Uniphore Help Center Portal

A2A

Agent-to-Agent (A2A) refers to the direct communication and interaction between two or more Artificial Intelligence (AI) agents. This is a key part of collaborative AI, where agents work together to solve a problem that no single agent could solve. It's the digital equivalent of a team of experts collaborating on a project.

Accuracy

Accuracy measures how accurately an AI model's output matches the target. It's a key metric for evaluating an AI's performance. For a chatbot, high accuracy means its answers are consistently correct and relevant to the questions asked. Low accuracy means the model often provides incorrect or irrelevant information.

AI Agent

An AI agent is an autonomous entity that can perceive its environment, make decisions, and take actions to achieve a specific goal. An agent is more than a simple program; it's designed to be proactive and intelligent, using tools and reasoning to complete complex tasks. For example, a customer service agent might not only answer questions but also open a ticket, review customer history, and follow up.

Agentic Runtime

An agentic runtime is the environment or platform where an AI agent executes its tasks. It provides the necessary resources, tools, and connections to external systems that the agent requires to operate effectively. It’s the engine that powers the agent’s decision-making and actions.

Agentic Workflow

An agentic workflow is a process or series of steps performed by one or more AI agents. Unlike a traditional, static workflow, it is dynamic and can adapt to new information and changing circumstances. This makes it ideal for complex, unpredictable tasks.

Agentic Search

An agentic search is an advanced information retrieval technique where autonomous AI agents actively navigate a Knowledge Graph to construct comprehensive answers. Unlike traditional search or basic Retrieval-Augmented Generation (RAG) that simply returns matching text chunks, agentic search involves agents dynamically deciding which connected information paths to follow and then building comprehensive answers.

Agent Distillation

A fine-tuning process that transfers a high-capability model's problem-solving behavior into a smaller, faster model optimized for a specific agent's tasks. The system observes how the larger model handles the agent's real-world inputs, collects those interactions as traces, and uses them to train a lightweight model that replicates that behavior. The result is reduced latency and cost while preserving output quality.

Agent Evaluation

An agent evaluation is the process of testing and measuring an AI agent's performance. This goes beyond simple accuracy and includes metrics such as how efficiently it completes tasks, how effectively it uses tools, and how well it handles unexpected situations. It’s about ensuring the agent is not just smart but also reliable and effective.

Agent OPA

An agent OPA (Open Policy Agent) is a system that enforces policy and security rules within an agentic workflow. It acts as a gatekeeper, ensuring that agents perform only authorized actions. This is a critical component for security and compliance, especially in multi-agent workflows.

BPMN

Business Process Model and Notation (BPMN) is a standardized graphical notation for modeling and documenting business processes. It's a flowchart-like notation that is easy for both humans and computers to understand. You use BPMN to map out the steps of a process, making it a perfect tool for designing and orchestrating complex multi-agent workflows.

Co-related RAG

Co-related Retrieval-Augmented Generation (RAG) is a strategy that goes beyond simple information retrieval. It's designed to identify and use not only the requested information but also any other data that is logically correlated with the query. For example, if you ask about a specific product, it might also retrieve related products or customer feedback that helps provide a more comprehensive answer. It's about building a richer context for the AI's response.

Cost of GPU

The cost of a Graphics Processing Unit (GPU) refers to both the initial purchase price of the hardware (Such as a high-end NVIDIA A100 or H100) and the ongoing operational costs, including electricity and cooling. These costs are a major factor in the overall expense of training and running AI models.

Cost of LLM/SLM

The cost of Large Language Models (LLMs)/ Small Language Models (SLMs) is primarily based on the computational resources required to run them. LLMs are very expensive due to their size, which requires powerful GPUs and significant energy consumption. SLMs are more economical to operate because they are smaller and faster, making them a more cost-effective choice for many business applications.

CP Server

A Control Plane (CP) Server is a central hub in a multi-agent system that manages and orchestrates the activities of different agents. It acts as a conductor, assigning tasks, monitoring progress, and ensuring that agents work together towards a common goal. It's the brain behind the entire operation.

Data Agents

A data agent is an AI that can autonomously perform a series of tasks to find, process, and analyze data. It acts like a digital assistant that understands a request (e.g., Find me all customer reviews about the product from the last month) and then executes the necessary steps to get the answer. It can connect to different data sources, run analysis, and present a summary all on its own.

Data Chunking

Data chunking is the process of breaking large unstructured data, such as long documents or videos, into smaller, more manageable chunks. This is especially important for AI systems, such as LLMs, that have a limit on the amount of information they can process at once. By chunking data, you make it easier for the AI to handle and get specific answers.

Data Enrichment

Data enrichment is the process of improving a dataset by adding information from other sources. For example, you might take a list of customer names and email addresses and enrich it by adding demographic data, purchase history, or social media profiles. This provides a more comprehensive view of your customers, which benefits targeted marketing and deeper insights.

