# Suvegasoft GenAI Glossary — AI & LLM Terminology

> Comprehensive glossary of GenAI, LLM, RAG, and AI terms. Clear definitions for concepts like embeddings, fine-tuning, hallucinations, and more.

- Canonical: https://suvegasoft.co.uk/glossary/

## AI Agents

> Autonomous systems that use LLMs to perceive their environment, make decisions, take actions, and work toward goals, often with access to tools and the ability to plan multi-step workflows.

_Also known as: LLM Agents, Autonomous Agents, Agent Systems_

AI agents go beyond simple question-answering by incorporating perception, reasoning, planning, and action. They can use tools, interact with APIs, and execute complex multi-step workflows autonomously.

### Agent Components

1. **Perception**: Understanding the environment/task
2. **Planning**: Breaking down goals into steps
3. **Action**: Executing steps using tools
4. **Memory**: Maintaining context and learning
5. **Reflection**: Evaluating and improving performance

### Popular Agent Frameworks

- **LangChain**: Python/JS framework for agents
- **AutoGPT**: Autonomous goal-oriented agent
- **BabyAGI**: Task management and execution
- **CrewAI**: Multi-agent collaboration
- **Microsoft Autogen**: Multi-agent conversations

### Common Agent Patterns

- **ReAct**: Reasoning + Acting in loops
- **Plan-and-Execute**: Upfront planning, then execution
- **Tool-using**: Agents with API/function access
- **Multi-agent**: Specialized agents collaborating

### Use Cases

- Customer support automation
- Research and data analysis
- Code generation and debugging
- Business process automation

Related terms: llm, prompt-engineering

## Context Window

> The maximum amount of text (in tokens) that an LLM can process in a single request, including both the input prompt and generated output.

_Also known as: Context Length, Context Limit, Token Limit_

Context windows determine how much information an LLM can "remember" during a conversation or task. Larger context windows enable more sophisticated applications but come with increased cost and latency.

### Common Context Sizes

- **GPT-4 Turbo**: 128K tokens (~300 pages)
- **Claude 3.5 Sonnet**: 200K tokens (~500 pages)
- **Gemini 1.5 Pro**: 1M tokens (~2,400 pages)
- **GPT-3.5**: 16K tokens (~40 pages)

### Context Window Components

```
[System Message] + [User Prompt] + [Retrieved Docs] + [Chat History] + [Output] ≤ Context Limit
```

### Managing Context

1. **Summarization**: Condense older messages
2. **Sliding window**: Keep recent N messages
3. **RAG**: Retrieve only relevant info, not everything
4. **Compression**: Use techniques like AutoCompressors
5. **Hierarchical**: Split tasks across multiple calls

### Trade-offs

- **Larger context**: More info, higher cost, slower
- **Smaller context**: Less info, cheaper, faster

Related terms: tokens, llm, rag

## Embeddings

> Dense vector representations of text that capture semantic meaning, allowing machines to understand and compare the similarity between different pieces of content.

_Also known as: Vector Embeddings, Text Embeddings, Semantic Embeddings_

Embeddings convert text into arrays of numbers (vectors) in a high-dimensional space where semantically similar text is positioned closer together. This enables semantic search, clustering, and recommendation systems.

### How Embeddings Work

1. **Input text**: "The cat sat on the mat"
2. **Embedding model**: Converts to vector \[0.23, -0.15, 0.87, ...]
3. **Vector space**: Similar meanings → similar vectors

### Common Embedding Models

- **OpenAI text-embedding-3-large**: High quality, API-based
- **Cohere embed-v3**: Multilingual support
- **BGE/E5**: Open-source alternatives
- **Instructor**: Task-specific embeddings

### Applications

- Semantic search in RAG systems
- Document clustering and organization
- Recommendation engines
- Duplicate detection

Related terms: vector-database, rag

## Fine-tuning

> The process of further training a pre-trained LLM on domain-specific data to adapt its behavior, knowledge, or style for specialized tasks or domains.

_Also known as: Fine-tuning, Model Fine-tuning, LLM Fine-tuning_

Fine-tuning takes a general-purpose LLM and specializes it by continuing the training process on a curated dataset. This allows the model to better understand domain-specific terminology, follow particular formats, or exhibit desired behaviors.

### When to Fine-tune

- **Domain expertise**: Medical, legal, technical terminology
- **Format adherence**: Specific output structures
- **Style consistency**: Brand voice, tone
- **Edge cases**: Handle uncommon scenarios

### Fine-tuning vs RAG

- **Fine-tuning**: Bakes knowledge into model weights
- **RAG**: Provides knowledge at query time

Often, the best approach uses both: fine-tune for behavior, RAG for knowledge.

