Architecture Reference

Appearance

AI Models and LLM Systems ​

AI Models and LLM Systems Graphics Coverage ​

Primary chapter graphic: LLM Glossary Map, Generative AI Learning Roadmap, Transformer Encoder Decoder Flow, AI Concepts Map, MCP Server Catalog, Frontier Model One-Pager, Generative AI Tech Stack, Reasoning Model Training Path, Open Model Inference Flow, RAG vs Fine-Tuning, Full Fine-Tuning vs LoRA vs RAG, Open Source AI Stack, How LLMs See Text, Visual Token Context Compression. Accepted graphics: 14. Reviewed non-signal pages: 5. Open graphics in review: 0. QA status lives in graphics audit and visual review ledger.

Corpus pages: p. 38, p. 43, p. 74, p. 131, p. 147, p. 158, p. 175, p. 194-195, p. 200-201, p. 230, p. 242, p. 266-267, p. 277, p. 326-327, p. 335-336, p. 343, p. 364, p. 373-374 Coverage: 24 pages; low-confidence extraction ranges: p. 364, p. 373-374

This chapter is part of Marius's owned architecture build corpus. The text routes decisions; durable implementation signal is carried by accepted graphics, reviewed non-signal decisions, and the linked QA audit.

Chapter Visuals ​

Accepted graphics carry the canonical design signal for this chapter. Each selected source page is either accepted as a graphic or explicitly marked non-signal in the source-faithful ledger. Review and QA state live in visual inventory, visual review ledger, and graphics audit.

LLM Glossary Map ​

LLM Glossary Map

Generative AI Learning Roadmap ​

Generative AI Learning Roadmap

Transformer Encoder Decoder Flow ​

Transformer Encoder Decoder Flow

MCP Server Catalog ​

MCP Server Catalog

AI Concepts Map ​

AI Concepts Map

Frontier Model One-Pager ​

Frontier Model One-Pager

Generative AI Tech Stack ​

Generative AI Tech Stack

Reasoning Model Training Path ​

Reasoning Model Training Path

Open Model Inference Flow ​

Open Model Inference Flow

RAG vs Fine-Tuning ​

RAG vs Fine-Tuning

Full Fine-Tuning vs LoRA vs RAG ​

Full Fine-Tuning vs LoRA vs RAG

Open Source AI Stack ​

Open Source AI Stack

How LLMs See Text ​

How LLMs See Text

Visual Token Context Compression ​

Visual Token Context Compression

Open Review Queue ​

  • none

Reviewed Non-Signal Pages ​

  • AI Models And LLM Systems: LLM + Vector Search Map: source p. 195; batch 08; status non-signal/reviewed; ledger reason in visual-review-ledger.json
  • AI Models And LLM Systems: LLM + Tool Map: source p. 230; batch 08; status non-signal/reviewed; ledger reason in visual-review-ledger.json
  • AI Models And LLM Systems: LLM + Tool Map: source p. 242; batch 11; status non-signal/reviewed; ledger reason in visual-review-ledger.json
  • AI Models And LLM Systems: LLM + Embedding Map: source p. 277; batch 12; status non-signal/reviewed; ledger reason in visual-review-ledger.json
  • AI Models And LLM Systems: Embedding + Pattern Map: source p. 131; batch 16; status non-signal/reviewed; ledger reason in visual-review-ledger.json

Use When ​

  • A workflow needs language understanding, extraction, classification, summarization, or generation.

Avoid When ​

  • A deterministic rule is cheaper, clearer, and more reliable.

Core Model ​

  • Model systems combine prompt, context, tool constraints, evaluation, fallback, and cost controls.
  • Prefer explicit ownership over accidental coupling. Every boundary should say who owns correctness, cost, data, recovery, and change.
  • Use corpus page pointers for inspection, and keep the chapter notes focused on reusable design decisions.

Implementation Guidance ​

  • Define task type, acceptable error, source context, output schema, review threshold, and regression set.
  • Write the smallest useful design note: purpose, inputs, outputs, state, failure behavior, observability, and rollback.
  • Choose the first implementation that can be tested against the real workflow without hiding a known production risk.

Tradeoffs ​

  • Fine-tuning can specialize behavior; retrieval keeps knowledge fresh without changing weights.
  • Centralization reduces duplicated work but can become a bottleneck when every team needs exceptions.
  • Specialized infrastructure helps at scale, but it must earn its operational cost.

Failure Modes ​

  • A model is evaluated by vibes instead of test cases tied to business risk.
  • The diagram shows boxes but not ownership, retry behavior, data freshness, or user-visible failure.
  • The system has no proof path for the highest-risk assumption.

Decision Checklist ​

  • Measure accuracy, refusal behavior, hallucination rate, latency, cost, and human review load.
  • Name the owner, source of truth, timeout, retry policy, and evidence that the path works.
  • Add one regression check for the failure mode most likely to recur.

Neutral Automation Examples ​

  • A classifier suggests categories with confidence and routes low-confidence items to a reviewer.
  • A neutral internal automation starts with fixtures, then adds credentials, permissions, and production scheduling only after the boundary is tested.
  • A customer-facing workflow keeps irreversible actions behind explicit approval until metrics show it is safe to automate further.