Data Graphs

Data graphs are a way to visualize data as a network of interconnected entities and relationships. Unlike a traditional spreadsheet, which uses rows and columns, a data graph illustrates the relationships between different pieces of information. Think of it like a mind map where each idea is a node, and the lines connecting them show how they are linked. This is great for understanding complex connections, like how different customers and products are related in a sales network.

Data Partitioning

Data partitioning is the process of dividing a large dataset into smaller, more manageable pieces. This is typically done to improve the performance and speed of data processing and analysis. For example, a massive customer database might be partitioned by region, with each region's data stored separately. This enables faster retrieval of information for a specific region, eliminating the need to sift through the entire dataset.

Data Preparation

Data preparation is the process of cleaning, transforming, and organizing raw data into a structured format suitable for analysis or AI model training. This foundational step ensures data quality, consistency, and relevance by handling tasks such as removing duplicates, standardizing formats, enriching datasets, and partitioning data into manageable segments.

Data Splitting

Data splitting is a specific type of partitioning used in machine learning. You divide your dataset into at least two parts: a training set and a testing set. The training set is used to teach the AI model, while the testing set evaluates how well it performs on new, unseen data. This is crucial to ensuring the model isn't just memorizing the data it was trained on.

Fine-tuning

Fine-tuning is the process of further training a pre-trained LLM on a smaller, specialized dataset to improve its performance on a particular task. For example, you might fine-tune a general LLM on your company's product manuals to make it an expert on your products. This makes the model more accurate and useful for a specific purpose without requiring a new model to be trained from scratch.

Golden Dataset (Q&A)

A Golden Dataset is a high-quality set of questions and answers used to test and evaluate an AI model's performance. It's considered the gold standard because the answers are known to be correct and reliable. By running your AI model against this dataset, you can accurately measure its accuracy and ensure it's providing the right information.

GPU (A100, H100)

A  Graphics Processing Unit (GPU) is a specialized electronic circuit designed to rapidly process images and graphics. In the context of AI, GPUs like the NVIDIA A100 and H100 are essential because their parallel processing power is ideal for the massive, complex calculations required to train and run large AI models. They are the workhorses of modern AI, enabling rapid handling of vast amounts of data and performing complex calculations.

Graph DB

A Graph Database (DB) is a type of database that stores data in a network of interconnected nodes (entities) and edges (relationships). Unlike a traditional database with rows and columns, a graph DB is designed to show how different pieces of data are connected. This is extremely useful for applications such as social networks, fraud detection, or recommendation engines, where the relationships between data are more important than the data itself. For example, it can easily show who a person's friends are, and who their friends' friends are.

Graph RAG

Graph RAG is an advanced version of RAG that utilizes a graph database to store and retrieve information. Instead of just searching for chunks of text, it uses relationships between data points (graph) to find more relevant, contextually rich information. This is particularly useful for complex questions that require connecting multiple facts.

Guardrail

A guardrail is a set of rules and policies designed to regulate an AI model's behavior. It prevents the model from generating harmful, inappropriate, or off-topic content. Guardrails can enforce safety standards, ensure compliance with brand guidelines, and keep the AI focused on its intended purpose.

Human In the Loop (HITL)

HITL (Human-in-the-Loop) is an AI development approach that involves human interaction and collaboration. The system automates certain tasks, but a human expert remains in the loop to review, validate, and refine the AI's output. This is particularly important for high-stakes decisions, as it ensures accuracy and accountability. For example, a human might review an AI's medical diagnosis before it's finalized.

Inference

Inference is the process by which a trained AI model uses new data to make a prediction or decision. Think of it as the answer part of the AI process. For example, if you trained a model to recognize cats, inference is the moment it sees a new picture and says, "That's a cat!" It's where the AI's learning is put into practice.

Inferencing Server

An inferencing server is a dedicated, powerful computer optimized to run AI models for inference. Instead of a general-purpose computer, these servers are built to handle the heavy computational load required to make predictions quickly and efficiently. Imagine a specialized machine in a warehouse whose sole purpose is to examine photos all day and identify objects much faster and more reliably than a regular computer.

Knowledge Base (KB)

An enterprise-grade repository that stores and organizes organizational content for AI-powered operations. It uses knowledge representation to structure information, enabling AI systems to understand, reason through, and apply it to solve business problems.

Knowledge Graph

A Knowledge Graph is a structured semantic network that represents real-world entities such as concepts, events, or objects, and the complex relationships between them. Instead of storing data in flat, isolated tables, a knowledge graph links related data points together, much like a web of interconnected nodes. This structure provides deep context and meaning, making it an ideal foundation for AI agents to traverse information, understand how different pieces of data relate to one another, and draw logical conclusions.

