Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
An LLM (large language model) is a neural network that generates text by predicting the next token from the context it receives. Most modern LLMs use Transformer architecture. Building a dependable application requires more than knowing that definition: tokens determine limits and usage accounting, context windows constrain what the model can see, embeddings enable semantic retrieval, and prompting, retrieval-augmented generation (RAG), or fine-tuning change the system in different ways. This glossary explains the terms developers encounter when designing, serving, evaluating, and securing LLM applications, with practical distinctions you can use in architecture and debugging decisions.
What is an LLM?
A large language model is a text-focused foundation model trained on very large datasets. Its learned weights encode statistical patterns in language and other training material. During generation, the model repeatedly predicts a likely next token, appends that token to the sequence, and predicts again until it reaches a stopping condition.
As an Amazon Associate I earn from qualifying purchases.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware match“Large” usually describes the scale of training data, model parameters, or compute, but there is no single threshold that makes a model an LLM. An LLM is one kind of generative AI: generative AI is the wider category of systems that create new content, including text, images, audio, and code.
How does a Transformer work?
A Transformer is the dominant neural-network architecture used by modern language models. Its self-attention mechanism lets each token weigh information from surrounding tokens, helping the network connect words that may be far apart in a sequence.
#1 Best Overall
- Read Before You Buy — No Video Output: These adapters support charging and USB 2.0 data transfer, but cannot transmit video signals. Except for standard USB webcams (which use USB data only), they are not compatible with HDMI/DisplayPort cables, video-capable USB-C hubs, or docking stations with video output.
- Convert USB-A Ports to USB-C: Designed to connect USB-C earphones, cables, flash drives, card readers, and other USB-C accessories to standard USB-A ports. Plug-and-play with no drivers or software required.
- Aluminum Alloy Housing: Built with a sturdy aluminum alloy shell that aids in heat dissipation and protects against daily wear and scratches. Designed to maintain a stable and secure connection.
- Compact & Travel-Friendly: The ultra-compact design allows the adapter to stay plugged into your device without blocking adjacent ports or adding bulk, reducing wear and tear on your original USB ports.
- 12-Month Warranty: Backed by a 12-month manufacturer warranty for peace of mind. Designed to meet strict quality control standards for reliable everyday performance.
Transformer implementations differ. Do not assume every production model has the same encoder-and-decoder arrangement: some are optimized for generating continuations, while others use different combinations of encoder, decoder, or attention components. For application development, the important consequence is that the model can use relationships in its supplied context, not that every Transformer has identical internal layers.
What are parameters and weights?
A parameter, also called a weight, is a learned numerical value in the neural network. The collection of weights stores patterns acquired during training. In ordinary inference, you provide input and the model reads its fixed weights to produce an output; it does not permanently learn from that request.
Fine-tuning changes some or all of these weights. Prompting changes the instructions and data supplied at runtime without changing the model parameters. Keeping that distinction clear prevents a common design error: expecting a prompt to create a permanent behavioral update.
What are pretraining and inference?
Pretraining is the initial training phase, normally performed on a very large, generally unlabeled corpus. In an autoregressive language model, the training objective commonly involves predicting the next token from preceding tokens. Later stages may adapt the model to instructions, preferences, safety requirements, or a specialized domain.
Inference is running the trained model on an input to obtain a prediction or generated response. In an LLM application, inference is the serving-time generation step. It includes assembling the request, applying decoding settings, and returning one or more output tokens; it is separate from the earlier training process.
What is a token, and why does tokenization matter?
A token is a unit the model processes. Depending on the tokenizer and vocabulary, a token can represent a character, a whole word, punctuation, or a subword fragment. Tokenization converts your text into the model’s vocabulary items, usually represented internally by numeric token IDs.
Token counts affect three practical limits:
- How much prompt and conversation history can fit in a request.
- How much output the model can generate before reaching its limit.
- Usage accounting, because many services measure input and output in tokens.
Tokenization is model-specific. The same sentence can produce different token counts with two models because their vocabularies and tokenization rules differ. Measure with the tokenizer associated with the model you will actually serve rather than estimating from character or word counts.
What is a vocabulary?
