Language Models (LLM)
What is an LLM?
A Language Model (Large Language Model or LLM) is an Artificial Intelligence system trained on immense volumes of textual data to understand and generate natural language. These models constitute the main engine for chatbots and LLM apps created with Wikit Semantics.
How LLMs Work
LLMs operate on a principle of contextual prediction: by analyzing a sequence of text (the prompt), they generate a coherent continuation by drawing on the knowledge acquired during their training. This capability allows them to:
- Understand user questions and requests
- Generate relevant and natural responses
- Perform various linguistic tasks (summarization, translation, analysis, etc.)
- Adapt to the desired context and style
Using LLMs in LLM Apps
Language models are the central engine of LLM apps created with Wikit Semantics. They are used in an orchestrated manner to transform natural language interactions into concrete and relevant actions. Within an LLM app, the LLM fulfills several essential functions: it analyzes and understands user intentions, generates natural and contextualized responses, responds to small talk, ... The Wikit Semantics platform automatically optimizes the prompts sent to the LLM based on context, business knowledge (data sources), and defined objectives (app type). This intelligent use of LLMs allows for the creation of fluid and relevant conversational experiences, while maintaining control over the generated responses.
The combination of LLMs with the platform's capabilities makes it possible to build robust LLM apps tailored to business needs.
LLMs available in Wikit Semantics
Wikit Semantics offers a selection of high-performance language models suitable for different needs – for example, Azure OpenAI GPT-4o-mini or Mistral Small.
The "Models" tab in the "Organization Settings" section lists the available LLMs. To activate other models, contact your CSM.
Limitations and Considerations
Although very powerful, language models (LLMs) have certain limitations that are important to consider when using them. They can sometimes generate "hallucinations," meaning they produce inaccurate or invented information, particularly when they lack context or reliable data. Their understanding of the world is limited to their training period, and they do not have real-time knowledge. LLMs can also exhibit biases inherent in their training data, which requires particular vigilance in certain usage contexts. Furthermore, their "black box" operation can make it difficult to precisely explain their decisions. Usage costs, especially for the most powerful models, must be taken into account in solution design. This is why Wikit Semantics integrates control and validation mechanisms that help limit these risks: source verification, contextual enrichment, and fine-tuning of interactions. Thoughtful use of LLMs, combined with these safeguards, allows for making the most of this technology while managing its limitations.