FREQUENTLY ASKED QUESTIONS ABOUT AI

A large language model (LLM) is a type of AI model trained on vast quantities of text data, often hundreds of billions of words, to learn statistical patterns in language. LLMs can generate coherent text, answer questions, summarize documents, translate languages, write code, and perform a wide range of other language tasks. Examples include OpenAI GPT-4o, Anthropic Claude, Google Gemini, and Meta Llama. LLMs are the foundation of most modern generative AI applications, including AI assistants, content generation platforms, and enterprise AI agents.

A large language model (LLM) is trained on massive datasets using billions to trillions of parameters, giving it broad general-purpose capabilities across many domains and languages. A small language model (SLM), such as Microsoft Phi, Google Gemma, or Mistral 7B, uses far fewer parameters (typically fewer than 10 billion) and is designed for specific tasks or domains. SLMs are faster, cheaper to run, and can be fine-tuned or deployed on-device, making them well suited for enterprise use cases where cost efficiency, data privacy, and latency matter. The trade-off is that SLMs may not perform as well as LLMs on complex, open-ended reasoning tasks.

A token is the basic unit of text that an AI model processes. Tokens are not the same as words; they are subword fragments determined by the model's tokenizer. In English, one token is roughly equivalent to four characters or three-quarters of a word. The sentence "Vertesia helps teams manage content intelligently" contains approximately eleven tokens. Tokens matter because AI APIs are priced per token (both input tokens and output tokens), and the number of tokens in your prompt and response directly affects cost and latency.

A context window is the maximum amount of text (measured in tokens) that an AI model can process in a single interaction. Everything the model "sees" (the system prompt, the conversation history, retrieved documents, and the user's question) must fit within the context window. Common context window sizes range from 8,000 tokens for older models to 1 million tokens or more for newer frontier models like Gemini 3 Pro. A larger context window allows the model to reference more information at once, which is critical for tasks such as analyzing long contracts, processing large codebases, or reasoning across multiple documents.

A prompt is any text input sent to an AI model to elicit a response. Prompts can be as simple as a single question or as complex as multi-page instructions that include role definitions, contextual information, examples, formatting requirements, and constraints. The quality of a prompt has a direct and significant impact on the quality of the model's output. This practice of deliberately crafting inputs to improve AI responses is known as prompt engineering.

A prompt template is a reusable, parameterized structure for constructing prompts programmatically. Instead of writing a new prompt from scratch for every request, a prompt template defines a consistent structure with placeholders that are filled in at runtime with specific values, such as a document to analyze, a user's name, or a product category. Prompt templates are essential for building reliable, scalable AI applications because they ensure consistency, make prompts easier to version and test, and allow non-technical users to interact with AI systems without needing to understand the underlying model. Vertesia provides a native prompt template management system that lets teams create, version, test, and deploy templates across multiple AI models.

Fine-tuning is the process of taking a pre-trained AI model and further training it on a smaller, domain-specific dataset so that it learns the vocabulary, style, and knowledge of a particular field or organization. A fine-tuned model is better suited for specialized tasks than a general-purpose model. For example, a law firm might fine-tune a model on legal documents so it responds with appropriate legal terminology. Fine-tuning requires significant data preparation effort and compute resources, but it can dramatically improve model accuracy for specific use cases.

A foundation model (also called a base model) is a large AI model trained on a broad, general-purpose dataset. Foundation models serve as the starting point for most AI applications. A fine-tuned model is a foundation model that has been further trained on a narrower, domain-specific dataset to adapt its behavior for a particular use case. Think of a foundation model as a highly educated generalist and a fine-tuned model as that same generalist after a specialized residency in a specific field.

Model inference is the process of running a trained AI model to generate predictions or responses based on new input. When you send a prompt to an LLM and receive an answer, you are performing inference. Inference is distinct from training (which is the process of teaching the model). For enterprise teams, inference cost and latency are key operational considerations, particularly at scale. Vertesia routes inference requests intelligently across model providers to optimize for cost, speed, and quality depending on the task.

