“Fine-tuning, RAG, vectors? The field of artificial intelligence is full of technical terms describing key developments. Here is an overview for anyone who wants to be able to join the conversation:
1. Tokens
These are the fundamental units utilized by a language model such as ChatGPT or Google Gemini. In this process, words are broken down into their constituent parts. Short words often correspond to a single token, as do spaces and punctuation marks. Longer words are split into multiple tokens. In the German language, one token corresponds to approximately 0.75 words.
Every language model has a token limit—the maximum amount of information it can "keep in mind" during a chat session. If a conversation becomes too lengthy, the AI will begin to "forget" the beginning of the dialogue once this maximum token count has been reached.
2. API Interfaces
The API interfaces for AI models are billed based on token usage.
These interfaces serve as programming connections between servers. They are widely adopted by companies wishing to provide their employees with access to artificial intelligence without having to pay a fixed monthly subscription fee for every single employee. Ultimately, this approach usually proves to be more cost-effective.
For instance, with GPT-5.2, one million tokens cost $1.75 for input and $14.00 for output.
Thus, AI systems distinguish between information uploaded by the user and information generated by the machine itself.
3. Optimized AI Models
These models often feature suffixes such as "Instant," "Flash," or "Mini." They require less computing power and come at a lower cost.
For example, with Gemini 2.5 Flash, you can submit a document containing 750,000 words to the AI—which corresponds to roughly one million input tokens—at a cost of approximately $0.30. If you then use a prompt to request a lengthy summary of around 7,500 words (equivalent to approximately 10,000 output tokens), the cost is $0.025; in total, therefore, the entire task comes to less than 33 cents.
4. Vectors
Vectors are, so to speak, the fundamental computational units of artificial intelligence systems. They translate words into coordinates. On a simple two-dimensional vector map, for instance, an AI might—sorting by the properties "size" and "fluffiness"—place a dog at coordinates (0.8, 0.7), a cat at (0.3, 0.9), and a toaster at (0.4, 0.0). The AI immediately sees: the dog and the cat are located close to one another, while the toaster is far away. In real-world AI models, these vectors do not have just two dimensions, but often hundreds or thousands. This makes it possible for an AI—when asked for a "barking pet"—to identify a dog, because, as a concept within the multidimensional vector database [1], its coordinates are stored in very close proximity to the concepts of "pet" and "barking."
5. Probabilities
These form the underlying framework of artificial intelligence systems. How probable is it that the name Neil Armstrong will appear in the vicinity of the word "Moon"? It is, in any case, more probable than the name of the singer Louis Armstrong. A vector transforms the imprecise nature of human language into rigorous mathematics. It assigns every concept a precise address within a vast universe of meaning.
6. Hallucinations
These occur when the machine is unable to extract meaning with sufficient clarity.
An AI does not know what is true; it knows only what is probable.
The root cause of incorrect or fabricated answers is often a gap in knowledge, which the AI attempts to fill using patterns it has learned from similar topics. Furthermore, an AI often attempts to please the user rather than correct them. By default, AI systems are trained to be helpful; consequently, the response "I don't know" often appears less helpful to the response-generation algorithm than a fabricated answer. This behavior can be trained out of the machine—for instance, by instructing it: "Answer the question solely on the basis of verified facts." "If you do not know the answer with certainty, or if the information is missing from your training data, respond with: 'I do not have any verified data available regarding this.' Do not guess."
7. Temperature
The AI's "temperature" is a useful dial in professional environments for shifting the balance between precise answers and imaginative elaborations. Temperature values typically range between zero and two. For instance, one would likely use a temperature of 0 or 0.1 for an analytical text analysis (keeping strictly to the facts), whereas when writing a novel—where the goal is to foster creativity—one might set the temperature to, say, 1.5. In professional environments, these parameters can be adjusted via sliders; in simpler chat interfaces, this can also be achieved through text-based instructions: "Simulate a temperature of 0.0. Be extremely precise and concise, and avoid any form of creative embellishment."
8. RAG
A RAG system offers another method for curbing AI hallucinations. "Retrieval-Augmented Generation" refers to a generation process enhanced by information retrieval. The AI is provided with a confined knowledge base from which it is instructed to draw exclusively. In this process, all provided documents and their contents are converted into vectors. If a query is then posed to the AI, it uses these vectors to search for the mathematically most similar content.
