Storage System Notes

[appatalks]

Using this as a scratch pad for ideas.

• Let's assume Context window is limited to 128000

[appatalks]

Idea 1

1. Contextual Retrieval

The model identifies the required context and invokes function to fetch specific data from the knowledge graph.

Example:

• Query: "What rank is Lt. Cmdr. Data, and which organization does he serve?"
• SQL API Call:

  SELECT json_extract(json_data, '$.properties.rank') AS rank,
         json_extract(json_data, '$.relationships')
  FROM knowledge_graph
  WHERE entity = 'Data';
  

Response:

{
  "rank": "Lieutenant",
  "relationships": [
    { "type": "serves_in", "target": 2 }
  ]
}

• When a learned fact is found, a function is used to update/add new line records.

2. Structure this further by Nodes and Relationships

[Data] --(rank: "Lieutenant")--> [Rank]
       --(species: "Android")--> [Species]
       --(serves_in)--> [Starfleet]
       --(uses)--> [Positronic Brain]
       --(promoted_to)--> [Captain]

[Starfleet] --(founded_by: "United Earth")--> [United Earth]
            --(focuses_on)--> [Exploration and Diplomacy]

[Positronic Brain] --(created_by)--> [Dr. Noonien Soong]
                   --(category: "Technology")--> [Technology]

• With example node details:

Node: Data

• Properties:
  • rank: "Lieutenant"
  • species: "Android"
• Relationships:
  • serves_in: Starfleet
  • uses: Positronic Brain
  • promoted_to: Captain

---

Framework Breakdown Idea

Core Framework Structure for Intelligence

1. Top-Level Nodes (Categories)

Broad categories that form the foundation of intelligence:

• Personality: Traits, preferences, and motivations.
• Knowledge: General, specialized, and evolving facts.
• Skills: Learned abilities or tasks.
• Reasoning: Logical processes and decision-making.
• Memory: Long-term (facts/events) and working memory (contextual).
• Sensory Input: Data from external environments (vision, hearing, etc.).

---

2. Sub-Nodes (Specialties)

Each top-level node branches into more specific domains:

Knowledge

• General:
  • History
  • Geography
  • Science
• Specialized:
  • Startrek fandom
  • Robotics
  • Programming

Skills

• Communication:
  • Language
  • Emotional intelligence
• Technical:
  • Coding
  • Problem-solving
  • Diagnostics

Memory

• Working Memory:
  • Recent queries
  • Current tasks
• Long-Term Memory:
  • Historical events
  • Acquired knowledge

Reasoning

• Logical:
  • Decision trees
  • Deductive reasoning
• Creative:
  • Hypothesis generation
  • Problem-solving heuristics

---

3. Relationships (Connections)

Semantic and functional links between nodes:

• Is-a: Defines categories (e.g., "Data is an Android").
• Part-of: Links components to larger entities (e.g., "Starfleet is part of the Federation").
• Uses: Functional relationships (e.g., "Data uses a Positronic Brain").
• Learns-from: Tracks sources of new knowledge or skills.
• Adapts-to: Represents flexibility and contextual changes.

---

Example Framework in Text Representation

[Personality]
  --(trait: "Logical")--> [Trait]
  --(trait: "Curious")--> [Trait]

[Knowledge]
  --(general)--> [History]
  --(specialized)--> [Starfleet]

[Skills]
  --(technical)--> [Coding]
  --(communication)--> [Language]

[Memory]
  --(working_memory)--> [Recent Queries]
  --(long_term_memory)--> [Historical Events]

[Reasoning]
  --(logical)--> [Decision Trees]
  --(creative)--> [Hypothesis Generation]

[appatalks]

When a learned fact is found, a function is used to update/add new line records.

I'll need to think of a retrieval and delivery method too.

I might need to have all outputs processed by another model in the background and trigger some post hooks.

[appatalks]

Must read goodness

https://cookbook.openai.com/examples/embedding_wikipedia_articles_for_search

[appatalks]

Vector Databases should be able to achieve this.

[appatalks]

This "thought": true response the gemini model returns I suspect has potential as an internal hook for data lookup. Need to see what the o3 models will be like too.

{
   "text": "The user is asking a simple mathematical question. I need to provide the correct answer.",
    "thought": true
},

[appatalks]

Idea 2

• Google's Notebook LM seems very powerful with its organization and what it does as a product. Perhaps there is some utility here.

[appatalks]

Idea 3

Similar to Vector Databases, OpenAI has Vector Stores .. How did I not know about that?! (Maybe cuz its Beta)

And I still rather have core learned offline information stored, well offline.