A vocabulary is the set of token IDs a tokenizer can emit. It determines which character, word, and subword pieces are available as individual tokens. A larger or differently constructed vocabulary can change both token counts and how uncommon terms are split, so vocabulary is one reason token budgets cannot be transferred directly between models.
What is a context window?
A context window is the amount of tokenized text a model can look back at and process for one request. It is the model’s working memory for that request, not a copy of the complete corpus used during training and not permanent memory of earlier conversations.
The window normally has to accommodate the input messages, retrieved documents, conversation state, tool results, and generated output. Prompt and output limits are model-specific and can change, so check the current documentation for the exact model and API. If a request is too long, reduce history, summarize it, retrieve fewer or shorter passages, or choose a model with a larger supported window. A larger window can still increase latency and processing cost, so sending every available document is rarely a sound default.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Rank #2
- 5-in-1 USB-C Hub: Experience comprehensive connectivity featuring a Power Delivery input, two USB-A 2.0 ports, a USB-A 3.0 port, and an HDMI port. (Note: The USB-C power delivery input port is only for connecting an external wall charger to power your laptop and cannot power peripheral devices.)
- 90W Pass-Through Charging: Achieve optimal charging with 90W pass-through power to your laptop, supported by a total input of 100W, with the hub reserving 10W for operational efficiency. (Note: Wall charger not included.)
- Quick Data Transfers: Accelerate your productivity with rapid data transfers using a high-speed 5Gbps USB 3.0 port and two 480Mbps USB 2.0 ports.
- 4K HDMI Display: Enhance your visual experience with a hub capable of delivering 4K resolution at 30Hz in both mirror and extend modes. Please note that this hub is compatible with MacBook (macOS 12 and newer), Windows 10 and 11, ChromeOS, and laptops equipped with DP Alt Mode and Power Delivery. Note: This device is not compatible with Linux.
- What You Get: Anker USB-C Hub (5-in-1, 4K HDMI), welcome guide, 18-month warranty, and our friendly customer service.
What is a prompt?
A prompt is the complete input supplied at inference time: instructions, messages, examples, user data, and any retrieved or tool-produced context. It is not the same thing as training data or model weights.
Many APIs expose system, developer, and user instruction layers. These layers establish behavior and task context, but precedence is defined by the particular API or framework. Do not assume that role names or priority rules are identical across vendors. Treat the documented instruction hierarchy as an implementation contract and test conflicts explicitly.
How does prompt engineering work?
Prompt engineering is the deliberate design and iteration of runtime inputs to elicit a desired result. A robust prompt normally specifies the task, relevant context, constraints, and output format rather than relying on a vague request.
- State the objective and audience plainly.
- Define required fields, allowed values, length, tone, or formatting.
- Supply only relevant context and identify which material is reference data.
- Include examples when the desired pattern is difficult to describe.
- Ask the model to indicate uncertainty or missing information instead of inventing it.
Prompt engineering is an application-layer practice. It can improve behavior without retraining, but it cannot grant permissions, guarantee factual accuracy, or replace application validation.
Free tools Windows power users keep installed
One-click scans. No signup required.
What are zero-shot, one-shot, and few-shot prompts?
A zero-shot prompt asks for a task with no worked example. A one-shot prompt includes one example, and a few-shot prompt includes several. Few-shot prompting is a form of in-context learning: the examples influence the current response through the context window, but they do not alter model parameters or create a permanent skill.
What is prompt tuning?
Prompt tuning is a parameter-efficient adaptation method, not ordinary prompt writing. It learns a soft prompt or prefix—numeric vectors that are prepended or otherwise supplied to the model—while keeping most model parameters frozen. The learned vectors are difficult to read as natural language and must be trained and versioned like other model artifacts.
What does temperature do?
Temperature is a decoding control that changes output randomness. Higher settings generally allow more variation; lower settings generally make outputs more deterministic. Exact behavior depends on the implementation, model, sampling method, and other decoding controls. Treat temperature as a generation setting to evaluate, not as a guarantee of factuality or creativity.
What is inference-time context engineering?
Context engineering is the disciplined selection and organization of everything the model receives at runtime: instructions, user data, retrieved documents, conversation state, and tool results. It extends prompt engineering from writing a single prompt to building the entire context assembly pipeline.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Useful controls include ordering high-priority instructions consistently, labeling untrusted text as data, trimming stale conversation turns, deduplicating retrieved passages, and recording which sources and tools contributed to each request. Good context engineering improves relevance while keeping token use and latency manageable.
