Your chatbot works, but it doesn’t sound like your product

You find the tone by talking

Defining the role, context and instructions gave us a starting point. Then came extensive use: trying different openings, changing subjects, answering a question, correcting an interpretation and seeing what happened several turns later.

A question can show interest. A chain of questions can make someone feel they are being interviewed. Cutting questions too far can leave the companion sounding flat and the other person carrying the entire conversation.

Humor needs the same attention. A joke can work in one moment and feel uncomfortable in another. Repetition matters too: several acceptable replies, each built from acknowledgment and paraphrase, eventually start to feel lifeless.

For an adult companion, this includes being able to discuss desire, pleasure and personal subjects naturally, without a childish or lecturing tone. That range belongs to the experience we are trying to sustain.

Test what follows the first reply

That use led to checks and tests for specific behaviors. Xiara has tests for receiving an answer without immediately asking another question, respecting a close and carrying a correction through a change of subject.

The sequence matters. If you ask for less repetition and the next reply improves, you still need to see what happens when the conversation moves to music or a personal subject. The adjustment needs to carry through those transitions.

We combine tests of the logic and instructions with scenarios sent to models and review of the responses. Each examines a different part: selecting a memory, preserving an instruction or checking how the whole thing sounds in conversation.

Memory gives the voice depth

Memory handling was crucial. In tests with little context, conversations could feel cold or repetitive. There was less information to recognize what mattered, what had been corrected and where to pick up a thread.

A relevant memory can change the reply: it lets the companion return to an open subject or keep a preference in mind without asking the person to rebuild the context. That continuity gives the conversation depth.

It also requires careful selection. A memory can be accurate and still arrive at the wrong moment. Xiara’s tests cover both retrieving context when someone returns to a subject and leaving it out when the conversation points elsewhere.

Corrections deserve the same care. A preference about how to talk should keep having an effect when the subject changes. Saving it, retrieving it and applying it appropriately are decisions you eventually hear in the response.

Evaluate models inside the experience

The work also involved APIs and specific models. Switching models could change the rhythm, length, language or use of context. We had to return to the conversations that mattered to us.

An internal comparison in July 2026 evaluated models from OpenAI, Mistral and xAI against the same scenarios. The report recorded differences in latency, continuity, language and memory acknowledgments. Decisions were made by task: conversation behavior and a structured music function were reviewed separately.

The evaluation itself kept one limit open: passing automatic checks does not tell you whether a response feels alive, natural or overly coached. That part needs use and editorial judgment.

Keep what you learn and continue adjusting

A correction can resolve one friction and create another. Shorter replies can become dry. More initiative can produce too many questions. Adding context means reviewing when it should enter the conversation.

The cycle was to converse, identify the mismatch, adjust and return to the same cases. Keeping those cases lets us examine whether an improvement holds and whether it affects something that already worked.

Reaching a baseline we felt was solid enough took months of development, testing and use. That is the craft involved: spending time with the replies, noticing what feels wrong and turning it into something we can review again.