Retrieval-augmented generation (RAG) is an AI architecture that combines a language model with a retrieval system. Instead of relying solely on knowledge the model learned during training, RAG first retrieves relevant documents or data from an external knowledge base and then passes that retrieved information to the model as context. This allows the model to answer questions based on up-to-date, organization-specific information rather than generic training data. RAG is the most widely adopted technique for building enterprise AI applications because it improves accuracy, reduces hallucinations, and keeps the model's knowledge current without requiring expensive retraining. Vertesia's platform is built natively around RAG workflows.

Hallucination refers to when an AI model generates information that sounds plausible but is factually incorrect, fabricated, or unsupported by the provided context. This happens because language models are trained to produce statistically likely text, not to verify facts. Hallucinations are one of the most significant risks in enterprise AI deployments. Mitigation strategies include retrieval-augmented generation (RAG), which grounds the model's responses in verified source documents, and output validation pipelines that check answers against known facts before they are surfaced to end users.

Model evaluation is the practice of systematically testing an AI model's outputs against a defined set of criteria, such as accuracy, relevance, groundedness, tone, and safety, to determine whether the model is performing as expected. Without rigorous evaluation, teams cannot reliably deploy AI in production or confidently compare different models. Vertesia includes built-in evaluation tools that allow teams to run test suites against multiple models simultaneously, track performance over time, and select the best model for each task based on objective results rather than intuition.

Multi-modal AI refers to models that can process and generate multiple types of data, such as text, images, audio, video, and structured data, within a single model. GPT-5 and Gemini 3 Pro are examples of multi-modal models. Multi-modal capabilities are increasingly important for enterprise content workflows because most organizational knowledge exists in diverse formats, including PDFs with embedded charts, scanned documents, audio recordings, and video. Vertesia supports multi-modal content ingestion and processing across its content intelligence pipeline.

Multi-model AI gives enterprise applications the flexibility to select and switch between different Large Language Models (LLMs) rather than relying on a single provider. Because AI models vary significantly in speed, intelligence, cost, and specialization, multi-model support ensures you can route each specific task to the optimal model. It eliminates vendor lock-in, optimizes operational performance and cost, and future-proofs your AI investment as new, more capable models hit the market.

Vector search (also called semantic search or approximate nearest-neighbor search) is a search method that represents documents and queries as high-dimensional numerical vectors (called embeddings) and finds the most similar vectors using mathematical distance metrics. Unlike traditional keyword search, which matches exact words, vector search understands meaning and context, so it can surface relevant results even when the query uses different terminology than the document. For example, a vector search for "ways to reduce staff costs" can return documents about "workforce optimization" and "headcount reduction" without those exact words appearing in the query. Vector search is the backbone of modern enterprise AI search and RAG systems.

Embeddings are numerical representations of text, images, or other data in a high-dimensional vector space, where items with similar meaning or content are placed close together. When you generate an embedding for the phrase "revenue growth strategies," the resulting vector is mathematically close to embeddings for "ways to increase sales" or "tactics to boost income" because they share semantic meaning. Embeddings are created by specialized embedding models and are stored in a vector database for fast retrieval. The quality of embeddings directly affects the accuracy and relevance of AI search results.

A vector database is a purpose-built data store optimized for storing, indexing, and querying high-dimensional vector embeddings at scale. Unlike relational databases that store structured rows and columns, vector databases are designed to find the most semantically similar vectors to a given query vector extremely quickly, even across millions or billions of records. Vector databases are a core infrastructure component of any RAG or semantic search system. Vertesia uses MongoDB as the database of the Vertesia platform.

A knowledge graph is a structured representation of information that maps entities (such as people, products, organizations, and concepts) and the relationships between them. Unlike a relational database that stores data in flat tables, a knowledge graph captures the connections between data points in a way that reflects real-world relationships. For example, a knowledge graph for a pharmaceutical company might connect a drug compound to the diseases it treats, the clinical trials that tested it, the researchers who conducted those trials, and the patents that protect it. AI systems can traverse knowledge graphs to answer complex questions that would be impossible to answer with keyword or vector search alone.

