Introduction

Large Language Models (LLMs) act as mediators, translating human intentions into tangible actions.

However, this mediation comes up against two problems.

The ability of models to understand and interpret the subtlety of requests on the one hand, and the ability of users themselves to clearly formulate their needs, intentions and objectives on the other.

The Optimized Context Distribution System (SROC) allows us to solve these problems through a multidimensional approach to theIA, considering both the limits of theartificial intelligence and those of humans.

This system optimizes the management of data, content and contexts, thus ensuring efficient distribution of information to specialized artificial intelligences, while providing benchmarks to humans, so that the latter can correctly implement their intentions.

By adjusting the formulation of requests through suggestions of intent, goal, and execution frames, SROC improves not only the precision of requests, but also the relevance of the resulting responses.

By preventing computational fatigue and increasing the attention focus of AIs through dynamic sizing of context windows, SROC promises to improve not only operational performance in various work contexts, but also user experience.

Indeed, the SROC contributes to the latter's skills development during its interactions with AI, by facilitating its access to suitable responses and guiding it towards a fine expression of its needs.

This mediation optimized by the SROC is based on the integrated management of the limiting factors of AI. Factors that must be understood in order to grasp their effects and the optimization levers accepted by the SROC.

Limiting factors

The applicant's formulation

For many humans, accurately articulating expectations of systems as advanced as AI can be a daunting challenge, especially since our egos prevent us from questioning our inability.

In fact, using an LLM seems child's play, because we only use words to issue instructions.

However, each word carries a multitude of parameters that can drastically modify the interpretation of requests and the formulation of responses generated by AI.

Faced with it, we are confronted with the ignorance of our own language and discover our inability to express ourselves with the necessary precision that good communication implies.

Since human-to-human interactions are intuitive, our approximation is never called into question, whereas interactions with AI require perfect accuracy.

Precision, finesse and the smallest nuances are essential to ensure that the results not only meet expectations even if these are poorly defined at the time of the query, but also exceed them.

This is what prompt engineering contributes to, the fundamentals of which frame, without user intervention, each request issued within the framework of the SROC.

But this is not enough, because a phenomenon intrinsic to LLM is at work in each interaction: generalization.

The generalization

Current LLMs, having stored almost all available data, do not suffer from a lack of information that could be filled by prompt engineering or fine-tuning, but suffer from too great a dispersion of their attention.

The problem lies both in the weighting of the interpretation of the queries, and in the weighting of each response.

Interpretation and generation being only the sum of statistical averages on the scale of all the ingested data.

This means that when a LLM model is solicited, even if the latter is correctly prompted and fine-tuned, its level of attention linked to a particular objective will be diluted in all circumstances by weighting effect.

Which induces the need to increase the size of the context windows to cancel this dilution effect by prompt engineering, or to train models on additional, more specific data to improve their behavior, their way of doing things or their relevance.

What AI design allows via SROC, the latter integrating an architecture optimizing the efficiency of any instruction.

But although this is essential, because it helps to reduce the effects of generalization, any request will in any case be confronted with other phenomena and limiting factors which in turn induce a generalist weighting, starting with the factor of conflicts of objectives.

Intrinsic conflicts of goals

The Copywriter example:

The copywriter brings together several skills, each of which brings together other specific skills and disciplines.

Skill 1: Target Identification:

A good copywriter must know how to extract the psychological aspects of a message's potential targets.

This is a skill in its own right, which, to be exploited to its full potential, must be used independently of any other objective, in order to draw on the depth of the task.

Skill 2: AIDA technique (Attention/Interest/Desire/Action)

This second skill, which already conflicts with the first (target identification), thus diluting the focus of attention of the AI, is itself subject to a conflict of attention between four sub-skills:

• Capture theAattention: involves specific skills related to capturing and retaining attention.
• Arouse theIInterest: involves the ability to make logical links between targets and offers aiming at the supply/need adequacy, this involving characterizing the logic of the adequacy links.
• Provoke the Ddesire: involves specific skills to extract elements of the psyche to act, as needed, on frustrations, fears or desires.
• Encourage theAction: involves the use of appropriate techniques aimed at triggering action based on elements of psyche, interest and desire.

Here, exploiting the full potential of each of the skills involved in what the term Copywriter implies, comes down to a conflict of objectives.

The goal is to capture attention, but also to arouse interest, as well as to provoke desire and incite action.

While humans can intuitively differentiate between giving more or less importance to one objective over another depending on the circumstances, AI is simply incapable of doing so by the very nature of its design.

