Skip to content

gptbrowsercodeactivity.md

Latest commit

 

History

History
322 lines (238 loc) · 8.34 KB

gptbrowsercodeactivity.md

File metadata and controls

322 lines (238 loc) · 8.34 KB

GPT Context Window, Session Memory, and Tool-Capable Assistants

(Anki Card, HTML View, Anki Internal Format, Python Reader/Writer)

This document explains:

  • How a context window works
  • How session memory is inserted
  • Which types of GPT-style assistants support which capabilities
  • How a program might process a JSON representation of a Q&A card
  • How to represent this in Anki fields, templates, and HTML
  • How Python pseudocode might simulate the assistant’s behavior

1. Context Window & Session Memory (Short Explanation)

A context window is the text the model can see at once.
It has a fixed size (e.g., 8k, 32k, 128k tokens).
Anything outside it is forgotten unless reinserted.

Session memory is external.
The system stores facts, summaries, or previous messages and reinserts them into the prompt.

No special syntax is required; tags like [System], [Memory], [User] are conventions.

2. Which Types of GPT-Style Assistants Support Which Capabilities?

Below is a general classification of assistant types and what they typically support.

A. Web-enabled assistants (e.g., Copilot, ChatGPT with browsing)

Can do:

  • Web search
  • Read first page of a web resource
  • Use built-in calculator
  • Use project files (if integrated into an IDE)

Cannot do:

  • Modify arbitrary local files without explicit user permission
  • Access private resources unless provided

B. Local offline LLMs (no tools)

Can do:

  • Pure text reasoning
  • Math only through internal reasoning (no calculator)

Cannot do:

  • Web search
  • Read web pages
  • Access files
  • Use external tools

C. Code assistants (IDE-integrated)

Can do:

  • Read project files
  • Apply changes to files
  • Use built-in static analysis
  • Sometimes use a calculator

Cannot do:

  • Web search (unless connected to a tool)

D. Math-enhanced assistants

Can do:

  • Use a built-in calculator
  • Perform symbolic math

Cannot do:

  • Web search
  • File operations

3. Example Anki Note Fields

Topic: GPT Context Window & Tool Capabilities

Background:
Different types of GPT-style assistants support different capabilities.
Some can search the web, some can use a calculator, some can read files.

UserMessage:
Which types of GPT assistants can perform web search, use a calculator, read a web page, or modify project files?

SessionMemory:
- User name: Alex
- Interest: AI internals
- Previous topic: context windows

ContextWindowExample:
[System] You are a helpful assistant.
[Memory] User name: Alex.
[History] User: Explain context windows.
[History] Assistant: A context window is...
[User] Which types of GPT assistants can perform web search, use a calculator, read a web page, or modify project files?

AnswerExplanation:
Web-enabled assistants can search and read pages.
Code assistants can modify project files.
Math-enhanced assistants can use calculators.
Offline LLMs cannot use external tools.

4. Anki Card Templates

Front (Q → A)

<div class="topic">{{Topic}}</div>

<h2>Q:</h2>
<div class="question">{{UserMessage}}</div>

<h3>Session Memory:</h3>
<div class="memory">{{SessionMemory}}</div>

<h3>Context Window Example:</h3>
<div class="context">{{ContextWindowExample}}</div>

Back (A)

<h2>A:</h2>
<div class="answer">{{AnswerExplanation}}</div>

<hr>

<h3>Background:</h3>
<div class="background">{{Background}}</div>

5. Reverse Card (A → Q)

<h2>A (Given):</h2>
<div class="answer">{{AnswerExplanation}}</div>

<hr>

<h2>Original Question:</h2>
<div class="question">{{UserMessage}}</div>

6. CSS for Anki or Browser

.topic {
  font-size: 1.2em;
  font-weight: bold;
  color: #3366cc;
}

.question, .answer, .memory, .context, .background {
  margin: 10px 0;
  padding: 8px;
  border-left: 3px solid #ccc;
  background-color: #f9f9f9;
}

