it again.

The sleek_ai_chatbot_retrieval_weight filter sets the weight of single posts, for example to raise one product page:

add_filter('sleek_ai_chatbot_retrieval_weight', function ($weight, $postId) {
    return $postId === 42 ? 1.5 : $weight;
}, 10, 2);

The hooks reference describes the filter’s arguments.

Custom sources

The sleekAi/chatbot/sources filter adds sources to the list. Each entry has an id, a label and a string value:

add_filter('sleekAi/chatbot/sources', function ($sources) {
    $sources[] = [
        'id' => 'openingHours',
        'label' => 'Opening hours',
        'value' => get_option('my_opening_hours', 'Mon-Fri 9:00-17:00'),
    ];

    return $sources;
});

A variable that uses the source is replaced with its value. SleekAI reads the value while it answers a visitor’s message, in a REST request, so the value cannot depend on the page the visitor is on; use the built-in sources for page data. The hooks reference describes the filter’s arguments.