Related terms: llm, rag

## GraphRAG

> An advanced RAG technique that combines knowledge graphs with traditional vector search to capture relationships between entities and enable more nuanced, context-aware retrieval.

_Also known as: Graph RAG, Graph-based RAG, Knowledge Graph RAG_

While traditional RAG retrieves similar documents, GraphRAG also considers the relationships between entities, enabling it to answer complex questions that require understanding connections and context.

### How GraphRAG Works

1. **Extract entities**: Identify people, places, concepts
2. **Build graph**: Connect related entities
3. **Hybrid retrieval**: Combine vector search + graph traversal
4. **Contextual generation**: Use both docs and relationships

### Advantages Over Traditional RAG

- **Relationship awareness**: "Who worked with whom on what?"
- **Multi-hop reasoning**: Follow chains of connections
- **Entity disambiguation**: Distinguish between similar names
- **Temporal context**: Track changes over time
- **Better coherence**: Maintain consistency across related facts

### When to Use GraphRAG

- **Complex domains**: Legal, scientific research, enterprise knowledge
- **Relationship-heavy**: Social networks, org charts, supply chains
- **Multi-entity queries**: Questions involving multiple connected entities
- **Citation tracking**: Following references and attributions

### Tools & Frameworks

- Microsoft GraphRAG
- Neo4j + LangChain
- Knowledge graph databases (Neo4j, Amazon Neptune)

Related terms: rag, embeddings, vector-database

## Hallucination

> When an LLM generates plausible-sounding but factually incorrect or nonsensical information, often with high confidence, presenting fabricated data as truth.

_Also known as: AI Hallucination, Model Hallucination, Confabulation_

Hallucinations occur because LLMs are trained to predict the next token based on patterns, not to verify truth. They don't have access to real-time facts and can't distinguish between accurate and inaccurate information during generation.

### Types of Hallucinations

- **Factual**: Incorrect facts, dates, statistics
- **Fabricated**: Invented references, citations, people
- **Contradictory**: Self-contradicting within response
- **Nonsensical**: Logically impossible statements

### Mitigation Strategies

1. **RAG**: Ground responses in retrieved facts
2. **Lower temperature**: Reduce randomness
3. **Explicit instructions**: "Only use provided context"
4. **Source citation**: Require references
5. **Confidence scores**: Request uncertainty estimates
6. **Human review**: Critical info needs verification

### Why It Matters

Hallucinations are the primary barrier to using LLMs in high-stakes applications like healthcare, legal, and finance. Production systems must implement robust mitigation strategies.

Related terms: llm, rag, temperature

## LLM

> Large Language Models are AI systems trained on vast amounts of text data to understand and generate human-like text, forming the foundation of modern GenAI applications.

_Also known as: Large Language Model, Large Language Models_

LLMs like GPT-4, Claude, and Llama are transformer-based neural networks with billions of parameters, trained to predict the next token in a sequence. This simple training objective enables them to perform a wide variety of language tasks.

### Common LLMs

- **GPT-4** (OpenAI): Multimodal, excellent reasoning
- **Claude** (Anthropic): Long context, strong safety
- **Gemini** (Google): Multimodal, fast inference
- **Llama** (Meta): Open source, customizable

### Key Capabilities

- Text generation and completion
- Question answering
- Summarization
- Translation
- Code generation
- Reasoning and analysis

Related terms: fine-tuning, temperature, tokens, context-window

## Prompt Engineering

> The practice of designing and optimizing input prompts to elicit desired outputs from LLMs, combining techniques like few-shot learning, chain-of-thought reasoning, and role-based prompting.

_Also known as: Prompting, Prompt Design, Prompt Optimization_

Effective prompt engineering can dramatically improve LLM performance without any model training. It's both an art and a science, requiring understanding of how models interpret instructions and context.

### Common Techniques

- **Zero-shot**: Direct instruction without examples
- **Few-shot**: Providing examples in the prompt
- **Chain-of-thought**: Asking model to show reasoning
- **Role prompting**: "You are an expert..."
- **System messages**: Setting behavior context
- **Formatting**: Using XML, JSON, markdown for structure

### Best Practices

1. **Be specific**: Clear, detailed instructions
2. **Provide context**: Background information
3. **Show examples**: Demonstrate desired format
4. **Constrain output**: Specify length, format, tone
5. **Iterate**: Test and refine prompts
6. **Use delimiters**: Separate instructions from content

### Advanced Patterns

- **ReAct**: Reasoning + Acting (for agents)
- **Tree of Thought**: Exploring multiple reasoning paths
- **Self-consistency**: Multiple samples, majority vote

Related terms: llm, temperature, tokens

## RAG

> Retrieval-Augmented Generation combines information retrieval with LLM generation to provide accurate, up-to-date responses grounded in external knowledge bases rather than relying solely on the model's training data.

