Skip to main content

Artificial Intelligence (AI)

AI refers to computer systems that can perform tasks that usually require human intelligence. They are trained on large amounts of data, learn patterns, and make decisions based on this information.

Context window

The context window is the maximum amount of text that can be processed at once. It is measured in tokens, which are word pieces the model uses to read and generate text. Depending on the model, context windows range from 128k to over 1 million tokens — that’s roughly 100 to 750+ pages of text. You can find a list of the context window sizes here. The context window gets filled up with your prompt, attached documents, and the previous messages in a chat (read more about this here). When the context window is full, some information from the chat or the documents are cut out, which reduces the answer quality and speed.
For the best performance, please open a new chat to regularly clear the context window

Data analysis

Data analysis is the process of working with tabular files such as CSV, Excel, and Google Sheets in chat. Langdock can inspect rows and columns, calculate metrics, create charts, and generate files. We describe this in more detail here.

Embedding

There are two ways how the models work with documents. One way is to send the entire text to the AI model with the prompt in the context window. The other one is to only send relevant sections in the context window. The second option is used for very long documents or a large number of files, for example in Knowledge bases. To find the relevant sections, the uploaded documents are sliced into many smaller sections. These sections receive a vector, a long number that describes the topic of the section. This section is then called embedding. The selection process, which sections are relevant is an embedding search (see next section). Later, the embeddings or the sections behind the embeddings are then put into the context window. An embedding search identifies different embeddings that are relevant for a specific prompt. If you ask about a specific topic, the embedding search checks which sections in attached documents are connected to this topic and attaches them to the prompt and sends it to the model. The relevance is checked with vectors, which is a mathematical way of depicting a specific topic (see section about vectors below).

Hallucination

Large Language Models are trained on data. They use this data to predict the most likely next token (or word piece) to generate a response. In some cases, for example, when the model does not have the correct context in the prompt or attached documents, it can rely too much on the training data to answer the prompt. Then, it generates a response with the training knowledge, whereas this might be incorrect in the situation you are currently in. Hallucination is made-up information by the model. Unfortunately, this behavior is (still) a part of how AI models work. Some models hallucinate less than others, but overall you should always check the output for correctness.

Large Language Model (LLM)

Large Language Models are AI models designed to understand and generate human language. They are trained on large amounts of text and information, which allows them to understand the context and answer based on it. Here is an article describing the basics of AI models.

Machine learning (ML)

Machine learning is a subset of AI where the performance of tasks is improving over time because the AI is learning how to better perform a task.

Prompt

A prompt is the message you as the user send to the AI model. It should prompt the model to perform a specific task, answer a question, or respond in any way. The best way to prompt is not only in individual words (like a Google search) but rather complete, natural sentences. You can learn more about prompting in our prompt engineering guide.

Retrieval-Augmented Generation (RAG)

To enhance the model capabilities beyond the training knowledge, AI tools use RAG. RAG helps to add information to the model. This can be adding information by attaching a document in the context window with the full text. One other form, which is often meant with “RAG,” is using the embedding search (see above) to search a vector database and attach the relevant vectors to the prompt.

Token

Tokens are word pieces used for text generation. The models generate the response token by token, so they start with the first word piece to write the next piece, then the next piece, etc. The longer a response is, the more tokens are used for generating the answer. There is a maximum amount of tokens that can be processed by the model (see context window above) and also a maximum limit of how many tokens a model can generate.

Tool calling / function calling

Large Language Models can use different tools to perform an operation. These tools are not part of the model, but a feature built by an engineer, which the model can communicate with. Some tools are the web search tool or the data analysis tool.

Vector

A vector is a mathematical object in a multi-dimensional space. It is a series of numbers and is assigned to a specific embedding or text section. The numbers describe the topic of the section in multiple categories. The vectors then can have a specific distance from each other. For example, a vector for a section about pizza is not as close to a vector of a section about artificial intelligence as a vector of a section about computers.

Vector database

A vector database stores uploaded files as vectors so embedding search can find relevant sections for a query. Knowledge bases and text documents in synced integration folders use vector search in the background. Synced spreadsheets use complete file reading instead. You can also connect your own vector database to an Agent.

Vision

Vision is the capability of AI models to process and analyze images. Not all models are capable of this. The models that can process images have this shown in the app when you select the model. They are also listed here.