Context window,

personalized,

and stopping criteria.

Acme Design is a full service design agency.

  • Context window
  • Personalized
  • Stopping criteria

Outline

Context window

A context window is the maximum span of tokens (text and code) an LLM can consider at once

Ref

Including both the input prompt and output generated by the model

--> Hallucination ?

Ref

Model by use cases

Case Example Model Pros Cons
Quality & Reasoning Advanced research assistants, complex code generation, highly nuanced content creation, medical/legal text analysis, strategic decision support,... - OpenAI GPT-4o
- OpenAI GPT-4 Turbo / Microsoft Azure GPT-4 Turbo		 - Highest quality
- Large context size		 - High price
- High latency
Complex Data & Long Conversations Document summarization (long papers, books), legal discovery, large codebase analysis,... - Google Gemini 1.5 Pro
- Google Gemini 1.5 Flash		 - Massive context size
- High quality		 - Medium price
- Very high latency
Real-Time Responses & Budget-Conscious Chatbots (especially high-volume customer service), interactive content generation, real-time code suggestions, dynamic content creation,... - OpenAI GPT-4o mini - Very cheap price
- Low latency		 - Low context size
- Medium quality

Comparision GPT model

Ref

Discussion?

Personalized

  • Users' interests, characters, and beliefs,... are different
  • Better user experience

Why

  • Tailoring the behavior and responses of LLMs to individual user preferences or specific contexts (language, tone, and content, etc.)
  • Respond to users in a way that feels customized and personalized to them

What

How

No personalization at all. One model for everyone

Baseline

Giving the model a short description of who the user is

Prompt-based personalization

Adjust the model’s internal behavior based on real user data (train it to understand a specific user’s preferences more deeply)

Fine-tuned personalization

Ref

Ref

The idea of parameter-efficient finetuning techniques

Method Required Input Concept Advantages Disadvantages
Baseline Basic prompt No personalization Simple, fast Not tailored to individual users
Prompt-based personalization Prompt + user attributes Use personal information inside the prompt Easy to implement, no retraining needed Limited effectiveness
Fine-tuning personalization Personal data + model Train a model specifically for each user High effectiveness Resource-intensive, requires personal data

Comparision

  • Social: Tiktok, Youtube, Spotify,...
  • ChatGPT, Gemini,...
  • Gmail's Smart Compose
  • ...

Where

Discussion?

Stopping criteria

Criterias to stop the generated responses

  • Efficiency: Avoid wasting computational resources on generating unnecessary tokens

  • Quality: Generated output is relevant, coherent, and avoids excessive repetition

  • User Experience: can be significantly improved by controlling the length and format of responses

Why

Limit the number of tokens the model can generate.

 

openai.ChatCompletion.create(
    model="gpt-4",
    messages=chat_history,
    max_tokens=512
)

Token Count / Max Tokens

How

OpenAI API

Ref

✅ Pros

  • Simple and effective.

  • Prevents infinite or excessively long outputs.

⚠️ Cons

  • Might cut off meaningful output mid-sentence.

  • Doesn’t adapt based on context or content.

Token Count / Max Tokens

How

Define one or more string sequences that, when generated, immediately stop further output.

 

openai.ChatCompletion.create(
    model="gpt-4",
    messages=chat_history,
    stop=["foo", "bar"]
)

Special Stop Sequences

How

✅ Pros

  • Works well with tools/functions integration.

⚠️ Cons

  • Must know or enforce these sequences in the prompt.

  • Fragile if the model never emits the exact string.

Special Stop Sequences

How

Stop when the generated content semantically completes the task (e.g., completes a paragraph, finishes a function, etc.).

 

if response.endswith((".", "!", "?")) and not continuation_needed(response):
    stop_generation = True

Semantic Stopping

How

✅ Pros

  • Feels natural and human-like.

  • More flexible for open-ended generation.

⚠️ Cons

  • Harder to implement — requires post-processing or heuristics.

  • May need a second model or logic to evaluate “completion.”

Semantic Stopping

How

Some LLMs can be guided with a schema — they stop generating when the function/tool-call is formed correctly.

{
  "functions": [
    {
      "name": "get_weather",
      "parameters": {
        "type": "object",
        "properties": {
          "location": { "type": "string" }
        }
      }
    }
  ]
}

Function Calling & Tool Use

How

✅ Pros

  • Great for building agentic systems.

  • Guarantees structured outputs.

⚠️ Cons

  • Overhead in defining function schemas.

  • Not for freeform text generation.

Function Calling & Tool Use

How

Repetition Penalties, Top-p, Temperature,...

 

openai.ChatCompletion.create(
    model="gpt-4",
    messages=chat_history,
    temperature=0.7,
    top_p=0.9,
    presence_penalty=0.6
)

Sampling-based Criteria

How

✅ Pros

  • More natural termination in creative writing or poetry.

  • Encourages diverse but coherent outputs.

⚠️ Cons

  • Indirect — you’re guiding rather than explicitly stopping.

  • Needs fine-tuning and experimentation.

Sampling-based Criteria

How

Thank You!

Questions?

Made with Slides.com