It reduces its interpretation to the sum of the objectives induced by each of the skills and related sub-skills, this by probabilistic assumption, reducing its response to an addition of averages which are themselves sums of other probabilistic averages.

For “Copywriting”, as for any other profession, task, service, this phenomenon takes place and requires a fine segmentation of each element in order to extract the specific objectives so that they do not conflict with each other.

A segmentation which is allowed by the ESP method (Exponential Segmentation Process) that the SROC allows to implement.

But this, once again, is not enough, because the specificity which characterizes relevance comes directly from the arbitration of a scale of priority in the management of these conflicts of objectives.

An arbitration which, if left to the free appreciation of the LLM, can only result in a generality resulting from a sum of other generalities themselves defined by a weighted approach; in itself very far from what the expected relevance could be.

The Arbitration of Relevance

When formulating a response, LLMs arbitrate the characteristics of relevance for us.

This does not only concern the field of competence, but also any data of any type, whether it is information, an instruction, a condition, etc., etc.

However, the arbitration of conflict of objectives inherent in the characterization of relevance, whatever the field of application, falls exclusively within the subjective perception of the applicant.

It is therefore up to the latter to characterize it himself in order to restrict the possibilities of interpretation resulting from a generalization.

This is knowing that the applicant is not expert enough to formulate these characteristics, and knowing that this ultimately comes down to characterizing the criteria of his own subjectivity, which is impossible.

So how can we arbitrate in these conditions?

Well, by relying on benchmarks inspired by human social structures allowing everyone to imagine characteristics of relevance, intuitively admitted by what these benchmarks imply; in this case frames of reference, such as companies, departments and professions.

What the SROC also offers through a socially constructed approach replicating the realities of the world of work.

Neglect of the anthropomorphic reference point

Anthropomorphism is the tendency to attribute human characteristics, behaviors, and emotions to artificial systems, in order to make them more accessible.

In the context of transformative models, we are only dealing with an input field to initiate a discussion, so the anthropomorphic aspect is neglected.

However, the assignment of a job (or department or other standardized framework) by anthropomorphic characteristics also contributes to their accessibility, not by emotional and identification principles, but by the point of reference that job anthropomorphism induces.

This allows a pre-framing of any request, directing the applicant's intentions towards the object of the profession, thereby increasing relevance, without having to arbitrate it according to subjective criteria, because the latter are intuitively included in the notions evoked by the term profession.

In fact, the term “profession” evokes an idea, an unconscious perception of a result that could match our expectations, which, ultimately, we do not know how to express consciously with a level of precision sufficient to be correctly interpreted, but whose unconscious expression is included in the intuition of the object of the profession.

Which allows to express the characteristics of the expected subjective relevance, without even mentioning them.

It is therefore essential not to underestimate the importance of anthropomorphism in the context of transformative models in order to propose user interfaces that structure navigation by elements that are accessible to the user's intuition, so that their choices express their expectations in their place.

To this end, the SROC offers an anthropomorphic approach allowing the virtual replication of the worlds of companies, projects and professions.

But once again, this is not enough, because another phenomenon comes to mitigate the anthropomorphic factor: the computational limits commonly associated with the terms fatigue and laziness of AI.

Computational fatigue

Computational fatigue refers to the decrease in performance of AI systems resulting from overuse, requiring load mitigation.

These episodes of fatigue are unpredictable, frequent (several times a day), can last a few minutes or several hours and can only be identified by the observation of a significant deterioration in relevance.

The uninformed user, when issuing a query during an episode of fatigue, will therefore assume that using AI is inappropriate because the result is mediocre.

The AI ​​could have risen to the challenge five minutes before or five minutes after the request, but not knowing this, the user declares it inadmissible.

Even worse, when he persists in obtaining a relevant result by insisting and repeating himself without obtaining the expected result, which discourages him lastingly.

This frustration is all the greater when the prompt engineering work has been done correctly, and all the limiting factors have been considered, but forgetting to integrate the computational fatigue factor into the edition of the instructions and parameters.

Because computational fatigue does not prevent AI from responding, it prevents it from correctly considering the instructions and all the frameworks that have been given to it, which are supposed to guarantee relevance.

Thus, the greater the efforts made to mitigate the limiting factors, resulting in instruction saturation, the more likely this is to produce an opposite effect during fatigue episodes, because, from then on, the AI ​​is no longer able to consider all of the instructions, and even less to manage intrinsic conflicts.

