Conversations

See the chats your agents had with real users, and ask Voxli AI about them.

What a conversation is

A conversation is a chat one of your agents had with a real user. It isn’t a test: there is no tester, no assertion and no score. You find them under Conversations, in the Observability group of the left sidebar.

A conversation reaches Voxli in one of two ways:

• Imported. Your team sends the chat’s transcript through the API. It shows what the user saw.
• From traces. When no transcript was sent, Voxli builds the chat from its traces, the records of what your agent did in each turn. Traces record what your agent’s code did, not what the user received, so a turn can differ from what the user saw. If the transcript is imported later, it replaces the one built from traces.

Your developers set up both. See Import Conversations and Send Traces.

The conversation list

The Conversations page lists the newest conversations first. Each row shows:

• Conversation - the first thing the user wrote, or the chat’s id when the user wrote nothing.
• Source - Imported, or From traces when Voxli built the chat from its traces.
• Messages - how many messages the user and your agent sent.
• Started - when the chat started.

Click Show more at the bottom of the list to load more. Click a row to open the conversation.

A conversation

The conversation page shows how many messages the user and your agent sent, how many tools your agent called (when it called any), how long the chat lasted (when the messages have times) and when it started.

Below that, the Transcript shows the chat in order, laid out like a test’s conversation in Results. Each tool call your agent made shows its arguments and what the tool returned. A transcript built from traces says so, since a turn can differ from what the user saw.

The Details section beside the chat shows:

• Source - Imported or From traces.
• Chat ID - your agent’s own id for the chat.

Any values your team sent along with the chat, such as its channel, show in a Metadata section below it.

Connecting Langfuse or LangSmith

If your agent’s traces are in Langfuse or LangSmith, connect it so Voxli can read them.

1. Click your workspace name in the top left corner, choose Settings, then open Integrations.
2. Click Add integration and choose Langfuse or LangSmith.
3. Fill in the connection and click Connect. Voxli checks the keys before it saves them.
   • Langfuse: the Host, then the project’s Public key and Secret key. The host is filled in for Langfuse Cloud EU. For the US region use https://us.cloud.langfuse.com, or enter your own host.
   • LangSmith: the API URL and an API key. The URL is filled in for LangSmith US. For the EU region use https://eu.api.smith.langchain.com. Fill in Workspace ID only for a key that works across several workspaces.

Click a connection’s row to see whether it still works, Test it, or replace its keys. You can connect more than one Langfuse project or LangSmith workspace.

A connection doesn’t copy every chat into Voxli. Voxli reads it one chat at a time, by the chat’s id: the Langfuse session or LangSmith thread with the same id. It does this:

• About a minute after each import. Voxli reads the chat’s traces from every connection. Importing the chat again reads them again, so a longer chat gets its newer turns.
• When Voxli AI looks up a chat. If Voxli has no traces for that chat, it reads them from your connections and keeps them. A chat that only your traces know then shows up in the list, marked From traces.

Asking Voxli AI about a conversation

Open a conversation, then click the Voxli AI button (the sparkles icon) in the top bar. The panel links the conversation, so you can ask about it right away.

Voxli AI reads the messages and, for each turn, its traces: the model calls, the tool calls and their results, the knowledge lookups and the guardrails. It can explain why your agent replied the way it did, or tell you when the traces don’t show it. A conversation without traces can still be read, but the reasons behind its replies weren’t recorded.

In any Voxli AI chat, you can also ask it to:

• Find conversations by what was said, or by date.
• Look up a conversation by any id you have for the chat: the chat id, a Langfuse session or LangSmith thread id, or the conversation’s id in Voxli.
• Replay a conversation against one of your agents, sending the user’s messages in order and comparing the replies with the original.
• Turn a conversation into a test that sends the user’s messages in order, with assertions on what your agent should have done.

Replaying a conversation and writing tests are work Voxli AI does in the background. It says what it will do and waits for your go-ahead, unless you already asked for exactly that. A conversation doesn’t record which agent had it, so Voxli AI asks which agent when that isn’t clear.

What’s next

• Agents - the agents Voxli AI can replay a conversation against.
• Assertions - what a test built from a conversation checks.
• Send Traces - how your developers get traces into Voxli.