Knowledge Base — Module 14.1

How This Lesson Fits the Module & Volume

The RAG pipeline is only as trustworthy as its knowledge base (KB)—the curated corpus of sources you are willing to retrieve from. Unlike a raw file dump, a KB has ownership, freshness SLAs, access rules, and quality gates. It feeds chunking and the vector store; indexing makes it searchable at scale.

Learning Objectives

By the end of this lesson, students should be able to:

Define a RAG knowledge base vs an ungoverned document pile.
List source types and ingestion patterns (batch, CDC, crawlers).
Set freshness, ownership, and deprecation policies.
Connect KB governance to metadata and ACLs.
Plan evaluation sets drawn from real KB questions.
Explain why garbage-in still yields fluent garbage-out with RAG.

Definition

A knowledge base for RAG is the governed collection of documents (and their derived chunks) that an organization designates as authoritative evidence for retrieval-augmented answers, including processes for update, access, and retirement.

KB Building Blocks

Element	Questions to answer
Sources	Which systems are in-scope? Who owns them?
Ingest	Batch nightly? Event-driven? Manual publish?
Quality	Deduped? PII scrubbed? Structured enough?
Access	Tenant / role visibility?
Lifecycle	How are obsolete docs removed from the index?

Curated KB

Owned pages
Review cadence
Higher trust

Raw lake

Everything indexed
Fast coverage
Noise & conflicts

Hybrid policy

Tier A/B sources
Boost curated
Common in prod

Code: Tiny KB Registry

from dataclasses import dataclass
from datetime import date

@dataclass
class KBDoc:
    doc_id: str
    uri: str
    owner: str
    tier: str          # "A" curated, "B" supplemental
    updated: date
    active: bool = True

class KnowledgeBase:
    def __init__(self):
        self.docs: dict[str, KBDoc] = {}

    def upsert(self, doc: KBDoc):
        self.docs[doc.doc_id] = doc

    def retire(self, doc_id: str):
        if doc_id in self.docs:
            self.docs[doc_id].active = False  # index job must delete vectors too

    def indexable(self) -> list[KBDoc]:
        return [d for d in self.docs.values() if d.active]

kb = KnowledgeBase()
kb.upsert(KBDoc("hr-pto", "s3://kb/hr/pto.md", "HR", "A", date(2026, 6, 1)))
print(len(kb.indexable()))

Strengths of a governed KB

Fewer conflicting answers
Clear owners for bad content
Safer ACL boundaries

Tradeoffs

Curation costs time
Coverage gaps hurt recall
Stale “active” flags poison RAG

Common Misconception

“If it is in SharePoint, it belongs in the KB.” Drafts, duplicates, and personal folders create contradictory evidence. Scope the KB deliberately; quarantine low-trust sources or mark them tier-B with lower boosts.

Knowledge Check

Short Answer: What makes a KB “governed”? Answer: Ownership, freshness, access, and retirement processes—not just files.
True/False: Retiring a doc requires removing its vectors too. Answer: True.
Multiple Choice: Tiered sources help by: (a) boosting trusted content, (b) removing GPUs, (c) CSS. Answer: (a).
Short Answer: Name an ingest pattern. Answer: Batch, CDC/event-driven, or crawl/publish (any).
True/False: RAG fixes contradictory source documents automatically. Answer: False.
Multiple Choice: Eval sets should come from: (a) real user questions on the KB, (b) only lorem ipsum, (c) DPI settings. Answer: (a).
Short Answer: Why track owners? Answer: Accountability when content is wrong or stale.
True/False: Personal draft folders are ideal tier-A sources. Answer: False.
Multiple Choice: Next lecture: (a) Indexing, (b) Vol. 1 only, (c) printers. Answer: (a).
Short Answer: How does metadata support the KB? Answer: Encodes tier, ACL, freshness, and provenance for retrieval.

Key Takeaways

A RAG KB is a governed evidence corpus, not a file dump.
Lifecycle and ACLs matter as much as embedding quality.
Garbage sources produce fluent wrong answers.
Next: Indexing.

Trainer’s Guide

Lab: Inventory 20 candidate sources; label A/B/exclude; justify three exclusions.

Discussion: Who has authority to publish into production RAG?

Recap: The knowledge base is RAG’s system of evidence. Continue with Indexing.