AI systems that autonomously plan, decide, and take sequences of actions to complete complex goals without continuous human guidance. Unlike simple chatbots, agents use tools, memory, and multi-step reasoning to accomplish tasks end-to-end. The dominant paradigm for enterprise AI in 2026, powered by frameworks like LangGraph.

The core mathematical operation behind transformer models — allows a model to weigh the importance of different input tokens when generating each output token. "Self-attention" lets the model relate every position in a sequence to every other position, enabling deep contextual understanding regardless of distance.

Automated Machine Learning — tools and frameworks that automatically select models, tune hyperparameters, and engineer features, reducing the need for manual ML expertise. Key examples include Azure AutoML, Databricks AutoML, and Google AutoML Tables. Best suited for tabular data problems with standard objectives.

Best Match 25 — a probabilistic ranking algorithm for keyword-based text retrieval. The backbone of traditional search engines (Elasticsearch, OpenSearch) and still essential in "hybrid search" alongside dense vector retrieval. BM25 excels at exact keyword matching where semantic search underperforms, making it a critical complement in production RAG systems.

Bidirectional Encoder Representations from Transformers — Google's 2018 model that reads text bidirectionally, enabling deep contextual understanding. While superseded by GPT-style models for generation, BERT-family models (RoBERTa, DeBERTa) remain the gold standard for cross-encoder re-ranking, NER, and classification tasks in production RAG pipelines.

A prompting technique where you instruct the LLM to reason step-by-step before producing a final answer. Adding "Let's think step by step" or providing worked examples dramatically improves performance on math, logic, planning, and multi-step reasoning tasks. Zero-shot CoT works surprisingly well; few-shot CoT is more reliable for consistent structured reasoning.

The process of splitting large documents into smaller segments before embedding them for RAG retrieval. Chunk size (256 vs 1024 tokens) and strategy (fixed-size, sentence-boundary, semantic, recursive character) critically affect retrieval quality and generation faithfulness. No universal optimal exists — it depends on document type and query distribution.

The maximum amount of text (measured in tokens) an LLM can process in a single interaction — including system prompt, conversation history, retrieved documents, and generated output. GPT-4o: 128K tokens. Claude 3.5 Sonnet: 200K tokens. Gemini 1.5 Pro: 1M tokens. Context window size is the primary constraint in RAG architecture and agentic system design.

Data Analysis Expressions — the formula language used in Power BI, Excel Power Pivot, and SSAS. Enables complex business calculations including time intelligence (YTD, MoM, rolling averages), dynamic segmentation, ratio analysis, and context-sensitive aggregations. Learning DAX is the single highest-leverage skill for Power BI professionals.

An open-source storage format built by Databricks that adds ACID transactions, schema enforcement, time travel (query data as it was at any past point), and versioning on top of data lakes like S3 or ADLS Gen2. The foundation of the Lakehouse architecture and the default table format for Azure Databricks and Snowflake Dynamic Tables.

Finding relevant documents using semantic meaning encoded in dense embedding vectors rather than exact keyword matching. A query embedding is compared to document embeddings using cosine or dot-product similarity. Outperforms BM25 for conceptual or paraphrased queries but may miss exact entity matches — which is why hybrid search (dense + BM25) consistently wins in production.

A dense numerical vector representation (typically 768–3072 dimensions) of text, images, or audio where semantically similar content is geometrically close in vector space. The foundation of semantic search, RAG, recommendation systems, and clustering. Key models: OpenAI text-embedding-3-large, Cohere embed-v3, BGE-M3 (open-source). Choice of embedding model is often more impactful than vector database choice.

The systematic process of measuring LLM and RAG output quality. Offline evaluation uses benchmark datasets with ground-truth labels. Online evaluation monitors production metrics. Key dimensions: faithfulness (is the answer grounded in retrieved docs?), relevance (does it answer the question?), groundedness (is it supported by evidence?). LLM-as-judge is now the dominant approach for scalable evaluation.

Providing 2–8 input/output examples in the prompt to demonstrate the desired output format or reasoning style to the LLM. More reliable than zero-shot for structured extraction, classification, and domain-specific tasks. Example selection matters enormously — diverse, high-quality demonstrations outperform random sampling by 20–40% on typical benchmarks.

Continued training of a pre-trained LLM on domain-specific data to specialise its behaviour for particular tasks or knowledge domains. Usually applied via LoRA or QLoRA for efficiency. Fine-tuning excels at consistent output style and domain vocabulary but cannot reliably add new factual knowledge — use RAG for knowledge grounding instead. Costs 10–100× more than prompt engineering.

An evaluation metric for RAG systems that measures whether the LLM's answer is supported by the retrieved context documents, as opposed to being generated from the model's parametric memory (hallucination). Measured by RAGAS framework. A groundedness score > 0.85 is typically considered production-ready. The primary quality signal for RAG system iteration.

When an LLM confidently generates plausible-sounding but factually incorrect information — fabricating statistics, citations, code, or events that don't exist. The primary risk in production LLM applications. Mitigation strategies: RAG (ground in retrieved facts), chain-of-thought (explicit reasoning), constitutional AI (self-critique), systematic evaluation pipelines, and human-in-the-loop for high-stakes outputs.