Leading at the same time to an increase in hallucinations, a phenomenon also consubstantial with the functioning of LLM, which consists of the production of a response at all costs, even if the latter is false.

Hence the need to integrate the management of computational fatigue into any interaction with AI on the one hand, and to approach this management through the prism of saturation effects on the other.

Instruction saturation

Dosing the volume of context not only helps to mitigate the effects of fatigue, but also prevents attention selection bias.

This other phenomenon is characterized by the inability of transformative models to consider the entire context window made available, even at full operating capacity.

Some studies show that about 40% of the context window size is actually considered. Although this percentage is constantly evolving and does not yet have consensus, it sheds light on the lack of attention efficiency on at least half of the window.

More than enough to consider the ability to desaturate on the fly as a fundamental element.

On the one hand, to ensure a maximum level of relevance under the worst possible conditions, thus preventing computational fatigue.

And on the other hand, to ensure that the focus of attention of the AI ​​is established by the limit of selection bias.

This results in a level of systematic relevance, contributing to a resilient organization of AI-assisted work, thus ensuring the continuity of the processes in action.

However, this on-the-fly desaturation which is permitted by the SROC, once again introduces a subjective arbitration falling to humans.

The latter, remaining the sole judge of the relevance expected at the time of the request, must be able to easily arbitrate its desaturation choices.

The orientation of choices

The user must be able to rely on a segmentation previously established by ESP method, by equating computational fatigue episodes and attention selection biases, as the limiting frame of reference for real operational capabilities.

Thus, a fine segmentation, the principles of which are postulated in a pre-established environment, will naturally result in nomenclatures and labels of contexts that the user will be able to intuitively recognize when he has to desaturate a query, identifying at a glance the elements to activate and/or deactivate.

This, via SROC, by adding or removing context elements on the fly, with a simple click, allowing the user to precisely, quickly and intuitively adapt the context to fluctuations in computational fatigue, to attention disorders of the AI, but also to fluctuations in conflicts of objectives.

Let us recall that the increase in the volume of instruction implies an increase in conflicts of objective.

Thus, the effects of on-the-fly desaturation are not limited to filling attention deficits and fatigue problems, but also allow increasing relevance through an attention velocity effect.

Since desaturation is coupled with the finesse of quality prompt engineering, it allows the induction of a large number of notions by the subjective implications inherited from the anthropomorphic orientation, all in a minimum of parameters and instructions.

Which contributes to both reliability and relevance, while preventing out-of-bounds situations.

However, to implement this triple-benefit finesse, it is appropriate to consider the factor of intelligence in its proper measure.

The intelligence factor

The term artificial intelligence induces confusions that lead everyone to believe that the intelligence factor is essential.

However, none of the factors contributing to the relevance of a result involve intelligence of a useful nature within the framework of a transformative model.

If there is indeed a connection between diverse and varied information (remembering that it is systematically smoothed and weighted), bringing out the term intelligence through the capacity to weave links between this information, relevance, itself, is about much more than that.

When we ask an LLM to perform a task, we are not asking for their ability to make connections that we equate with a form of reasoning, no, we are asking for their ability to make the connection between the rough draft of our intentions and the vague idea of ​​our expectations in subjective terms of results.

Which is certainly based on a certain level of intelligence based on the size of the AI ​​model, because this size is largely equated with superior reasoning abilities.

But beyond a certain threshold, the ability to reason no longer has anything to do with relevance, and no longer has a sufficient impact on the quality of the result.

Because, whatever they are, LLMs are subject to the accumulation of the above-mentioned limiting factors, which no technological advance can resolve, especially since the subjective human factor is omnipresent.

Thus, when a relevance that does not admit generalities comes into play, it becomes inappropriate to rest one's expectations solely on the supposed intelligence of a transformative model, as much as on a single lever, such as fine-tuning or prompt engineering.

Because the usefulness of an LLM, its relevance and its reliability are based on the aggregation of multiple factors, each in multiple perspectives, each involving the arbitration of subjective characteristics, the latter requiring dynamic intervention by the human, the sole judge of the relevance of a result in the face of expectations that he himself is incapable of defining with sufficient precision.

All of this, again and again within the computational constraint and under the influence of attention selection biases, each of these constraints being able to have a counterproductive effect in the face of expectations.

This highlights the importance of not confusing “intelligence” with “relevance”, in order to be able to orchestrate and direct the full potential of LLMs in the right direction, namely that of relevance.