_Also known as: Retrieval Augmented Generation, Retrieval-Augmented Generation_

RAG works by first retrieving relevant documents from a knowledge base, then using those documents as context for an LLM to generate a response. This approach significantly reduces hallucinations and allows AI systems to access information beyond their training cutoff date.

### How RAG Works

1. **Query Processing**: Convert user query into embeddings
2. **Retrieval**: Find relevant documents using semantic search
3. **Augmentation**: Add retrieved context to the prompt
4. **Generation**: LLM generates response based on context

### Benefits

- Reduces hallucinations by grounding responses in facts
- Enables access to current, domain-specific information
- More cost-effective than fine-tuning for knowledge updates
- Provides source attribution for generated answers

Related terms: embeddings, vector-database, llm

## Temperature

> A parameter controlling the randomness of LLM outputs, where lower values (0-0.3) produce deterministic, focused responses and higher values (0.7-1.0) generate more creative, diverse outputs.

_Also known as: Sampling Temperature, Model Temperature_

Temperature works by adjusting the probability distribution when the model selects the next token. At temperature 0, the model always picks the most likely token (deterministic). At higher temperatures, less likely tokens have a better chance of being selected.

### Temperature Ranges

- **0.0**: Deterministic, repeatable (great for code, facts)
- **0.1-0.3**: Mostly focused, slight variation
- **0.5-0.7**: Balanced creativity and coherence
- **0.8-1.0**: Creative, diverse, sometimes surprising
- **>1.0**: Highly random, often incoherent

### When to Use What

- **Low temperature (0.0-0.3)**:
  - Factual Q&A
  - Code generation
  - Data extraction
  - Classification

- **High temperature (0.7-1.0)**:
  - Creative writing
  - Brainstorming
  - Marketing copy
  - Story generation

Related terms: llm, tokens

## Tokens

> The basic units of text that LLMs process, typically representing parts of words, whole words, or characters, used for input processing, output generation, and usage billing.

_Also known as: Tokenization, Token Count, Subword Tokens_

LLMs don't process text as human-readable words. Instead, they use tokenization to split text into chunks. One token roughly equals 4 characters in English, or about 0.75 words.

### Tokenization Examples

- "Hello, world!" → ~4 tokens
- "AI implementation" → 3 tokens
- "GPT-4" → 2 tokens (`GPT`, `-4`)
- "antidisestablishmentarianism" → 6-8 tokens (split into parts)

### Why Tokens Matter

1. **Cost**: API pricing is per token (input + output)
2. **Context limits**: Max tokens per request (e.g., 8K, 128K)
3. **Performance**: More tokens = slower, more expensive
4. **Prompt design**: Optimize prompts to reduce token count

### Tokenization Strategies

- **Character-level**: One character = one token (rare)
- **Word-level**: One word = one token (rare in modern LLMs)
- **Subword (BPE)**: Balance between characters and words (most common)
- **SentencePiece**: Language-agnostic tokenization

### Tools

- OpenAI tokenizer (tiktoken)
- Hugging Face tokenizers
- Token counters for cost estimation

Related terms: llm, context-window, temperature

## Vector Database

> Specialized databases optimized for storing and querying high-dimensional vectors (embeddings), enabling fast similarity search at scale for RAG and semantic search applications.

_Also known as: Vector DB, Vector Store, Embedding Database_

Vector databases use specialized indexing algorithms (like HNSW, IVF) to efficiently find similar vectors among millions or billions of embeddings, making them essential for production RAG systems.

### Popular Vector Databases

- **Pinecone**: Fully managed, easy to use
- **Weaviate**: Open-source, GraphQL API
- **Qdrant**: Rust-based, high performance
- **Chroma**: Simple, developer-friendly
- **pgvector**: PostgreSQL extension

### Key Features

- **Similarity search**: Find nearest neighbors
- **Filtering**: Combine vector search with metadata
- **Hybrid search**: Mix keyword and semantic search
- **Scalability**: Handle billions of vectors
- **Real-time updates**: Add/update vectors on the fly

Related terms: embeddings, rag