Hierarchical Navigable Small World — the approximate nearest-neighbor (ANN) algorithm used by most vector databases (Pinecone, Qdrant, Weaviate, Chroma) for fast, scalable similarity search across millions or billions of embeddings. Achieves sub-millisecond query times with >95% recall. The M and ef_construction parameters control the recall/speed trade-off at index build time.

Combining dense (semantic vector) retrieval with sparse (BM25 keyword) retrieval, typically fused via Reciprocal Rank Fusion (RRF) or weighted linear combination. Consistently outperforms either method alone across diverse RAG benchmarks — especially for mixed query types (some conceptual, some exact-match). Now supported natively by Pinecone, Azure AI Search, Qdrant, and Elasticsearch.

The ability of large language models to learn new tasks from examples provided directly in the prompt, without any gradient updates or weight changes. The foundation of few-shot and chain-of-thought prompting. ICL capability scales strongly with model size — smaller models (<7B parameters) typically need fine-tuning for reliable task performance where large models ICL successfully.

A feature of OpenAI, Anthropic, and Google LLM APIs that forces the model to always return syntactically valid JSON. Critical for reliable structured output in production pipelines, tool calling, and agentic systems where downstream code parses the response. Combined with Pydantic schema validation, JSON mode dramatically reduces agentic system failures from malformed outputs.

Key-Value Cache — stores previously computed attention states during LLM inference to avoid recomputation on repeated prefixes. Dramatically reduces latency (40–70%) and cost for multi-turn conversations and shared system prompts. Prompt caching is now available via Anthropic and OpenAI APIs, making long system prompts effectively free on repeated calls. Critical for cost-efficient agentic systems.

The most widely adopted Python framework for building LLM-powered applications. Provides abstractions for chains, agents, memory, tools, vector stores, and RAG pipelines. LangChain Expression Language (LCEL) enables composable, streamable pipelines. While LangGraph (built on LangChain) handles stateful agentic workflows, LangChain remains essential for simpler chains, RAG retrieval, and tool integration.

A graph-based framework (built on LangChain) for building stateful, multi-agent AI systems. Represents agent workflows as directed graphs where nodes are Python functions and edges define control flow, enabling cycles, conditional branching, and persistent state. The industry standard for production agentic AI in 2026, with built-in support for checkpointing, human-in-the-loop, and streaming.

Large Language Model — a deep learning model (typically transformer-based) trained on vast text corpora to predict and generate human language. Modern LLMs (GPT-4o, Claude 3.5, Gemini 1.5) exhibit emergent capabilities including reasoning, coding, mathematics, and instruction-following that weren't explicitly trained. The capability foundation of all modern AI applications.

Low-Rank Adaptation — a parameter-efficient fine-tuning technique that trains small, low-rank adapter matrices inserted into the model's attention layers, rather than updating all model weights. Reduces trainable parameters by 99%+ while achieving comparable results to full fine-tuning. QLoRA adds 4-bit quantization, enabling fine-tuning of 70B models on a single A100 GPU. Now the standard method for domain adaptation.

A data design pattern in Lakehouse platforms (Databricks, Synapse, Snowflake) with three progressive quality tiers: Bronze (raw, as-is data), Silver (cleaned, deduplicated, joined), and Gold (business-ready aggregates and dimensional models). Each layer adds quality and structure. Enables both real-time streaming ingestion and batch analytics from the same storage layer.

An architecture where multiple specialised AI agents collaborate on complex tasks — a researcher agent finds information, a writer agent drafts content, a critic agent reviews quality, an executor agent takes action. Enables task parallelism and specialisation beyond single-agent capabilities. Key patterns: supervisor-worker, peer-to-peer, hierarchical delegation. Implemented in LangGraph, AutoGen, and CrewAI.

The coordination of multiple LLM calls, tool uses, memory retrieval operations, and human approvals in an agentic pipeline. Frameworks like LangGraph handle orchestration by managing state transitions, conditional routing, and error recovery. Key challenges: determinism vs flexibility, cost control, latency budgeting, and graceful degradation when individual components fail.

A fully managed vector database service optimised for production-scale similarity search. Offers serverless and pod-based deployments, metadata filtering, hybrid search (dense + sparse), and multi-tenancy via namespaces. Popular choice for production RAG systems due to its managed infrastructure eliminating operational overhead. Alternatives: Qdrant (self-hosted), pgvector (Postgres), Weaviate (multi-modal).

A security attack where malicious text embedded in user input, web pages, documents, or API responses overrides the LLM's system prompt instructions. The SQL injection of the AI era — particularly dangerous in agentic systems with tool use. Mitigations: input sanitisation, prompt structure hardening, LLM output validation, privilege separation, and human-in-the-loop for irreversible actions.

The ETL (Extract, Transform, Load) engine in Power BI, Excel, and Azure Data Factory that uses the M formula language to connect, clean, reshape, and combine data from virtually any source. Power Query's step-by-step transformation history makes data cleaning auditable and reproducible — critical for enterprise BI governance. Proficiency in Power Query is essential for any serious Power BI development.