Keyword search matches documents based on exact or near-exact word matches between the query and the indexed content. It is fast and deterministic but brittle. It fails when users phrase queries differently from how documents are written. Semantic search uses embeddings to capture the meaning behind both the query and the documents, enabling it to surface relevant results regardless of the specific words used. Most modern enterprise AI search systems combine both approaches (a technique called hybrid search) to get the precision of keyword matching and the contextual intelligence of semantic search.

Hybrid search combines keyword-based retrieval with vector-based semantic search to produce more accurate and relevant results than either method alone. The scores from both systems are merged (a process called reciprocal rank fusion (RRF) or score normalization) and the top-ranked results are passed to a re-ranking model that further orders them by relevance before they are returned to the user or fed into an LLM. Hybrid search is considered the current best practice for enterprise RAG systems because it handles both precise lookups (e.g., product codes, names) and fuzzy conceptual queries equally well.

A re-ranker (also called a cross-encoder) is a model that takes a query and a set of candidate documents retrieved by an initial search and re-scores them based on their relevance to the query. Unlike bi-encoder embedding models that compare pre-computed vectors, re-rankers process the query and each document together, allowing them to make much more nuanced relevance judgments. Adding a re-ranker to a search or RAG pipeline typically improves answer quality significantly, at the cost of some additional latency. Vertesia's search pipeline includes configurable re-ranking to balance quality and performance requirements.

Chunking is the process of splitting large documents into smaller segments before generating embeddings and indexing them in a vector database. Because embedding models have token limits (typically 512 to 8,192 tokens), and because similarity search works best on focused, coherent units of text, the way you chunk documents has a major impact on search quality. Common chunking strategies include fixed-size chunking, sentence-level chunking, paragraph-level chunking, and semantic chunking (which uses an AI model to identify natural topic boundaries). Vertesia's content ingestion pipeline applies semantic chunking strategies tailored to each document type for the most accurate results.

An embedding model encodes text into a vector representation in a single forward pass, optimized for fast retrieval across large corpora. It does not compare a query to each document directly. A reranking model (cross-encoder) takes both the query and a specific candidate document as a combined input and produces a relevance score, which allows for much richer comparison at the cost of being slower and more computationally expensive. In a typical production pipeline, an embedding model handles the initial broad retrieval (top 50-100 candidates) and a reranking model refines the final ordering (top 5-10 results) before they are used in generation.

An AI agent is a software system in which an LLM is given a goal and a set of tools, and autonomously decides how to use those tools (in what sequence, with what inputs) to achieve the goal. Unlike a simple chatbot that responds to a single turn, an agent can plan multi-step workflows, call external APIs, read and write files, query databases, trigger actions in third-party systems, and self-correct when it encounters errors. AI agents are the foundation of autonomous AI applications that go beyond question-answering into actual business process execution.

Agent tools (also called function calls or actions) are specific capabilities that an AI agent can invoke during its reasoning process. Tools can include things like: searching a knowledge base, querying a database, calling a REST API, sending an email, creating or updating a document, running a code snippet, browsing the web, or triggering a workflow in a business system. The agent's LLM decides which tool to call, with what parameters, and in what order, based on the goal it has been given. The quality and breadth of an agent's tool set directly determines what tasks it can perform autonomously. Vertesia provides a rich, extensible tool ecosystem for enterprise AI agents, including native integrations with content repositories, CRMs, ERPs, and search infrastructure.

Agent orchestration refers to the coordination of multiple AI agents (or of an AI agent across multiple steps, tools, and conditional branches) to accomplish complex goals. Orchestration handles task routing (which agent or tool handles which subtask), state management (keeping track of what has been done and what remains), error handling (what to do when a tool fails or returns unexpected results), and human-in-the-loop checkpoints (when to pause and ask a human for approval). Vertesia's native agent runtime provide the infrastructure that makes multi-agent, multi-step workflows reliable enough for production deployment.