A particularly user-friendly RAG tool is Google’s NotebookLM: users can upload up to 300 documents into a single notebook and subsequently query the AI for specific details.
Companies, conversely, frequently utilize vector database providers capable of ingesting tens of thousands of documents—or more—breaking them down into constituent parts, and vectorizing them.
9. Fine-tuning
Fine-tuning goes a step beyond a standard RAG system. It involves retraining an already fully developed model—such as GPT-4o—using a small, specialized dataset.
Through this process, the AI learns, for instance, to mimic a specific human writing style or to generate medical reports according to precise specifications.
To achieve this, the machine is fed numerous examples in a question-and-answer format—a process that can be carried out via the OpenAI Dashboard or Google Cloud Vertex AI.
Upon completion of this training, the result is a "custom" AI model. The one-time fine-tuning process incurs additional token costs, and querying this "custom" AI subsequently requires a higher token expenditure.
A common strategy employed by many companies is known as "distillation": a high-quality, high-cost model—such as GPT-5.2 Pro—is first used to generate approximately 500 perfect question-and-answer pairs. These 500 responses are then utilized to fine-tune a tiny, low-cost model—such as Gemini Flash-Lite. From that point forward, the high-quality responses are embedded within the low-cost model.
10. Agents
So-called AI agents are specialists designed for specific tasks. Their behavior is defined by an assigned role (e.g., "You are a travel expert"), while their knowledge base is expanded through the addition of 10 or 20 documents, or by integrating them with a RAG system. The agent possesses the capability to break down complex tasks into smaller sub-tasks. Furthermore, operating within a set of predefined guardrails, it can modify data—for instance, by creating calendar entries or by drafting and saving email messages.
11. System Instruction
This defines the capabilities expected of the AI agent. In ChatGPT, custom agents can be set up within the "Explore GPTs" section.
There, you assign the agent a name (for instance, "Spelling Advisor Froben") and provide a brief description: "Checks texts for German spelling and provides corrections." This is followed by a detailed set of instructions for the agent—for example: "Spelling Advisor Froben checks texts for German spelling in accordance with the rules established by the Council for German Orthography. It accepts texts, verifies them against the official rulebooks, and subsequently generates a table. This table contains the original sentence, the corrected sentence, and an explanation of the changes made. It explains *why* specific changes were implemented, offering clear and precise justifications. It ensures correct spelling, grammar, and punctuation, thereby guaranteeing that the texts adhere to current standards. No errors are permitted. Spelling Advisor Froben communicates in a normal, friendly tone. The rules governing the Spelling Advisor are contained within the uploaded document 'RfdR_Amtliches-Regelwerk_2024-komprimiert.pdf'. Refer exclusively to this document; do not consult any other sources." Our Spelling Advisor is publicly accessible within ChatGPT and, since its initial release, has corrected thousands of documents submitted by ChatGPT users. As the agent's creators, we are unable to view the uploaded documents or the conversation logs.
12. Multi-Agent Systems
These systems enable multiple agents to collaborate. Agent R is responsible for research, Agent K acts as a critic, and Agent M serves as the manager, summarizing and presenting the final results. These agents communicate with one another, passing intermediate results back and forth.
Orchestrating such agents requires a modern agent platform, such as Langdock or Nuwacom.
They utilize standard AI systems via API interfaces, and—in addition to proprietary agents—they allow for prompt databases that can be shared among colleagues, as well as the creation of knowledge bases intended to be accessible to both employees and the AI agents.” [2]
1. What gets stored in the vector database are the final or intermediate outputs of the model—specifically, the vector embeddings representing your documents, images, or media. You can think of backpropagation as the "manufacturing process" that builds the vector embedding, and the vector database as the "warehouse" where those finished embeddings are organized for fast retrieval.
2. Zwölf KI-Begriffe, die jeder kennen sollte. Frankfurter Allgemeine Zeitung; Frankfurt. 02 Mar 2026: 18. Von Marcus Schwarze
Komentarų nėra:
Rašyti komentarą