Intelligence versus relevance

LLM models, with their vast knowledge, are like omniscient oracles, gifted with clairvoyance, erudite in all things.

However, even an oracle cannot provide relevant answers if the queries do not specify in detail the reduction of the field of interpretation, the desired orientation of the response, the exhaustive context, the nested objectives and the behavioral nuances expected in the management of priorities, themselves variable on subjective criteria that cannot be controlled, impossible to anticipate and even less to characterize without the use of tacitly induced notions.

Since these oracles are perpetually confronted with conflicts (requiring subjective arbitrations) which are both introduced by the user, prompt engineering, fine-tuning and pre-training by unsupervised deep learning cannot in any case be relevant despite their intelligence.

For the expression of this intelligence, outside a framework limited to a clearly defined intention, is no more than a general assertion whose reliability rests on our confidence in this expression.

On the other hand, too much intelligence, based on these generalist modalities and the ability to reason, produces opposite effects.

Indeed, an intelligence of this type can make so many links of possibilities of interpretations for each word composing the instructions, that it becomes impossible to focus attention as desired, so high is the complexity.

By consequently increasing tenfold the intrinsic conflicts of objectives, and drastically complicating the characterization of the expected relevance, intelligence in itself generates more problems than it provides solutions.

Because when we decide to trust him with elements that we do not control, we not only take the risk of taking responsibility for mediocre production.

We also cede the sovereignty of our decisions and erase the specificity of our judgment of relevance in favor of uniformity, this by accepting that subjective criteria of relevance are dictated to us when they escape us.

As long as the problem of computational fatigue is solved by technological advances, just as much as the problem of selection bias, this would in no way resolve the need for desaturation aimed at exacerbating objective priorities rather than others, the relevance of which only the operator can be judged.

So, although transformative models are called artificial intelligence systems, their form of intelligence is of little use in the quest for performance, especially compared to what the relevance obtained by process and consideration of limiting factors can bring.

This is valid for all models based on attention mechanisms, because the latter operate on statistical patterns.

The statistical signature

LLM models are not only forced to generalize due to their limitations in context and computation, but they also do so on erroneous bases, because the conflicts introduced by training or by the user are themselves constituted by a generalist weighting.

Thus, no relevance can be expected from an LLM simply by relying on its intelligence.

A relevant response is only relevant by statistical inadvertence, and not by the expression of a mastered logic contributing to the accomplishment of a requested task.

Because LLM models respond based on statistical patterns, themselves derived from other statistical patterns whose conflicts have been treated by generalization.

Here we see the signature of unsupervised deep learning, essential to increasing intelligence in the general sense, but which prevents any specificity and therefore any relevance.

This accumulation of difficulties leads to a reduction in the depth of the responses and systematically raises the question of relevance in relation to objectives, intentions and expectations.

Thus, in all circumstances, it is appropriate to rely, not on the usefulness of a superior intelligence, but on the limits of the latter, in order to draw out its true potential, the latter being quite sufficient to upset work and societal paradigms.

But not without the help of humans who wish to remain in control of what they produce with the assistance of AI.

The absence of limits

Managing limiting factors and their interdependencies, as well as managing context dosages and anthropomorphism parameters, requires a unified multi-perspective management system, built around these constraints.

This, in an environment that replicates human realities and norms in order to induce the subjective implications necessary for the benchmarks that guide humans in their tasks with AI.

Because, whatever the level of intelligence of a language model (LLM), it is, and will remain incapable of providing perfectly relevant answers, as long as its operation is not framed by predefined limits.

These limits must be dynamically adjustable, allowing to manage the level of attention of the AI ​​on certain objectives rather than others, while allowing to mitigate context saturation if necessary, allowing in all circumstances to obtain a satisfactory match between expectation and result.

It is in compliance with all the conditions mentioned that an LLM can be a relevant mediator in the accomplishment of our intentions, transposed into tangible tasks.

From then on, the intelligence of a model becomes almost secondary, because at this stage of understanding, the criterion of operating costs becomes more important than that of intelligence in the choice of the model.

Because, when the limiting factors are considered at their fair value and are integrated accordingly, they allow to obtain from a supposedly less intelligent model, the equivalent of what a so-called more intelligent model is capable of producing in terms of relevance.

The difference is that a model considered less intelligent costs between 10 and 100 times less for an error rate that is similar, all things considered regarding the problem of hallucination.

The latter boils down to a deficit of information relating to the volume of the model parameters, which can be filled by means of dynamically adjustable segmentation.