A multi-agent system is an architecture in which multiple specialized AI agents collaborate, each handling a specific domain or task, under the coordination of an orchestrator. For example, a contract review workflow might involve a legal compliance agent, a risk assessment agent, a summarization agent, and a redlining agent, all managed by an orchestrator that assembles their outputs into a final deliverable. Multi-agent systems enable parallel work streams (different agents working simultaneously), specialization (each agent is optimized for its task), and scalability (adding more agents to handle additional domains or volume). Vertesia's agent runtime is designed for enterprise multi-agent deployments.

Agent observability is the practice of capturing and analyzing detailed records of how an AI agent behaves: every decision it makes, every tool it calls, the inputs and outputs of each step, the time taken, the tokens consumed, and any errors encountered. Without observability, AI agents are black boxes that are difficult to debug, audit, or improve. Robust observability includes structured logging, distributed tracing (following a request across multiple agents and tools), metrics (latency, cost, error rates), and evaluation (whether the agent's final output was correct and aligned with the goal). Vertesia includes built-in observability for all agent runs, giving teams the visibility they need to operate AI reliably in production.

A reactive agent responds to explicit user inputs or triggered events. It waits for a prompt and then acts. A proactive agent (also called an autonomous or goal-directed agent) operates on longer time horizons, taking initiative based on a defined goal or schedule without needing a human to prompt each step. For example, a reactive agent might answer "summarize this document" when asked; a proactive agent might monitor a content repository overnight, identify newly published documents relevant to a defined topic, summarize them, and push a digest to stakeholders, all without human intervention. Vertesia supports both modes, allowing teams to deploy agents that operate reactively in user-facing workflows and proactively in background automation pipelines.

Function calling (also called tool use or tool call) is a capability built into many modern LLMs that allows the model to output a structured request to invoke a specific function or API rather than generating a plain text response. The calling application intercepts the function call, executes the requested action (e.g., querying a database or calling an API), and returns the result to the model, which then continues reasoning. Function calling is what enables AI agents to interact with the real world in a structured, predictable way. It is the mechanism that connects the model's reasoning to actual system actions.

In an agent context, the system prompt is the foundational set of instructions that defines the agent's identity, capabilities, behavioral guardrails, available tools, and operating rules. The system prompt is sent to the LLM at the beginning of every conversation or task run, and it shapes how the agent interprets user inputs and decides which actions to take. A well-designed system prompt is one of the most important factors in agent reliability and safety. Vertesia's agent builder allows teams to create, version, test, and deploy system prompts across multiple agent configurations with full audit trails.

Human-in-the-loop (HITL) refers to a design pattern where an AI agent pauses at defined checkpoints and requests human review or approval before continuing. HITL is essential for high-stakes tasks, such as approving a contract, sending a financial transaction, or publishing content, where the cost of an AI error is too high to allow fully autonomous execution. HITL balances the efficiency of automation with the accountability of human oversight. Vertesia's workflow engine supports configurable HITL checkpoints, with routing to appropriate reviewers based on task type, risk level, and organizational role.

Generative AI refers broadly to AI systems that generate new content (text, images, code, audio) in response to prompts. Agentic AI is a more specific and advanced category in which AI systems not only generate content but also take actions, use tools, make multi-step decisions, and operate autonomously over time to achieve goals. All agentic AI uses generative AI at its core (typically an LLM for reasoning), but not all generative AI is agentic. The shift from generative to agentic AI represents one of the most significant transitions in enterprise software, from AI as a creative assistant to AI as an autonomous collaborator.

Content intelligence is the application of AI and machine learning to automatically analyze, classify, enrich, and extract insights from content and documents at scale. Rather than treating content as static files, content intelligence transforms unstructured text, images, and documents into structured, searchable, and actionable knowledge. For enterprises with large content repositories (knowledge bases, product catalogs, legal libraries, marketing assets, etc.) content intelligence is the foundation for building AI search, AI agents, and personalized content experiences. Vertesia applies AI to every document that flows through it, extracting entities, generating summaries, classifying topics, and producing semantic embeddings automatically.