7. Browser HTML Rendering

<div class="card">
  <div class="topic">GPT Context Window & Tool Capabilities</div>

  <div class="label">Q:</div>
  <div class="box">
Which types of GPT assistants can perform web search, use a calculator, read a web page, or modify project files?
  </div>

  <div class="label">Session Memory:</div>
  <div class="box">
- User name: Alex
- Interest: AI internals
- Previous topic: context windows
  </div>

  <div class="label">Context Window Example:</div>
  <div class="box">
[System] You are a helpful assistant.
[Memory] User name: Alex.
[History] User: Explain context windows.
[History] Assistant: A context window is...
[User] Which types of GPT assistants can perform web search, use a calculator, read a web page, or modify project files?
  </div>
</div>

8. Internal Anki JSON-like Representation

{
  "note_type": "GPTContextWindowTools",
  "fields": {
    "Topic": "GPT Context Window & Tool Capabilities",
    "Background": "Different types of GPT-style assistants support different capabilities.",
    "UserMessage": "Which types of GPT assistants can perform web search, use a calculator, read a web page, or modify project files?",
    "SessionMemory": "- User name: Alex\n- Interest: AI internals\n- Previous topic: context windows",
    "ContextWindowExample": "[System] You are a helpful assistant.\n[Memory] User name: Alex.\n[History] User: Explain context windows.\n[History] Assistant: A context window is...\n[User] Which types of GPT assistants can perform web search, use a calculator, read a web page, or modify project files?",
    "AnswerExplanation": "Web-enabled assistants can search and read pages. Code assistants can modify project files. Math-enhanced assistants can use calculators. Offline LLMs cannot use external tools."
  }
}

9. Python Pseudocode: Reader, Writer, JSON Converter, and Assistant Simulation

# -----------------------------------------
# 1. Load the Anki-like JSON card
# -----------------------------------------

import json

with open("card.json", "r") as f:
    card = json.load(f)

fields = card["fields"]

# -----------------------------------------
# 2. Render templates (simplified)
# -----------------------------------------

from string import Template

front_template = Template("""
<div class="topic">${Topic}</div>
<h2>Q:</h2>
<div class="question">${UserMessage}</div>
""")

front_html = front_template.substitute(fields)

# -----------------------------------------
# 3. Simulate different GPT-style assistants
# -----------------------------------------

def simulate_assistant(assistant_type, card_fields):
    """
    assistant_type: "web", "offline", "math", "code"
    """

    if assistant_type == "web":
        # Web-enabled assistant
        # Can perform web search, read pages, use calculator
        return {
            "capabilities": ["web_search", "read_web_page", "calculator"],
            "response": f"Using web tools, I can answer: {card_fields['AnswerExplanation']}"
        }

    if assistant_type == "offline":
        # Offline LLM
        return {
            "capabilities": [],
            "response": "I cannot use external tools, but I can reason about the question."
        }

    if assistant_type == "math":
        return {
            "capabilities": ["calculator"],
            "response": "I can compute things but cannot search the web."
        }

    if assistant_type == "code":
        return {
            "capabilities": ["read_project_files", "apply_file_changes"],
            "response": "I can modify project files based on the question."
        }

# -----------------------------------------
# 4. Convert card to JSON for the assistant
# -----------------------------------------

assistant_input = {
    "question": fields["UserMessage"],
    "memory": fields["SessionMemory"],
    "context": fields["ContextWindowExample"]
}

assistant_json = json.dumps(assistant_input, indent=2)

# -----------------------------------------
# 5. Run through mock assistant
# -----------------------------------------

result = simulate_assistant("web", fields)
print(result)

10. Summary

This document shows:

  • How context windows and session memory work
  • How different types of GPT-style assistants support different capabilities
  • How to represent this in Anki fields, HTML, and JSON
  • How Python pseudocode might simulate the assistant’s behavior

Everything is structured so you can study, modify, or extend it.