Knowledge Representation

Knowledge representation is the process of structuring information so an AI system can understand and use it to solve problems. It's about taking human knowledge, such as facts and relationships, and presenting it in a format that a computer can process. This can be achieved through tools such as rules, logical statements, or graphs. The goal is to provide the AI with a clear and organized understanding of the world.

Latency

Latency is the time it takes for a system to respond to a request. In the context of AI, it’s the delay between submitting a query and receiving the complete answer. Lower latency means faster response times, which are crucial for real-time applications such as chatbots and automated customer service.

Lexical/Symbolic

Lexical/Symbolic refers to the traditional, rule-based approach to AI, in which the system understands and processes information using explicit symbols and rules. This is like a computer following a detailed instruction manual. It focuses on the meaning of words and symbols themselves rather than on their context or underlying patterns. For example, a system might be programmed with the rule: "If the word is 'hot', and the word is 'dog', it refers to a food item". It's very precise, but can't handle ambiguity.

LLM

A Large Language Model (LLM) is an AI model trained on massive amounts of text and data to understand, generate, and interact with human language. These powerful models underpin technologies like Open GPT, which can write emails, summarize documents, and answer complex questions.

LLM as a Judge

LLM as a judge is a method for evaluating the quality of an AI's output. Instead of using a human or a fixed set of rules, you use a more powerful, well-regarded LLM to rate the responses of another, typically smaller, LLM. This provides a fast, scalable, and often very effective way to assess the quality of an AI's answers.

LORA

LORA (Low-Rank Adaptation) is a popular Parameter-Efficient Fine-Tuning (PEFT) technique. It works by freezing most of the LLM's original weights and adding a small number of new, trainable parameters. This makes fine-tuning much more efficient, allowing you to quickly and cheaply adapt a large model to a new task without the massive overhead of traditional fine-tuning.

Markdown Language

Markdown is a simple way to format plain text. It uses special characters, like asterisks (*) or hashes (#), to make text bold or italic, or to create headings. It's not a full-fledged programming language; it's just a quick and easy way to add formatting to documents, emails, or notes without using complicated software. Think of it as a shortcut for making text look good.

MCP

Model Context Protocol (MCP) is a standardized protocol that allows AI models to securely connect with external data sources, tools, and services. MCP enables AI systems to access real-time information and perform actions beyond their built-in capabilities.

Model Drift

Model drift happens when an AI model's performance deteriorates over time. This is usually because the real-world data it's processing changes, and the model's original training data is no longer representative. For example, a fraud detection model trained on historical data might become less accurate as new types of fraud emerge. It’s a key reason why AI models need continuous monitoring and retraining.

Multi-agent Orchestration

Multi-agent orchestration is the process of coordinating and managing the activities of multiple AI agents working together to solve a complex problem. This involves assigning tasks, managing communication between agents, and ensuring their efforts are synchronized. It's the project management for a team of AI agents.

Overfitting

Overfitting is a problem that occurs when an AI model learns the training data too well, including its noise and random fluctuations. As a result, it performs exceptionally well on the data it was trained on but poorly on new, unseen data. Think of it as a student who memorizes a test without truly understanding the subject. They do well on that specific test but fail another one.

PEFT

Parameter-Efficient Fine-Tuning (PEFT) is a family of techniques that allows you to fine-tune an LLM without modifying all of its parameters. This is a significant benefit because it drastically reduces the memory and computational resources required for fine-tuning, making the process faster and more cost-effective. It's like only adjusting a few knobs on a radio instead of rebuilding the entire thing.

Prompt

A prompt is the input or instruction you give to an AI model to get it to generate a response. It can be a question, a command, or a piece of text. The quality and clarity of the prompt are crucial, as a well-written prompt yields better, more relevant responses from the AI.

Prompt Optimization

An agent improvement process that automatically refines an agent's prompt based on the results of an Agent Evaluation experiment. The system analyzes evaluation traces and generates a recommended prompt update, which you review and either approve, edit, or reject before it takes effect. Approved changes replace the agent's existing prompt; rejected changes leave it unchanged. Prompt Optimization should be run before Agent SLM Fine-Tuning, as a well-optimized prompt produces higher-quality traces for the fine-tuning process.

QLORA

QLORA is an even more advanced version of LORA. It combines the LORA technique with quantization, reducing the LLM's parameter precision (its memory footprint). This allows you to fine-tune very large models on consumer-grade hardware, making advanced AI techniques more accessible and affordable.

Quantization

Quantization is a technique for reducing the size and memory footprint of an AI model by storing its parameters with fewer bits. For example, it might convert a 32-bit number into an 8-bit number. This makes the model smaller, faster, and more efficient to run, especially on devices with limited memory.