Quantized Low-Rank Adaptation — fine-tuning with 4-bit NormalFloat quantization combined with LoRA adapters. Reduces memory requirements by 4–8× compared to full precision, making it possible to fine-tune 7B–70B parameter models on consumer-grade or single-GPU setups. The dominant fine-tuning method for open-source models (Llama, Mistral, Falcon) in 2026. Implemented via Hugging Face PEFT and bitsandbytes.

Retrieval-Augmented Generation — a technique that retrieves semantically relevant documents from an external knowledge base at query time and injects them into the LLM's context window before generating a response. Grounds LLM outputs in current, specific information and dramatically reduces hallucination vs purely parametric generation. The dominant architecture for enterprise LLM applications in 2026.

Retrieval-Augmented Generation Assessment — an open-source evaluation framework (pip install ragas) that measures RAG pipeline quality across four dimensions: faithfulness (answer grounded in context?), answer relevancy (answers the question?), context recall (retrieved all needed information?), and context precision (retrieved only relevant information?). The standard evaluation toolkit for production RAG systems.

Reasoning + Acting — a prompting and agent design pattern where the LLM alternates between Thought (explicit reasoning about the current state), Action (tool call or external step), and Observation (tool result). Introduced by Yao et al. 2022, ReAct remains the foundation pattern for virtually all tool-using agents and is built into LangChain, LangGraph, and AutoGen agent implementations.

A second-stage retrieval step that takes the top-K initial candidates from vector search and re-scores them using a more powerful cross-encoder model (Cohere Rerank, BGE-Reranker) that jointly processes the query and each document. Significantly improves precision — cross-encoders consistently achieve 10–25% better MRR than bi-encoders on standard benchmarks, justifying the additional latency (50–150ms).

A cloud-native data warehouse platform with a multi-cluster shared data architecture that separates compute from storage, enabling instant elasticity, zero-copy cloning, and cross-cloud data sharing. Features include Time Travel (query past states), Fail-Safe, Snowpark (Python/Java inside Snowflake), Cortex AI (LLM functions directly on your data), and Marketplace (data product exchange). The dominant enterprise data platform in India in 2026.

Snowflake's developer framework that brings Python, Java, and Scala execution directly inside the Snowflake platform. Enables building ETL pipelines, ML models, and UDFs using familiar dataframe APIs without data movement. Snowpark ML provides scikit-learn-compatible APIs for model training inside Snowflake. Critical for teams wanting to consolidate their ML and data engineering stack on Snowflake.

LLM responses forced into a specific predefined schema (JSON, XML, YAML) via API features (OpenAI Structured Outputs, Anthropic tool_use) or constrained decoding. Critical for reliable agentic tool calling and downstream programmatic processing. Combined with Pydantic model validation, structured output virtually eliminates parsing failures that plagued early LLM integrations. Best practice: always prefer structured output over prompt-only format instructions in production.

A sampling parameter (0–2, default varies by model) that controls the randomness of LLM outputs by scaling the probability distribution over tokens. Temperature 0: highly deterministic (always picks highest probability token). Temperature 1: sample proportionally from distribution. Temperature >1: increased randomness and creativity. Rule of thumb: 0 for data extraction/code, 0.3–0.7 for Q&A, 0.8–1.0 for creative writing.

The atomic unit of text that LLMs process and generate. English text averages ~¾ word per token (roughly 4 characters). "ChatGPT" ≈ 2 tokens. LLM pricing and context window limits are measured in tokens. GPT-4o costs $2.50/M input tokens. Understanding token economics is essential for cost-effective LLM application design, especially for long-context and high-volume use cases.

The ability of LLMs to call external functions, APIs, databases, or code interpreters by generating structured tool-call requests. The mechanism that transforms LLMs from text generators into agents capable of taking real-world actions — web search, code execution, database queries, API calls. Supported by GPT-4, Claude 3.5, Gemini 1.5. The most important capability for agentic system design.

A database system optimised for storing, indexing, and searching high-dimensional embedding vectors using approximate nearest-neighbor (ANN) algorithms like HNSW and IVF. The core infrastructure component of any production RAG system. Key options: Pinecone (managed), Qdrant (open-source, Rust), Weaviate (multi-modal), Chroma (dev-friendly), pgvector (Postgres extension). Choice depends on scale, self-hosting vs managed, and metadata filtering requirements.

Asking an LLM to perform a task with no provided examples, relying entirely on its pre-trained knowledge and instruction-following capabilities. Works well for simple tasks (summarisation, basic classification) with large models. For structured extraction, domain-specific tasks, or consistent formatting, few-shot prompting almost always outperforms zero-shot. "Zero-shot CoT" ("Think step by step") is a powerful middle ground.

A multi-dimensional clustering command in Snowflake that physically co-locates related data across multiple columns on storage micropartitions. Running CLUSTER BY (col1, col2) followed by ALTER TABLE ... CLUSTER BY dramatically reduces the micropartitions scanned for filter-heavy analytical queries, cutting costs by 60–90% for high-cardinality filter patterns. Particularly effective for date + category combinations in large fact tables.