AI Settings

Search accuracy and long conversations

Understand how Promptor finds the right content for each question, how to get more accurate answers, and how conversation memory works.

Good answers depend on Promptor finding the right parts of your content before the AI writes a reply. This page explains how that search works, what you can do to improve it, and how Promptor handles long conversations.

How Promptor finds your content

When you sync a knowledge base, Promptor splits each page, post, product or file into passages of up to about 1,000 characters. A Q&A entry is kept as one passage with its question and answer. It then turns each passage into an embedding: a numeric "fingerprint" of its meaning, created with OpenAI. This is semantic search, so a question matches content with the same meaning even when it uses different words.

For each visitor question, Promptor:

  1. translates the question into your site's language, if the visitor writes in another language
  2. creates an embedding of the question
  3. compares it with every passage in the knowledge base the chat uses
  4. sends the most relevant passages (up to 10) to the AI along with the question
  5. asks the AI to answer in the visitor's language and to recommend matching services, products, articles and FAQs

If no passage is relevant enough, the AI still answers, but without your content to rely on, and the question is marked Unanswered under AI Conversations. That view is the quickest way to find gaps in your content. See Review unanswered questions. Short greetings such as "hi" or vague one- or two-word questions get a friendly reply that asks the visitor what they are looking for, without an AI call.

Get more accurate answers

Most accuracy problems come from the content, not the settings. Check these first:

  • Is the content in the knowledge base? Open Knowledge Bases › Manage, click Manage on the card and check the item is selected. See Choose content and set content roles.
  • Was it re-indexed after the last change? Selected pages are re-indexed automatically when you update them. If the text changed without the page being saved in WordPress, for example after an import, click Clear Index and then Sync to rebuild the index. See Sync, re-index and clear a knowledge base.
  • Does the chat use the right knowledge base? Each shortcode points to one knowledge base, for example [promptor_search context="support"]. See Shortcode reference.
  • Is the answer written down in plain text? Promptor reads the text of your pages. Text inside images, or content loaded later by JavaScript, is not indexed. Spell out prices, hours and policies as plain text.
  • Are the content roles right? Pages marked Service become service suggestions with a quote form; Product items become product cards; FAQ and Blog Post content is used for answers and linked as related content.
  • Is the page focused? A page about one topic matches better than a long page about everything. A short FAQ page with one question per heading often helps the most.
  • Is the answer missing altogether? Write it as a Q&A entry. A Q&A entry is one short passage, so it matches its question very well. See Add Q&A entries.

For more fixes, see Answers are wrong or miss your content.

The Similarity Threshold setting

Settings › AI Behavior has a Semantic Search Engine card with a Similarity Threshold slider.

SettingWhat it doesDefault
Similarity ThresholdThe screen describes it as how semantically similar a content chunk must be to the question to be used. Range 0.5–0.9.0.7

We recommend leaving it at the default and improving the content instead, using the checklist above. In 2.1.0 the slider doesn't change the search: Promptor uses the most relevant passages that pass its own built-in relevance check, whatever the slider shows.

Long conversations

Conversation memory

With Conversation Memory on, the AI sees the earlier messages of the conversation, so a visitor can ask "how much does it cost?" after talking about a service. You find it in Settings › UI Settings. It is on by default.

SettingWhat it doesDefault
Conversation MemoryWhen enabled, the AI uses the recent messages of the conversation as context for better responses. Note: this increases API token usage.On

When memory is off, each question is answered on its own.

Limits that keep long conversations under control:

  • Promptor sends at most the last 20 messages to the AI.
  • Each message is cut to 4,000 characters when it is sent to the AI.
  • Visitors see "Your message is too long…" when they type more than 5,000 characters.

Where the conversation is kept

The visitor's conversation is stored in their own browser for the current browser session, up to the last 15 messages, together with any services they selected. It survives moving between pages and closes with the browser tab. Visitors can clear it with the new-conversation button in the chat header. See What visitors see in the chat.

Conversation summarization

The Conversation Management card on Settings › AI Behavior also contains summarization settings:

SettingWhat it doesDefault
Enable Conversation SummarizationThe screen describes it as automatically summarizing long conversations to save tokens.Off
Summarization ThresholdNumber of messages (user + AI). Minimum 2.6
Summarization PromptThe system prompt used to ask the AI to create a summary."Briefly summarize the key points of the preceding conversation."

In 2.1.0 these three settings are saved but don't change how conversations are handled, so we recommend keeping the defaults. To limit token use in long conversations, use usage limits or turn off Conversation Memory.

Still need help?

Our support team answers every ticket. Tell us your WordPress version, Promptor version and what you see.

Read the FAQSee what changed in each version

Open a support ticket

All help topics