RAG

Retrieval-Augmented Generation (RAG) is an AI technique that combines a search engine-like retrieval system with an LLM. Instead of relying solely on the LLM's internal knowledge, RAG first finds relevant information from a trusted data source (e.g., a company's internal documents) and then uses it to formulate its answer. This makes the LLM's responses more accurate, up-to-date, and grounded in facts, thereby reducing the likelihood of hallucination or fabrications.

RAG Evaluation

RAG evaluation is the process of testing and measuring the performance of a RAG system. This involves using a golden dataset of known-good questions and answers to assess how well the RAG system retrieves the correct information and generates accurate responses. This helps ensure the system is reliable and trustworthy.

Reasoning

Reasoning is the ability of an AI model to think through a problem, apply logic, and arrive at a conclusion. Instead of just repeating information it has been trained on, an AI that reasons can solve new problems by breaking them down into steps, evaluating different options, and making a logical choice. It's what enables an AI to transition from simply knowing facts to applying those facts to solve a complex puzzle.

Self-learning

Self-learning refers to an AI system's ability to improve its performance and adapt without human intervention. This can be achieved through various methods, such as reinforcement learning, where the model learns by receiving rewards or penalties for its actions. It allows an AI to continuously get better at a task by learning from experience.

Semi-structured Data

Semi-structured data is a mix of both. It doesn't fit into a rigid, relational database table but contains organizational tags or markers that make it easier to process than unstructured data. This data has some internal structure, but it's not a strict row-and-column format. A good example is an XML file or a JSON document, which uses tags to define elements. Think of it as a set of neatly labeled folders, but without the need for a traditional filing cabinet.

Skill

A skill is a specialized ability of an AI agent, often created by linking a prompt to a specific tool or function. For example, a weather agent might have a weather-lookup skill that uses a tool to retrieve weather data. Skills are what make agents versatile and able to perform a variety of tasks efficiently.

SLM

A Small Language Model (SLM) is a language model that is much smaller than an LLM. While it's less powerful with a smaller knowledge base, it offers key advantages: faster performance, lower operational costs, easier fine-tuning, and the ability to be specifically trained for niche tasks with high accuracy. SLMs are ideal for applications that require quick, efficient, focused AI that doesn't require the extensive knowledge of an LLM, such as customer service for a specific product line.

SLM Evaluation

SLM evaluation is the process of measuring an SLM's performance and effectiveness. Unlike LLMs, SLMs are often evaluated on their ability to perform specific, narrow tasks with high accuracy and speed. This ensures the model is not only fast and cost-effective but also reliable for its intended purpose.

Structured Data

Structured data is data that has a predefined, organized format, such as a spreadsheet or a database table. It's neatly arranged into rows and columns, making it easy for computers to read and analyze. Examples include customer names, addresses, and sales figures. Think of it as data that fits perfectly into a table.

Token

A token is the basic unit of text that an LLM understands. It can be a word, a part of a word, or even a punctuation mark. When you give an LLM a prompt, it first breaks down the text into these tokens. The LLM then processes each token in turn to generate a response. The cost and speed of an LLM are often measured by the number of tokens processed.

Tool/Function call

A tool/function call is the mechanism that allows an AI model or agent to interact with external systems and resources. For example, an agent might call a function to search the internet, look up a record in a database, or send an email. This extends the AI's capabilities beyond its internal knowledge, enabling it to perform real-world actions.

Unstructured Data

Unstructured data is information that doesn't have a pre-defined format. It's the opposite of structured data and is often harder for traditional computers to process. Examples include text documents, emails, social media posts, images, and audio files. Imagine a pile of loose notes, photos, and voice recordings that's called unstructured data.

Vector Database

A vector database is a specialized type of database designed to store and search for data based on its meaning, rather than just keywords. It stores data as vectors, which are numerical representations of items such as text, images, or audio. This allows it to quickly find items that are Semantically Similar, even if they don't use the exact same words. This is crucial for tasks such as those of a chatbot that require understanding the intent behind a question.

Vectorization

Vectorization is the process of converting data (text, images, or audio) into numerical representations called vectors. This is a fundamental step for many AI applications, as computers can only process numbers. For example, a sentence like "I love this product" is converted into a list of numbers that represents its meaning, allowing an AI to compare it numerically with other sentences. It's the key to making data understandable to AI.

Visual LLM

A Visual LLM is a type of AI that can understand images in addition to text. While a standard LLM only processes language, a Visual LLM can take an image as input and answer questions about its content, describe what's in it, or even generate new images. It's an AI that can read and see the world, making it much more powerful for tasks like analyzing charts or recognizing objects.

Workflow

A workflow is a set of interdependent tasks or processes that work in tandem to achieve a common business goal. This series of tasks assigned within a workflow has preset, user-defined rules and conditions. The processes are linear and can be completed sequentially or in parallel.