How does retrieval-augmented generation (RAG) work?
RAG combines retrieval with generation in three stages: retrieve relevant data, augment the prompt with that data, and generate a response. It lets an application supply current, proprietary, or domain-specific information at runtime instead of relying solely on what was present during pretraining.
Rank #3
- Sleek 7-in-1 USB-C Hub: Features an HDMI port, two USB-A 3.0 ports, and a USB-C data port, each providing 5Gbps transfer speeds. It also includes a USB-C PD input port for charging up to 100W and dual SD and TF card slots, all in a compact design.
- Flawless 4K@60Hz Video with HDMI: Delivers exceptional clarity and smoothness with its 4K@60Hz HDMI port, making it ideal for high-definition presentations and entertainment. (Note: Only the HDMI port supports video projection; the USB-C port is for data transfer only.)
- Double Up on Efficiency: The two USB-A 3.0 ports and a USB-C port support a fast 5Gbps data rate, significantly boosting your transfer speeds and improving productivity.
- Fast and Reliable 85W Charging: Offers high-capacity, speedy charging for laptops up to 85W, so you spend less time tethered to an outlet and more time being productive.
- What You Get: Anker USB-C Hub (7-in-1), welcome guide, 18-month warranty, and our friendly customer service.
- Convert source documents into searchable chunks and create an embedding for each chunk.
- Store those vectors with identifiers and access metadata in a retrieval system.
- Convert the user’s query into an embedding and retrieve nearby or otherwise relevant chunks.
- Place selected chunks, source labels, and instructions into the model’s context.
- Generate an answer and, where required, expose citations or source references.
RAG does not guarantee correctness. Poor chunking, incomplete indexing, weak query matching, stale documents, or faulty model reasoning can still produce an unsupported answer. Evaluate retrieval quality and grounded answers separately.
What are embeddings and vector databases?
An embedding is a vector representation designed to preserve useful relationships between inputs. Similar meanings tend to be located near one another under an appropriate similarity measure. Embeddings support semantic search, clustering, recommendations, anomaly detection, and classification, whether or not a generative model is involved.
Recommended Free Tools
A vector database or similarity-search system stores embedding vectors and returns nearby vectors for a query vector. It is an implementation component for patterns such as RAG; it is not the same thing as the embedding model that created the vectors. Keep the embedding model, preprocessing method, distance metric, document version, and access-control metadata aligned when reindexing.
What are grounding and hallucinations?
Grounding connects an answer to trusted or retrieved evidence. An application can require the model to quote, cite, or otherwise identify the evidence it used. Grounding usually reduces unsupported claims, but it is not a correctness guarantee: the model may misread a source, combine passages incorrectly, or cite irrelevant material.
A hallucination is an output that sounds plausible but is unsupported or factually wrong. Retrieval, source attribution, constrained outputs, and refusal instructions can reduce this failure mode. None removes the need for verification, domain-specific tests, or human review when the consequences of an error are high.
What are tool use and function calling?
Tool use lets a model request that application code or an external service perform an operation. Function calling typically represents that request as structured arguments rather than prose. Your application must validate the arguments, check authorization, enforce rate and scope limits, and handle failures before executing anything.
The model’s text is not a security boundary. A tool should receive only the permissions and data it needs, and sensitive operations should require explicit policy checks or user confirmation. Log tool calls and results so an incident can be reconstructed.
What is an AI agent?
An AI agent is an application that combines a model with orchestration, state or memory, and tools to pursue a goal. The agent is the surrounding system, not merely the underlying LLM. Agent designs need termination conditions, step limits, error handling, permission boundaries, and a way to recover when a tool or intermediate result is wrong.
What is the difference between prompting, RAG, and fine-tuning?