Document preparation (also called content conditioning or data preparation) is the process of transforming raw documents into a format that AI models can effectively process and reason over. It includes steps such as

• extracting text from PDFs, Word documents, and HTML pages
• normalizing formatting and encoding; removing noise (headers, footers, boilerplate)
• segmenting content into appropriate chunks
• enriching content with metadata (author, date, topic, language)
• generating embeddings
• indexing content in a vector or full-text search store

The quality of document preparation directly determines the quality of AI outputs (garbage in, garbage out). Vertesia automates the entire document preparation pipeline, handling 150+ document formats out of the box.

Unstructured data is content that does not have a predefined, machine-readable schema, including documents, emails, presentations, PDFs, images, audio files, and web pages. Approximately 80-90% of enterprise data is unstructured. Traditional databases and analytics tools struggle with unstructured data because they are designed for tabular, structured formats. AI models excel at interpreting unstructured content, but they require careful ingestion and preparation to do so reliably at enterprise scale. Vertesia specializes in transforming unstructured enterprise content into AI-ready knowledge assets.

Metadata enrichment is the process of automatically adding structured descriptive information, such as topic tags, entities (people, places, organizations), sentiment, language, document type, reading level, and custom taxonomy labels, to content, using AI models. Rich metadata dramatically improves search relevance, content discoverability, governance compliance, and the quality of AI-generated responses, because it gives retrieval systems and agents more precise signals for finding and contextualizing content. Vertesia applies AI-generated metadata enrichment to every document ingested into its platform, making content immediately more useful for both humans and AI agents.

A content type is a structured schema that defines the properties, metadata fields, and behaviors associated with a specific category of content, such as a contract, a product specification, a press release, or a support ticket. Content types enforce consistency across large repositories, making it easier to search, govern, and apply AI workflows to specific categories of content. In Vertesia, content types are fully configurable with custom schemas, and AI models can automatically classify documents into their correct content types during ingestion.

Automated document ingestion is the process of importing documents from source systems, such as SharePoint, Dropbox, Google Drive, Salesforce, databases, or web URLs, into an AI platform for processing. A robust ingestion pipeline handles format conversion, text extraction, metadata extraction, chunking, embedding generation, and indexing. The speed, fidelity, and coverage of document ingestion directly affect how quickly an organization can make its content AI-ready. Vertesia's ingestion pipeline is designed for enterprise scale, capable of ingesting 100+ million documents per day.

Content generation refers to using AI to create new content from scratch, such as drafting a blog post, writing a product description, or generating a sales email. Content intelligence, by contrast, focuses on AI's ability to understand, analyze, classify, and extract value from existing content, such as summarizing a contract, classifying documents by topic, or identifying compliance risks in a policy document. Both capabilities are valuable, and the most sophisticated enterprise AI platforms support both. Vertesia offers content intelligence to help organizations unlock the knowledge trapped in their existing content while also enabling AI-powered content generation workflows.

A content workflow is an automated sequence of steps that content goes through from creation or ingestion to final use or publication, with AI applied at specific stages to add value. For example, an AI-powered content workflow for a legal team might automatically classify incoming contracts, extract key clauses, flag unusual terms, generate a summary, route to the appropriate reviewer, and update a case management system, all without manual intervention. Vertesia's workflow engine allows teams to design, deploy, and monitor AI content workflows visually, with support for conditional branching, human-in-the-loop steps, and integration with external systems.

A semantic layer is an abstraction that maps raw content and data to a consistent set of business concepts, entities, and relationships that are meaningful to end users and AI systems alike. For content, a semantic layer might normalize how products, customers, geographies, and topics are referenced across thousands of documents so that an AI agent searching for "Q3 revenue for the APAC region" understands that documents referring to "third-quarter sales in Asia-Pacific" are relevant, regardless of terminology differences. Building a semantic layer on top of enterprise content is one of the most impactful investments an organization can make to improve the quality and reliability of AI-powered knowledge management.