NOTQIN — OT Track

Augmented Intelligence and RAG

3 min read

Augmented Intelligence and RAG

Definition

Augmented Intelligence — using AI (specifically large language models) as a prosthesis for human cognition rather than as a replacement. The system retrieves, summarizes, drafts, and reasons over a knowledge corpus, but a human decides what to do.

Retrieval-Augmented Generation (RAG) — the dominant pattern (since Lewis et al. 2020) for grounding LLM responses in a specific knowledge base:

User question
    → retrieval over corpus (semantic search / vector DB / wikilink graph)
    → relevant chunks injected into LLM context
    → LLM generates answer grounded in retrieved chunks
    → response with citations

RAG makes the LLM honest about which knowledge its answer is based on, and lets you update the corpus without retraining the model.

Why it matters for NOTQIN

Reading the vault through Claude Code IS RAG-by-conversation. Every time Ahmed says “summarize what we know about CBAM” or “find inconsistencies between the Operating Thesis and the Path-to-Pilot,” the agent retrieves from the vault corpus + reasons over it.

This collapses two formerly expensive phases of the KM Lifecycle:

PhaseBefore LLMsWith LLM+RAG
Synthesis (Combination, see SECI Model)Hours of human reading + writingMinutes of prompt + verification
RetrievalManual search + skimSemantic / question-answering

Three formerly-hard phases remain hard:

  • Externalization (tacit → explicit) — humans still must articulate
  • Internalization (explicit → tacit) — humans still must learn
  • Socialization (tacit → tacit) — humans still must shadow each other

Strategic implication: the high-leverage KM investment is now in the three modes LLMs can’t shortcut.

The Smartex Agent angle

The eventual NOTQIN Agent is a RAG system applied to the factory knowledge corpus — UNS data + sector pack + historical events + maintenance records — that surfaces decision-grade answers (“which line is the bottleneck this week,” “which defect cluster needs an 8D first”).

This is what makes NOTQIN industrial intelligence vs. industrial dashboards: the AI layer is RAG-grounded in the customer’s actual operational corpus, not a generic LLM hallucinating.

Building a RAG over a vault — minimum viable stack

  1. Corpus — markdown vault + git versioning ✅ (already)
  2. Chunking strategy — by note, by section, or sliding window
  3. Embeddings — text → vector (OpenAI ada-002, Voyage AI, local sentence-transformers)
  4. Vector store — Qdrant, Pinecone, Weaviate, pgvector, or in-memory for small corpora
  5. Retrieval — top-k semantic similarity + optional metadata filters (e.g., by audience: tag)
  6. Generation — LLM with retrieved chunks in context + a system prompt that demands citations
  7. Evaluation — retrieval@k, faithfulness, answer relevance (RAGAS framework)

For NOTQIN, this is a future module. Today’s “RAG” is Claude Code’s filesystem access — slower per query but zero infrastructure cost.

Common mistakes

  • Letting the LLM answer without retrieval → hallucination
  • Retrieving with bad chunking → relevant info split across chunks, model misses context
  • No citations in responses → can’t audit; trust collapses on first bad answer
  • Treating RAG as a replacement for KM hygiene → garbage corpus = garbage answers
  • Using RAG to skip the thinking — answers are good for retrieval; decisions still need a human
  • Ignoring access control → RAG over a vault with mixed sensitivity leaks knowledge to wrong audience (this is why Portal Audience Routing 2026 matters)

See also

Graph View

Backlinks