What Is LLM Full Form?

LLM = Large Language Model.

What Is a Large Language Model?

An LLM is a machine learning model trained on a large amount of text to predict the next token in a sequence.

A sequence is the response generated from the prompt that we send to an LLM.

What Are the Use Cases of an LLM?

  1. Answering

  2. Translating

  3. Summarizing

  4. Reasoning

  5. Generating code

What Is the Difference Between an LLM and a Database?

An LLM predicts the next token in a sequence. It does not necessarily retrieve an exact stored answer like a database.

What Happens When We Send a Prompt to an LLM?

The general process looks like this:

  1. Prompt

  2. Tokens

  3. Model processing

  4. Next token possibilities

  5. Token generation

  6. Response

The model generates a response token by token until it reaches an end condition.

What Is Training and Inference?

Training

Training is the process in which a model is fed a large amount of data and its parameters are adjusted to produce better results.

Inference

Inference is the process of sending a prompt to an LLM to generate a result.

What Is a Prompt?

A prompt can contain:

  1. Question

  2. Context

  3. Constraint

  4. Data

  5. Example

  6. Instruction

For example:

You are a technical interviewer.

Ask me one C# interview question at a time.

Start with dependency injection.

That's a prompt.

What Is Prompt Engineering?

Prompt engineering is the practice of designing instructions and context so that an LLM produces more useful, reliable, and consistent outputs.

For example, instead of:

Summarize this document.

you might use:

Summarize this document in exactly 5 bullet points.

Focus on:
1. Business objectives
2. Technical requirements
3. Risks
4. Cost
5. Timeline

Do not introduce information that isn't present in the document.

The second prompt gives the model much more structure.

What Are System, User, and Assistant Messages in an LLM?

System Message

A system message defines the role and expected behavior.

Example:

Suppose you are a financial analyst.

Never fabricate any data.

User Message

A user message is the actual instruction from the user.

Example:

Analyze gold price for 2026.

Assistant Message

An assistant message is the response generated by the LLM.

Example:

Gold price moved sideways between Amt1 to Amt2.

Why Should System and User Messages Be Defined Separately?

  1. A system message defines the application-level behavior that remains consistent across users and requests.

  2. User messages are specific to an individual request.

Can an LLM Remember Previous Conversations?

Not inherently across independent API calls.

An LLM generally processes the context that is provided to it.

For example:

Request 1

User: My name is Harsh.

Request 2

User: What is my name?

If the second request does not contain the previous conversation or context, the model generally has no way to know that information.

A chat application creates the appearance of memory by sending previous messages back to the model or by retrieving stored information.

Application
     ↓
Conversation history / memory
     ↓
LLM context
     ↓
Response

This distinction becomes very important in Semantic Kernel.

You Send the Same Prompt Twice and Get Different Answers. Why?

Because LLM generation is generally probabilistic.

At each generation step, the model can have multiple possible tokens:

"The capital of France is"

Paris       95%
London       2%
Other        3%

The generation configuration determines how those possibilities are selected.

Important parameters include:

  • Temperature

  • Top-p / nucleus sampling

  • Randomness/seed support, depending on the model/API

Higher temperature generally means more variation.

Lower temperature generally means more deterministic behavior.

How Can We Reduce Hallucinations?

Common techniques include:

  • Give the model reliable context.

  • Use RAG.

  • Use embeddings for semantic retrieval.

  • Use tool/function calling.

  • Require citations/references where appropriate.

  • Use structured output.

  • Validate model responses.

  • Don't allow the model to invent missing information.

  • Use deterministic application logic for critical decisions.

What Is a Token in an LLM?

Tokens are units of text used when processing prompts. Depending on the tokenizer, a token may represent a word, part of a word, punctuation, a number, or another unit of text.

Suppose we ask an LLM:

Who is prime minister of India?

The prompt goes to a tokenizer and may be divided into units similar to:

"Who", "is", "prime", "minister", "of", "India", "?"

The exact tokenization can vary depending on the tokenizer used by the model.

Your Application Needs Deterministic Output. What Would You Do?

  • Set temperature to 0 or as low as supported to reduce randomness.

  • Use a fixed model/version so behavior does not unexpectedly change.

  • Use structured output / JSON schema when the response has a defined format.

  • Give the model clear and specific instructions.

  • Keep the prompt and input context consistent.

  • Avoid unnecessary randomness such as sampling where configurable.

  • Use validation in application code to verify the LLM's output.

  • For critical business logic, don't rely solely on the LLM; use deterministic C# code or rules where possible.

What Is a Tokenizer?

The purpose of a tokenizer is to convert a prompt into tokens.

Those tokens are then mapped to token IDs from the tokenizer's existing vocabulary. These token IDs are numerical values.

The token IDs are then sent to the LLM.

The IDs are essentially integer indexes that identify tokens in the tokenizer vocabulary.

If a complete token does not exist in the vocabulary, the text can be broken down into smaller tokens or subwords until recognizable token representations are found. Depending on the tokenizer implementation, the process may eventually use character-level or byte-level representations.