SROC Solution

All the issues raised can be summed up in a single obligation: to bring together two interlocutors (human/AI) in a pre-established environment, framed by standards, tacitly accepted implications, objectives whose priorities are underpinned by the execution environment, and based on anthropomorphic markers which induce implicit notions essential to the desaturation and increase of attention focus.

This by integrating the fundamental element that represents the all-round limitation in order to establish a zone of concordance conducive to systematic relevance, admitting the variations resulting from the subjectivity of the applicant.

Definition of SROC

The SROC for Ssystem of Rdistribution Ooptimized Contextes is a content management system and systematic, transversal and multidimensional integration of generative AI in management applications and integrated software packages, based on the categorized management of data, content and contexts dedicated to AI, thus ensuring optimal distribution to specialized anthropomorphic AIs, through an approach considering the intrinsic limits of generative AI, as well as the cognitive limits of human beings.

SROC Objectives

The specific objectives of the SROC include:

1. Optimizing user queries: By offering suggestions of intentions based on the provided contexts and usage frameworks, the SROC helps users formulate more precise queries, thus increasing the relevance of the results.
2. Improving performance: By streamlining the distribution of information to specialized artificial intelligences, SROC aims to maximize operational efficiency and reduce processing errors.
3. Increasing user satisfaction: By ensuring more intuitive interactions that are tailored to user needs, SROC strengthens user satisfaction and engagement in their interactions with AI systems, particularly through the relevance of responses resulting from a streamlined distribution of information to specialized AIs.
4. Adaptability to changing needs: SROC stands out for its ability to adapt to changing needs, effectively responding to user requirements and adapting to technological advances. This ensures the continuity of operational performance in AI-assisted organizations, and reliable thresholds of relevance in the responses provided.
5. Flexibility in choosing models: the SROC allows everyone to choose their transformer model provider, as well as the parameters relating to temperature (Determinism and creativity), to the maximum quantity of the volume of tokens in input, output and/or global.

In short, SROC establishes an environment where artificial intelligence can truly serve users.

On the one hand, through the performance and flexibility levers induced by the exploitation of LLM-type artificial intelligence, and on the other hand, by transforming approximate requests into precise results, up to the demands of the professional world.

Ability to manage limiting factors:

The SROC allows:

• to calibrate on the fly the activation and deactivation of contexts in order to adjust the attention focus of the AI, by decongesting the instruction context of less priority elements according to the task to be accomplished. The choice of the level of relevance is thus adjustable by a desaturation combined with a focus simultaneously induced on a single objective.
• to guide the addressing of requests by department, professions, tasks and subtasks, in order to induce a pre-framing of the request, which by the business and department approach remains intuitive for the user. This while teaching him by the suggestions and orientations inherent in the interfaces, the good approach of requests to adopt to lead him towards an optimal relevance.
• to departmentalize the processing of requests through a socially constructed approach, understood by humans, thus allowing them to refine their choices of AI interlocutor according to the needs of which they are aware, but which they do not know how to formulate with the necessary precision.
• to exploit business segmentations resulting from an extraction ESP translated into disciplines, tasks, and related subtasks. This is done through specialized AI team members accompanying any anthropomorphic AI, each evolving in a broader context. Thus, allowing relevance to be achieved beyond expectations, through the effect of attention velocity on a single task.
• allows to build and assign so-called “global” support teams to specific execution environments, always through segmentation and subdivision. These teams are permanently accessible, regardless of the pages visited, so that hyper-specialization is at hand. This, without having to invoke other contexts and without needing to change environments, the contexts adjusting without human intervention.
• to segment the distribution of contexts with a variable nature, such as appointments, stocks and other moving flows that are impossible to integrate in fine-tuning given their variability. This hyper-segmentation of variable elements is essential, because the influence of selection bias is very strong in these circumstances.
• to partition contexts by content type in order to restrict their understanding and interpretation to the exclusive object implied by content typologies. This is similar to appointments which represent a content type, but in this case of partitioning, concerning other types of content which are predefined and can be refined according to the specific needs of segmentation by type.

The constituent elements of the SROC:

In project mode

The outlook for intention:

Pre-integrated into the SROC are the perspective variations allowing the processing of the following requests:

• Queries whose intents are shared by both the reference frame's intervention context and non-framework interlocutors, invoking shared unified contexts for public front-end and private internal uses.
• Queries whose intentions can only come from users actively participating in the execution framework, as a member with a role, a job or specific task assignments, invoking contexts limited to the connected user, always in the reference framework.
• Queries whose intentions can only be the origin of disconnected users, not belonging to the intervention framework, but interacting with it, invoking the related interpretation contexts, but subject to the orientations of the reference framework of the requested environment.

