Voiceflow Adds Transcript keyword search for Faster AI Agent Debugging.

Voiceflow is giving teams a faster way to investigate what their AI agents say, and why. The company has added Transcript keyword search alongside knowledge base logs for evaluations, two observability upgrades aimed at reducing the manual work involved in maintaining production agents.

The update, announced on LinkedIn, addresses a routine but costly problem. Once an agent is handling a meaningful volume of customer conversations, finding a specific failure can mean opening transcripts individually or relying on broad metadata filters. Even after locating the conversation, a team may struggle to determine whether the agent consulted the right company information before producing its answer.

Voiceflow’s new tools divide that investigation into two steps: search identifies the relevant conversations, while logs expose what happened inside the agent.

Transcript keyword search lets users narrow their transcript list using phrases that appeared in the conversation itself. Searches can be scoped to messages from the customer or messages generated by the agent, providing a more focused view of each side of the interaction.

That distinction is useful in practice. A support team could search customer messages for a product name, cancellation request or recurring error phrase. It could instead search agent messages to find cases where an outdated policy, incorrect instruction or undesirable response appeared.

The feature should make several common maintenance tasks more manageable:

• Finding conversations connected to a reported customer problem
• Checking how often customers use particular wording
• Locating transcripts in which an agent gave a known incorrect response
• Separating customer demand signals from the agent’s own language
• Building a targeted set of conversations for evaluation and review

Voiceflow describes the feature as phrase-based filtering that narrows the transcript list to conversations containing the specified language. The announcement does not say whether it supports semantic matching, fuzzy matches or advanced search operators, so teams should not assume it can automatically group differently worded requests.

The second change adds knowledge base logs to an evaluation’s Log Visibility settings. When enabled, the model performing the evaluation can see the knowledge base searches carried out by the agent during a conversation.

Previously available visibility covered playbook, workflow and tool logs. Adding knowledge base activity provides another layer of evidence, particularly when an evaluator needs to judge whether an answer was grounded in approved content. It can show that the agent searched the knowledge base, reveal the searches it attempted and help teams distinguish a retrieval problem from a response-generation problem.

For example, an agent might produce an incomplete answer because it formed a poor search query, because the relevant material was absent from the knowledge base, or because it failed to use retrieved information correctly. Exposing search activity gives the evaluating model more context for separating those possibilities. It also complements Voiceflow’s earlier work on query rewriting for ambiguous knowledge base searches, which focused on improving retrieval before an answer is generated.

The practical impact is less about adding another analytics screen and more about shortening the debugging loop. A team can search for a problematic phrase, open the relevant transcripts, then run or inspect evaluations with knowledge base activity available to the evaluator. That creates a clearer path from a customer complaint to the underlying agent behavior.

There are limits to what has been disclosed. Voiceflow has not specified plan requirements, pricing implications or a staged rollout schedule in the announcement. It also has not claimed that the new logs prove an answer is correct. They provide evidence about what the agent searched, but teams still need evaluation criteria and accurate source material to judge the resulting response.

What is Voiceflow Transcript keyword search?

Voiceflow Transcript keyword search is an observability feature that filters agent transcripts according to phrases used in a conversation. Users can scope a search to customer messages or agent messages to isolate relevant interactions more quickly.

What do knowledge base logs add to Voiceflow evaluations?

Knowledge base logs let the model performing an evaluation see the searches an agent ran against its knowledge base. They join playbook, workflow and tool logs in the evaluation’s Log Visibility settings, helping teams assess whether an answer was connected to their supplied content.

Which Voiceflow plan includes these observability upgrades?

Voiceflow’s announcement did not identify plan restrictions or additional pricing for either feature. Teams should check the platform or Voiceflow’s official changelog for current availability details.