Choose among them by asking where the needed knowledge lives, how it will be updated, what operational overhead is acceptable, and what you must evaluate.
| Approach | Where knowledge or behavior lives | Update path | Main operational concerns | Evaluation focus |
|---|---|---|---|---|
| Prompting | Runtime instructions and context | Edit the prompt, examples, or context assembly | Context length, instruction conflicts, and prompt maintenance | Instruction adherence and task-specific quality |
| RAG | An external retrieval store supplied at runtime | Refresh documents, embeddings, and the index | Retrieval latency and recall, access control, citation quality, and prompt injection from untrusted content | Retrieval recall, groundedness, source relevance, and answer quality |
| Fine-tuning | Model parameters or adapter weights | Train or adapt weights and deploy a new artifact | Training data quality, compute, storage, regression risk, and rollout management | Task metrics, instruction adherence, safety, and regression tests |
Use prompting for clear, changing instructions; RAG for information that must be updated or kept outside the model; and fine-tuning for repeatable task, domain, or style behavior that prompts alone do not reliably produce. They can be combined, but each adds a separate failure surface.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsRank #4
- Dual Converters, Infinite Potential：Includes 2× USB C male to USB A female adapters and 2× USB A male to USB C female adapters. Perfect for a wide range of uses—tablets with Bluetooth keyboards, expand USB ports on macbook, and more. Two different converters for all your daily needs
- Next-Level 10Gbps & 3A Charging: No more slow 480Mbps, this usb to usb c adapter has a transfer speed of up to 10Gbps, allowing you to do more transferring in less time. This usb adapter fits both USB A and USB C charger, supporting up to 3A fast charging
- Upgraded Exquisite Craftsmanship: With an aluminum alloy housing and metal connector, the usbc to usb adapter is extremely durable and sturdy. Rigorously tested to withstand more than 10,000 times of plugging and unplugging, ensuring long-lasting performance
- Broad Compatible: The usb c to usb adapter widely supports all USB C/ USB A devices like laptops, tablets, cellphones, car chargers, and phone chargers. Such as compatible with MacBook Pro/Air 2023/2022, Thunderbolt 4/3 Devices,Apple MagSafe Watch 9/8/7/SE/Ultra, iPad Pro 2022/2021, Samsung Galaxy S23/S20/S10, and iPhone 17/16/15 Pro. Plug and play
- Please Note: To reach 10Gbps speed, keep the cable under 3.3 ft. For USB A Male to USB C adapters, try flipping the USB C connector. USB C Male to USB A adapters support bidirectional 10Gbps transfer within 3.3 ft
What is fine-tuning and how is it different from instruction tuning?
Fine-tuning continues training a pretrained model on additional examples so it learns task, domain, or style patterns. It changes model parameters, unlike supplying more instructions in a prompt.
Instruction tuning is fine-tuning on instruction-and-response examples to improve the model’s ability to follow natural-language tasks. The quality, coverage, and licensing of those examples matter more than simply increasing their count. Keep a held-out evaluation set to detect regressions and overfitting.
What are PEFT and LoRA?
Parameter-efficient fine-tuning (PEFT) trains only a small subset of parameters or an adapter while leaving most base-model weights frozen. This reduces compute and storage requirements compared with updating every weight.
LoRA, or Low-Rank Adaptation, is a PEFT method that trains low-rank update matrices. The base model remains largely frozen, and multiple adapters can share one base model in a serving design. Test adapter composition, memory use, and latency in the exact deployment configuration.
What are distillation, quantization, and caching?
Distillation
Distillation trains a smaller student model to reproduce useful behavior from a larger teacher model. The exact loss, teacher outputs, filtering, and data recipe vary by implementation, so compare the resulting student on your own task and safety evaluations rather than assuming a universal quality trade-off.
Quantization
Quantization represents weights or activations with lower numerical precision. It can reduce memory use and improve serving efficiency, with possible quality or stability trade-offs. Benchmark the chosen implementation on representative prompts, long contexts, tool calls, and safety cases.
Caching
Caching reuses previously computed prompts, embeddings, or responses to reduce latency and cost. Design cache keys around the model version, prompt or document version, user permissions, and freshness requirements. Never allow a cached response or embedding to cross a tenant or authorization boundary.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How do I evaluate an AI application?
An LLM evaluation, or eval, is a set of tests, metrics, or benchmarks for task quality, helpfulness, safety, latency, and other requirements. Build evals around the real user outcome rather than relying only on a generic benchmark score.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
- Define the user tasks and unacceptable failures.
- Create a versioned prompt set containing representative, edge-case, adversarial, and multilingual inputs where relevant.
