Conversation AI Query Detection in Sparkforge

By Sparkforge Support Team ·Last reviewed May 3, 2026

Conversation AI now reviews the last five messages of a thread on every request, scores the query, and pulls from your sub-account's bot training to respond accurately. No setup required.

What changed

Conversation AI got an upgraded training flow. The previous version leaned on suggestion flow and the LLM's tool calls, which sometimes misread the customer's intent. The new flow reads recent conversation context and adapts to your sub-account data.

How it works

Conversation analysis: Each AI request reviews the last 5 messages to decide if the contact wants support or has a business question (pricing, services, hours).
Query extraction: The flow pulls the actual question out of the thread and assigns an analysis score.
Training data lookup: Pulls the matching answer from your sub-account's bot training so responses fit your business.

Flow steps

Automated analysis runs on every inbound message.
The query analysis score reflects whether the contact needs support or is asking a buying question.
The bot answers using the most relevant training entry for that query.

What you need to do

Nothing. The change is server-side. If you've already trained your bot, the new flow uses that data. If your responses still feel off, refresh your training data and check that pricing, services, and FAQs are current.

Frequently Asked Questions

Do I need to retrain my bot to get this update?

No. The new query detection flow runs automatically on every Conversation AI request. Existing training data is used as-is.

How many messages does the AI look at?

The last 5 messages in the thread. That window gives enough context to identify the real question without dragging in old, unrelated topics.

What is the analysis score?

It's a confidence indicator that flags whether the contact needs human support or is asking a business question the bot can answer from training.

My bot still gives wrong answers — what next?

Refresh your training data, audit crawled website content, and add explicit Q&A pairs for the questions that miss. The flow only answers as well as the data it has.