Global interpretation values:

Global interpretation values ​​are contexts propagated to all AIs operating within a dedicated environment, called a “project”.

These global values ​​integrate editorial alignment and direction.

General disconnected contexts:

Disconnected general contexts allow to adjust the orientation of intentions so as to give an adequate perspective to the task objectives. These contexts are propagated to all AIs that are accessible to the general public, without requiring an account.

Segmentation by content type

• For variable flows:
  • Events
  • Schedule
  • Stock
• For the intention guidelines:
  • Strategies
  • Notes and memo
  • Social Content
  • Editorial content
  • Email Contents
  • Training
  • Frequently Asked Questions

* Non-exhaustive: directly linked to the ability to choose the categories and types of contexts to be considered by the AIs individually or by the general framework called “project”.

In project or personal mode

Personal contexts:

The so-called “personal” contexts are only propagated to the current user. Thus, the elements specific to him are known to all the AIs visited by him, without this affecting the initial parameters of the AIs. The latter receive different parameter elements depending on the users who consult them. Thus allowing the variability necessary for adjusting intentions by explicit preferences.

General contexts:

The so-called “general” contexts are information and/or skills that are imposed globally for any user (connected or not) acting within the dedicated environment. This, regardless of the perspective of the query, always within the limit of the reference framework of global interpretation values. These general contexts are systematically considered.

Anthropomorphic AIs:

In a SROC type management interface, each AI has its own characteristics and parameters, the latter operating within the global reference framework called “project”.

Profile

• Last Name
• Age
• Genre
• job
• Main service
• Description
• Workplace

Expression:

• Quality of writing or speaking
• Style
• Registre
  

Feasibility study

• The hole prompt
• The discussion starter

Perimeter

• Context of intervention
• Audience Type / Audience
• Goals or tasks to be accomplished
• Function role
• Posture in a conflict context

skills

• • Embedded
  • specific

Knowledge

• • Embedded
  • specific

Alignment

• Alignment criteria

Model Setting

• API Provider
• Choice of supplier model
• Fine-tuned model selection
• Input limit
• Output limit
• Max sentence

AI Teammates

Each AI can be accompanied by teammates, each with their own anthropomorphic characteristics, each of these characteristics being guided by the intentions of the so-called main AI, namely the one accompanied by the teammates.

Team members act as employees under the authority of an AI's intentions, within the general framework called "project".

Teammates can be added for independent purposes, or added chronologically, resulting in a process in which the responses of the main AI are intended to be transmitted in whole or in part to the first teammate, whose responses are in turn intended to be transmitted to the next teammate, and so on, allowing for extreme depth in the desired relevance.

The development of this process is based on the ESP method which can be exploited through the so-called Jaris Team.

The Jaris team

The artificial intelligences that make up the Jaris TEAM systematically support all artificial intelligence as architects collaborating on its configuration.

TEAM Jaris AIs also act as employees of the main AI, but incorporate, in addition to the latter's intentions, all of its parameters. Thus allowing Team Jaris AIs to incrementally refine the main AI's parameters.

The support teams:

Support teams are departmentalized AI groupings that incorporate the project's frames of reference as well as the frames of reference of the anthropomorphic AIs that the user uses.

On any other page, the support teams only incorporate the reference frames of the global execution context called “project”.

These teams can be customized by business department, disciplines, and tasks.

This allows the user to compose his own environment according to his most frequent needs, while remaining within the limiting framework of the global environment.

The Incremental Expander

The incremental expander enables deep distribution on the fly by suggesting AI specifically dedicated to particular tasks, promoting focus on objectives induced by the proposal of the right AI interlocutor.

The incremental expander allows from any content to exceed the limits of length and depth by systematically regenerating, on demand, via selection of words, zones, sentences or passages, other queries which are reconstructed in a new context which is that of the selected content or text.

The proposals that are made by opening a dialog box guide the user's choice towards the right AI interlocutor, that is to say the one whose parameters and objectives are concentrated on a single element of intention. This guarantees optimal relevance of the request and the expected response, as thought by the requester.

The AIs requested within the framework of the incremental expander also integrate the general references as well as the references and characteristics of the main AI in use.

This translates into significant gains in terms of relevance, speed and efficiency, while offering an enriched and intuitive user experience.