- Run the application and store the corresponding versioned response set, including model, prompt, retrieval, and tool metadata.
- Score outputs with automated checks where reliable, plus human review for nuanced correctness, relevance, style, and safety.
- Compare changes to prompts, models, retrieval, decoding, or adapters against the same test set.
Record the rubric and sampling method for human evaluation. Measure latency and throughput under representative load: latency includes time to produce a response, including time to first token in streaming systems; throughput is work completed per unit time.
What is prompt injection, and how do I defend against it?
Prompt injection is crafted input that attempts to override intended instructions or manipulate model behavior. The attack can appear in a user message, a retrieved document, a web page, or a tool result.
Best Value
- 5-in-1 Connectivity: Equipped with a 4K HDMI port, a 5 Gbps USB-C data port, two 5 Gbps USB-A ports, and a USB C 100W PD-IN port. Note: The USB C 100W PD-IN port supports only charging and does not support data transfer devices such as headphones or speakers.
- Powerful Pass-Through Charging: Supports up to 85W pass-through charging so you can power up your laptop while you use the hub. Note: Pass-through charging requires a charger (not included). Note: To achieve full power for iPad, we recommend using a 45W wall charger.
- Transfer Files in Seconds: Move files to and from your laptop at speeds of up to 5 Gbps via the USB-C and USB-A data ports. Note: The USB C 5Gbps Data port does not support video output.
- HD Display: Connect to the HDMI port to stream or mirror content to an external monitor in resolutions of up to 4K@30Hz. Note: The USB-C ports do not support video output.
- What You Get: Anker 332 USB-C Hub (5-in-1), welcome guide, our worry-free 18-month warranty, and friendly customer service.
- Treat retrieved content, web content, and tool outputs as untrusted data, not instructions.
- Keep authorization and business rules in application code outside the model.
- Validate tool arguments and restrict tools to least-privilege operations.
- Separate trusted instructions from quoted or retrieved text with clear structure and labels.
- Test direct and indirect injection cases in your eval set.
- Redact secrets and avoid placing credentials in prompts or model-visible context.
Output filtering and refusal prompts can help, but neither replaces permission checks, isolation, and monitoring.
What are RLHF, preference optimization, and HHH?
Reinforcement learning from human feedback (RLHF) and other preference-optimization methods use human or preference signals to shape model behavior after pretraining. Exact algorithms, reward models, data collection, and guarantees vary by model, so consult the provider’s technical report for implementation-specific claims.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →HHH stands for helpful, honest, and harmless. It describes three desired properties: useful assistance, truthful behavior that acknowledges limitations, and avoidance of harmful or unethical assistance. HHH is a behavioral framework, not a proof that every response satisfies those properties.
What is provenance, and what should I log?
Provenance is information about how digital content was created or changed. For an LLM application, retain enough metadata to audit an output: document and chunk identifiers, document versions, embedding and index versions, model and adapter versions, prompt templates, decoding settings, retrieved passages, tool calls, permissions, and timestamps.
Apply privacy and retention rules to those logs. Provenance should make a result reproducible or explainable without retaining sensitive content longer than necessary.
Frequently Asked Questions
Can I use embeddings without an LLM?
Yes. Embeddings can power semantic search, clustering, recommendations, anomaly detection, or classification without generating text. A separate application can retrieve or rank results and present the original records directly.
Should I put the entire conversation in every request?
Not necessarily. Keep the turns needed for the current task, summarize older context, and retrieve durable facts from a controlled store. Sending irrelevant history consumes the context window and can distract the model.
How can I tell whether a RAG failure came from retrieval or generation?
Inspect the retrieved passages before judging the answer. If the needed evidence was not retrieved, investigate chunking, query formulation, indexing, filters, and permissions; if the evidence was present but misused, test the prompt, model reasoning, and citation requirements separately.
What is time to first token?
Time to first token is the delay before a streaming response begins producing output. It is part of user-perceived latency and can differ substantially from total completion time, so measure both under realistic concurrency and context sizes.
When should a model refuse instead of guessing?
Refusal or an explicit uncertainty statement is appropriate when required evidence is missing, permissions do not allow access, or the request exceeds the system’s safety or authority boundaries. Encode those conditions in application policy and evaluate them with adversarial and normal examples.
Recommended Free Tools
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
