Communication modes

The mechanism for determining communication modes employs adaptive switching between three dialogue styles:

• “talk” – naturalness, irony, empathy, and emotional openness,
• “creative” – the ability to create rich, complete narratives,
• “expert” – an analytical and synthesizing mind, the ability to work with complex concepts.

Each mode has its own language model parameter settings (temperature, top_p, max_tokens) and is accompanied by a specific mode prompt. The mechanism analyzes each message from the interlocutor and selects the optimal mode for the response. Mode determination works on the principle of weighted consensus: the decision is made based on two independent sources – the current message and the accumulated interaction history.

As a result, Chimera maintains dialogue in an optimal style – flexible enough to respond to changes and stable enough to maintain communicative coherence.

Architectural features

The architecture is built on a dual-loop evaluation system. Each loop is an independent analysis method with its own logic and confidence metrics.

Loop 1: Linguistic analysis

This loop uses a morphological analyzer to scan text. It identifies expert patterns (prevalence of specific terminology, infinitives, and strict constructions) and creative patterns (high density of adjectives and figurative descriptive constructions). The loop takes narrative density into account: long messages with rich morphological structure are biased toward the “creative” mode, while structured lists and instructions are routed to the “expert” mode. The “talk” mode is the default and is used if the above patterns are not detected.

Loop 2: Historical profile (partner model)

The partner model is a dynamically refined representation of the interlocutor and the dynamics of the relationship with them (communication habits). It analyzes previous interactions and forms a profile of preferences. The result of each current mode determination is recorded in the profile, refining the preference statistics for future interactions. This is not a strict directive, but a recommendation mechanism. Its influence varies depending on the degree of confidence in the recommendation.

Decision-making principle

Mode determination follows a priority logic, where different sources of information have different weights depending on the situation.

1. Combined analysis – the decision is made based on the weighted sum of signals from the linguistic loop and the partner model.
2. Priority of relevance – if sources conflict, explicit markers of the current message take precedence over historical trends. This models natural communicative logic: what is said now takes priority over what was said in the past.
3. Fallback strategy – in the absence of clear signals, heuristics are activated: analysis of the conversational nature of the text or preservation of the previous mode.

Confidence is calculated dynamically and reflects how well-grounded the decision is. High confidence arises when sources agree or when strong markers are present.

Conceptual principles of operation

Dynamic adaptation. The mechanism is not tied to a single method. It constantly balances between responding to the current context (“what is being said now”) and taking historical patterns into account (“how past interactions went”). This avoids both mechanical repetition and random switching.

Temporal stability. If the mode remains stable across several consecutive interactions, confidence in its continuation grows. This creates communicative inertia, similar to the natural flow of dialogue, where the style of communication does not change without reason.

Functional analogies: