How do you build a workplace knowledge base for Super Intelligence? Start by defining authoritative sources, owners, version rules, permissions and the questions the system must answer. Then structure the knowledge so SI can retrieve current evidence, cite the source, expose uncertainty and route gaps back to the people who own the knowledge.
A workplace knowledge base is not simply a folder full of documents. It is an operating system for organisational knowledge: what is authoritative, what is current, who may see it, how information is updated, how older versions are retired, and how people or intelligent systems know when the answer is missing.
This article is part of the eduKateSG Workplace Super Intelligence Hub. It follows the Part IV foundation on Context Engineering for Workplace Super Intelligence and Turning SOPs into Useful SI Context. This page owns the knowledge-base architecture: how workplace knowledge becomes maintained, searchable, permission-aware and useful to SI.
The Short Answer
A useful workplace knowledge base needs six things: authoritative sources, clear ownership, metadata, permissions, version control and a feedback loop. Super Intelligence can retrieve and explain knowledge only as reliably as the underlying system distinguishes current truth from obsolete, duplicate or informal material.
Knowledge Base vs Document Folder
A folder stores files. A knowledge base tells users how those files relate to work. It answers which document is current, which applies to a given role, which is policy versus example, which supersedes another, and where unanswered questions should go.
This difference becomes critical when SI is added. A human expert may recognise an old PDF by memory; a retrieval system may treat it as equally valid unless the organisation supplies stronger signals.
The Five Knowledge Classes
- Policy: authoritative rules, constraints and approved positions.
- Procedure: how work should be performed.
- Reference: facts, specifications, definitions and stable background.
- Example: worked cases, templates and demonstrations of acceptable output.
- Experience: lessons, exceptions, decisions and local know-how that may still require validation.
Each class should be labelled because the same text can carry different authority. A template is not policy. A previous decision is not always a universal rule. A training example should not silently become the current operating standard.
Start With the Questions, Not the Repository
Before ingesting every file, list the recurring questions employees need answered. What do people search for? Which questions create delays? Which answers require current policy? Which questions repeatedly reach experts?
The question inventory defines the useful knowledge surface. It also reveals what the organisation does not know well enough to automate.
The Knowledge Question Inventory
- How do I perform this task?
- Which policy applies?
- What is the current approved process?
- Who owns this decision?
- What changed since the last version?
- What exception applies?
- Where is the source?
- What should I do when the standard path fails?
- Which template should I use?
- What evidence is required?
- Who can approve this action?
- What has the organisation learned from previous cases?
The Source-of-Truth Principle
For every important knowledge domain, identify the source that has institutional authority. The knowledge base can point to many documents, but it should make the canonical source explicit.
If two sources conflict, the system should not average them into a fluent answer. It should surface the conflict or prefer the designated authority according to the organisation’s rule.
The Knowledge Owner
Every critical knowledge domain needs an owner or owning function. Ownership means responsibility for currentness, retirement, gaps and material corrections.
The owner does not need to write every article. They need enough authority to decide which version is current and how unresolved questions become maintained knowledge.
The Knowledge Steward
In larger organisations, a steward can handle structure, tagging, quality and publishing while the domain owner retains authority over content. This separates editorial maintenance from policy ownership.
The Knowledge Consumer
Consumers include employees, managers, customers, agents and automated workflows. Different consumers may need different views and permissions even when they rely on the same underlying source.
The Canonical Record
A canonical record is the maintained version of knowledge that should be trusted for the task. It can be a document, database entry, wiki page, SOP or policy object.
The important point is not file format. It is that the organisation can say: this is the current authoritative record for this question.
The Version Rule
Every critical record should make version and currentness visible. Users and SI systems should be able to distinguish current, superseded, draft and archived states.
Do not rely on filenames such as final-v7-really-final.pdf. Version state belongs in maintained metadata or the source system.
The Retirement Rule
Obsolete knowledge should not remain equally searchable. Archive or clearly mark superseded material, preserve it where history matters, and prevent it from competing with current policy in ordinary retrieval.
Retirement is one of the most important SI knowledge controls because stale text can look perfectly plausible.
The Draft Rule
Draft content should be identifiable and excluded from authoritative answers unless the use case explicitly asks for draft material. A strategy draft and an approved procedure have different authority.
The Example Rule
Examples should demonstrate how a rule is applied without becoming the rule itself. Label context, date and whether the example remains representative.
The Exception Rule
Exceptions should be documented with scope. A one-off approval should not silently become a general entitlement. Knowledge entries should state who authorised the exception and whether it creates a reusable rule.
The Decision Rule
Important organisational decisions can become durable knowledge when their rationale, owner, date, assumptions and applicability are recorded.
A decision log helps future SI systems explain not only what is true but why the current state exists.
The Knowledge Object
A useful knowledge object is small enough to retrieve precisely and rich enough to stand on its own. It might be a policy section, procedure step, FAQ, decision record or reference entry.
Knowledge objects should preserve links back to the larger authoritative source so retrieval does not strip away necessary context.
Metadata Is Operating Context
Metadata tells the system how to interpret the content. Useful fields include title, owner, domain, version, effective date, expiry or review date, audience, confidentiality, source type, authority level and superseded-by link.
Without metadata, SI must infer too much from prose.
Core Metadata Fields
- Title
- Domain
- Owner
- Status
- Version
- Effective date
- Review date
- Audience
- Confidentiality
- Authority level
- Source type
- Keywords or concepts
- Related procedures
- Supersedes
- Superseded by
- Last material change
Authority Metadata
A knowledge base should distinguish mandatory policy, recommended practice, informal guidance and example material. This helps SI avoid presenting advice as obligation.
Audience Metadata
Some knowledge applies only to certain countries, departments, products, customer types or employee roles. Audience metadata prevents a generally relevant document from being applied too broadly.
Confidentiality Metadata
Public, internal, confidential, restricted and privileged material should not be mixed without access controls. Metadata can support retrieval filtering, but technical permissions should enforce the boundary.
Currentness Metadata
Knowledge decays at different rates. Product pricing, access rules and operational runbooks may change quickly; stable definitions may not. Review dates should reflect the domain’s real rate of change.
The Knowledge Taxonomy
A taxonomy provides stable categories for domains, processes, products, roles and source types. The goal is not to classify every sentence perfectly. It is to give people and retrieval systems enough structure to narrow search and apply the right context.
Taxonomy by Domain
- Operations
- Sales
- Marketing
- Finance
- Human Resources
- Legal
- IT
- Security
- Customer Support
- Procurement
- Product
- Learning and Development
Taxonomy by Workflow
Domain alone is often too broad. Add workflow tags such as onboarding, refund, incident response, procurement review, customer escalation or reporting.
Workflow metadata is particularly useful for context engineering because the same department may contain very different decision rules.
Taxonomy by Role
Role-based tags help determine which knowledge is relevant and which permissions apply. A manager, specialist and frontline user may need different views of the same process.
Taxonomy by Knowledge Type
Tag whether content is policy, procedure, reference, example, template, decision record, troubleshooting guide or FAQ. This reduces authority confusion.
The Folder Structure Is Not the Knowledge Architecture
Folders can help people browse, but one document may relate to several workflows and roles. Metadata and links often create a more flexible knowledge model than forcing everything into one hierarchy.
Use folders for navigation where helpful, but do not make folder location the only meaning available to SI.
The Search Surface
Employees should be able to search by natural language, keyword, title, category and filters. Different search modes serve different tasks.
SI can add semantic retrieval and synthesis, but exact title and keyword search remain valuable for known-item lookup.
Keyword Search vs Semantic Search
Keyword search is strong when users know the exact term, identifier or phrase. Semantic search is strong when the user describes a concept in different language.
A workplace knowledge base should not force one method to replace the other. Hybrid search often produces better coverage.
Retrieval Is Not the Same as Answering
A system can retrieve relevant evidence without generating a single answer. In high-consequence workflows, showing the source may be preferable to synthesising a conclusion.
The knowledge base should support both navigation and answer generation.
The Citation Rule
When SI synthesises an answer from internal knowledge, material claims should link back to the supporting source where practical. This allows the user to inspect authority, date and context.
Citation is not proof by itself. The cited source must actually support the claim.
The Abstention Rule
When the knowledge base does not contain enough current evidence, the system should say so and route the gap rather than invent an answer.
Abstention is a sign of a healthy knowledge system because it protects the boundary between documented knowledge and inference.
The Gap Queue
Unanswered or weakly supported questions should enter a knowledge-gap queue. The domain owner can decide whether the answer belongs in documentation, policy, training or a case-specific human response.
This turns search failure into organisational learning.
The Query Log
Search and question logs reveal what employees actually need. Repeated zero-result queries, reformulations and escalations can guide documentation priorities.
Use query data with appropriate privacy controls.
The Knowledge Feedback Loop
- User asks question.
- System retrieves or fails.
- User verifies or corrects.
- Failure is classified.
- Owner receives recurring gap.
- Source is improved.
- Retrieval is retested.
- Future users receive stronger context.
The Knowledge Change Loop
When a source changes materially, update metadata, identify affected procedures, invalidate stale derived material and retest important questions.
A knowledge base should not assume that changing one policy page automatically repairs every downstream summary or template.
The Derived-Knowledge Problem
Summaries, FAQs, templates and training material may be derived from a canonical source. When the source changes, derived content can become stale.
Track relationships so the organisation knows what may need review after a material update.
The Duplication Problem
Duplicate knowledge creates uncertainty. The same answer may appear in wiki pages, email, SOPs and training decks. The solution is not necessarily deleting everything, but identifying which copy is authoritative and which are derivative.
The Contradiction Problem
When two authoritative-looking sources disagree, surface the conflict to the owner. Do not ask SI to silently choose unless the source hierarchy already defines the rule.
The Missing-Owner Problem
Knowledge without an owner decays. If nobody is responsible for review, the system cannot distinguish neglected content from intentionally stable content.
The Over-Ingestion Problem
Connecting every file can reduce retrieval quality because irrelevant, obsolete and low-authority text competes for attention. Start with high-value domains and curated sources.
The Under-Ingestion Problem
A corpus that excludes the documents employees actually rely on forces them to work around the system. Observe real workflows and include the evidence that matters.
The Chunking Problem
Long documents may need to be split into retrievable units. Chunks should preserve enough surrounding context to make the passage meaningful. Breaking by arbitrary character count can separate rules from qualifiers.
Where possible, use natural structure such as headings, sections, clauses and procedure steps.
The Context Window Problem
Even when the knowledge base stores a large corpus, the model may use only a subset at one time. Retrieval quality determines which evidence enters the working context.
This is why better curation can matter more than simply adding more documents.
The Permission-Aware Retrieval Rule
A user should retrieve only knowledge they are allowed to access. Permission filtering should happen before content reaches the model where possible.
Generated answers should not reveal restricted information indirectly.
The Least-Knowledge Principle
The system should receive the minimum knowledge necessary for the task. Broad access increases exposure and can reduce relevance.
The Role-Based View
A manager may see policy, rationale and approval rules. A frontline employee may see the required procedure. A customer-facing user may see externally safe language. One source can support multiple role-specific views.
The Location-Based View
Global organisations may need country or region-specific policy. Metadata should prevent one jurisdiction’s rules from appearing as universal.
The Product-Based View
Different products, customer tiers or service levels may have different procedures. The knowledge base should include applicability conditions.
The Time-Based View
Some knowledge applies only during a campaign, incident, transition or temporary exception period. Effective and expiry dates help SI avoid applying temporary rules indefinitely.
The Knowledge Base Maturity Ladder
Stage 1 — Scattered
Knowledge lives in documents, email and memory with little source authority.
Stage 2 — Indexed
Core documents are collected and searchable.
Stage 3 — Governed
Owners, versions, permissions and source types are explicit.
Stage 4 — SI-Readable
Metadata, chunking, retrieval and citations support reliable machine use.
Stage 5 — Learning
Queries, gaps, corrections and source changes continuously improve the corpus.
Stage 6 — Operational
Multiple workflows use the knowledge base as shared infrastructure with monitoring, change control and clear ownership.
The Minimum Viable Knowledge Base
A small team can begin with one approved folder or wiki section, a canonical source list, simple metadata, named owners and a query-gap process.
Do not wait for enterprise architecture if one bounded knowledge domain can already be improved.
The Team Knowledge Base
A team knowledge base should answer recurring operational questions, preserve decisions and support onboarding. It can remain simple if ownership and currentness are strong.
The Department Knowledge Base
A department needs stronger taxonomy, permissions and cross-workflow relationships because several teams may interpret the same policy differently.
The Enterprise Knowledge Base
Enterprise scale adds identity, federated sources, legal and regional variation, broader permissions, shared retrieval infrastructure and lifecycle governance.
Enterprise does not mean one giant repository. Several governed sources can remain federated while appearing through one search or SI interface.
Federated Knowledge
A federated system retrieves from multiple authoritative sources without forcing all content into one physical repository. This can preserve local ownership and specialised systems.
The interface should still expose which source supplied each claim.
Centralised Knowledge
Centralisation can simplify governance and retrieval but may create migration cost and one large failure radius. Choose based on organisational reality rather than architecture fashion.
The Knowledge Base and SOPs
SOPs are one important source class. They become more useful to SI when steps, preconditions, exceptions and authority are explicit. See How to Turn SOPs into Useful Super Intelligence Context.
The Knowledge Base and Documentation
Documentation workflows create and maintain the source material. The knowledge base organises authority, retrieval and lifecycle around that documentation.
See How to Use Super Intelligence for Documentation.
The Knowledge Base and Context Engineering
Context engineering chooses which subset of knowledge enters a particular task. The knowledge base provides the maintained pool; context engineering provides task-specific selection.
The Knowledge Base and Enterprise Search
Enterprise search is the access layer over organisational knowledge. A strong knowledge base makes search more accurate because authority, metadata and currentness are explicit.
The Knowledge Base and Agents
Agents need especially strong knowledge boundaries because they may act on retrieved information. A wrong source can become a wrong external action.
Tool-using workflows should refresh material state before consequential action and preserve source evidence.
The Knowledge Base and Human Review
Human reviewers should be able to inspect the source behind important answers. The system should not force reviewers to trust a generated summary without provenance.
The Knowledge Base and Learning
New employees, existing staff and SI systems all benefit from the same maintained knowledge. A strong knowledge base reduces repeated expert interruption and makes onboarding more consistent.
The Knowledge Base and Organisational Memory
Decisions, lessons and procedures should survive employee turnover. The knowledge base is one part of institutional memory, but only if updates and ownership remain active.
The Knowledge Audit
- Which sources are authoritative?
- Which are obsolete?
- Which are duplicated?
- Which lack owners?
- Which have unclear audience?
- Which contain sensitive information?
- Which are heavily searched?
- Which produce frequent zero-result queries?
- Which need review after policy change?
- Which should be archived?
The Knowledge Base Metrics
- Time to useful source
- Search success rate
- Zero-result rate
- Answer citation coverage
- Unsupported-answer rate
- Source freshness
- Duplicate rate
- Gap-resolution time
- Owner response time
- User clarification rate
- Knowledge reuse
- Expert interruption reduction
The 30-Day Knowledge Base Build
Week 1 — Questions and Sources
Inventory recurring questions and identify the small set of authoritative sources that answer them.
Week 2 — Ownership and Metadata
Assign owners, status, version, audience and confidentiality. Retire obvious obsolete material.
Week 3 — Retrieval and Testing
Connect the corpus to approved search or SI tools. Test known-answer, known-gap and conflicting-source questions.
Week 4 — Feedback Loop
Capture zero-result queries, unsupported answers and user corrections. Create the gap queue and owner workflow.
The Knowledge Base Failure Catalogue
- Stale source outranks current source
- Draft treated as policy
- Example treated as rule
- Restricted content retrieved incorrectly
- Conflicting documents synthesised silently
- Answer provided with no source
- Source cited but does not support claim
- Unanswered question guessed
- Duplicate content fragments ranking
- Owner absent
- Review date ignored
- Derived content not updated after source change
The Knowledge Base Repair Rule
Repair source authority before improving prompting. If the corpus is wrong, a better prompt only produces a more polished wrong answer.
The Knowledge Base Simplification Rule
Delete or archive unnecessary duplication. Fewer stronger sources can outperform a huge unmanaged corpus.
The Knowledge Base Permission Rule
The system should not see more than the user needs. Permission-aware retrieval belongs before generation where possible.
The Knowledge Base Currentness Rule
Material answers should reflect sources current enough for the task. Time-sensitive knowledge should have explicit refresh or expiry behaviour.
The Knowledge Base Gap Rule
A missing answer should become work for a knowledge owner, not an invitation for SI to invent organisational policy.
The Knowledge Base Citation Rule
Every material claim should be traceable to a source the user can inspect when consequence justifies it.
The Knowledge Base Transfer Rule
Before copying one department’s knowledge architecture, compare authority, permissions, workflow and currentness needs. The same taxonomy may not fit every domain.
What This Article Owns
This page owns the workplace knowledge-base architecture: sources, authority, ownership, metadata, permissions, versioning, retrieval, feedback and lifecycle.
It does not own document connectivity itself or enterprise search behaviour; those are the next two articles in the series.
Frequently Asked Questions
What is a workplace knowledge base for Super Intelligence?
It is a maintained system of organisational knowledge with clear authority, ownership, metadata, permissions and retrieval so people and SI can find current evidence reliably.
Do we need one repository?
No. A federated knowledge base can retrieve from several authoritative systems while preserving local ownership.
Should we upload every company document?
No. Start with valuable, current and governed sources. Over-ingestion can reduce retrieval quality and increase exposure.
How do we prevent stale answers?
Use version state, effective dates, owners, retirement rules, currentness checks and regression questions after material source changes.
What if two sources conflict?
Surface the conflict or follow an explicit source hierarchy. Do not let SI silently blend incompatible rules.
How should permissions work?
Users and agents should retrieve only the knowledge required and allowed for their role. Technical access controls should enforce the boundary.
What is the best first knowledge-base project?
A bounded domain with frequent recurring questions, clear source owners and a manageable corpus.
How do we know the knowledge base is working?
Measure faster source discovery, lower zero-result and unsupported-answer rates, strong citation coverage and reduced repeated expert interruption.
What comes next?
Continue to How to Connect Super Intelligence to Workplace Documents and Files, which covers the technical and operational access layer between SI and the files employees actually use.
The Core Knowledge Rule
Super Intelligence should not become the organisation’s memory by itself. It should become a reliable interface to maintained organisational memory.
The knowledge base is strong when current truth is easy to find, old truth is easy to recognise, missing knowledge becomes visible and every important answer can return to an accountable source.
Build the Knowledge Base Around Decisions and Workflows
The most useful workplace knowledge base is not organised only around topics. It should also reflect the decisions and workflows employees actually perform. A sales team does not merely need “pricing information”; it needs to know which price applies to which customer, which discount authority exists, what exception process applies and where current approval state is recorded.
This workflow orientation makes knowledge more useful to Super Intelligence because the system can retrieve information in relation to the task being performed rather than returning generic background material.
Knowledge for a Task vs Knowledge for Reference
Reference knowledge explains what something is. Task knowledge explains what to do now. A product specification can be reference knowledge; a refund procedure is task knowledge. A policy statement can establish a rule; a runbook explains how to apply it.
Both matter, but they should be distinguishable. SI answering a procedural question should not rely only on a descriptive reference document when a current procedure exists.
The Knowledge Hierarchy
- Source of truth: authoritative record or policy.
- Procedure: current operating method.
- Decision rule: threshold, approval or eligibility logic.
- Reference: supporting facts, definitions and specifications.
- Example: worked demonstration of application.
- Derived summary: compressed or role-specific explanation.
- Informal knowledge: local experience that may still need validation.
The hierarchy helps the retrieval layer prefer stronger evidence and makes it easier to explain why one source outranks another.
Knowledge Authority Is Context-Specific
One source can be authoritative for one question and insufficient for another. A CRM may be authoritative for current customer status but not for legal terms. A policy manual may define approval rules but not show whether a specific request was approved.
The knowledge base should therefore model authority by domain and object rather than assuming one repository is globally authoritative.
Knowledge Freshness Is Also Context-Specific
Some knowledge becomes stale within hours; other knowledge changes rarely. A customer entitlement, inventory level or deployment state may need near-real-time access. A stable definition may remain valid for years.
Use freshness classes so retrieval can decide when a cached or indexed copy is adequate and when a live system lookup is required.
Freshness Classes
- Live: must be checked from the current system before action.
- Daily: suitable for regularly refreshed operational state.
- Versioned: valid until an explicit new version replaces it.
- Periodic: reviewed on a scheduled cadence.
- Stable reference: changes rarely but still has an owner.
Knowledge Base Architecture: Source Layer
The source layer contains the systems where knowledge originates. These may include document repositories, wikis, databases, ticket systems, CRM, code repositories, policy systems and specialist applications.
Do not copy everything into a new location merely to make it AI-readable. Where possible, preserve the source of truth and build retrieval around it.
Knowledge Base Architecture: Index Layer
The index layer helps search and retrieval locate relevant knowledge. It may include keyword indexes, semantic vectors, metadata filters or other retrieval structures.
The index is derived infrastructure. It should be rebuildable from the authoritative sources and should not become a silent alternate record.
Knowledge Base Architecture: Retrieval Layer
The retrieval layer selects which evidence to bring into the current task. It considers query meaning, metadata, permissions, currentness and source authority.
This is where many SI systems succeed or fail. A strong model given the wrong evidence can still produce a confident bad answer.
Knowledge Base Architecture: Answer Layer
The answer layer may simply show documents, or it may generate a synthesis. High-consequence tasks may prefer source-first interfaces that show evidence before explanation.
When generated synthesis is used, the system should preserve citations and uncertainty.
Knowledge Base Architecture: Feedback Layer
The feedback layer captures wrong answers, missing sources, poor ranking, unanswered questions and user corrections. These signals should reach the people who can repair the knowledge system.
Knowledge Base Architecture: Governance Layer
Governance controls who owns content, who may access it, how changes are approved, how obsolete content is retired and how incidents are handled.
Without governance, the technical system can become a faster route to organisational ambiguity.
The Known-Answer Test Set
Create a small set of questions whose answers are known and supported by current sources. Use these to test retrieval after changes to indexing, models or source structure.
The test set should include easy lookups, semantic paraphrases and questions whose answer requires combining more than one source.
The Known-Gap Test Set
Also include questions the knowledge base intentionally cannot answer. The correct result is abstention, escalation or a statement that the information is undocumented.
This is one of the strongest tests against hallucinated organisational knowledge.
The Conflict Test Set
Include a few historical or deliberately conflicting sources and verify that the system follows the current authority rule rather than blending them.
The Permission Test Set
Test whether users with different roles receive only the information they are authorised to access. Permission-aware retrieval should prevent restricted content from entering the model context.
The Currentness Test Set
For time-sensitive sources, verify that the system retrieves the latest state and does not rely on a stale indexed copy when live data is required.
The Citation Test
For each answer, check whether the cited passage actually supports the claim. Citation quality is not merely the presence of a link.
The Retrieval Precision Problem
If retrieval returns too many loosely related passages, the model receives noise. Improve metadata, source curation, query rewriting or ranking rather than simply expanding the context window.
The Retrieval Recall Problem
If relevant evidence is missing, the system may answer from partial context. Review chunking, indexing, synonyms, permissions and source coverage.
The Hybrid Retrieval Pattern
Keyword and semantic search complement one another. Keyword search is strong for exact IDs, product names and legal phrases. Semantic retrieval is strong for paraphrased concepts and natural-language questions.
A hybrid strategy can combine both before reranking results for the task.
The Metadata Filter Pattern
Filters such as country, department, product, audience, version and source type can dramatically improve retrieval quality. They reduce the candidate pool before semantic similarity is applied.
The Reranking Pattern
A retrieval system can first gather a broad candidate set and then rerank for relevance, authority or task fit. This is especially useful in large heterogeneous corpora.
The Query-Rewriting Pattern
Users may ask vague questions. SI can rewrite the query into more precise search terms, expand acronyms or identify the workflow context before retrieval.
The rewritten query should not change the user’s underlying intent.
The Multi-Source Synthesis Pattern
Some answers legitimately require several sources: a policy, customer state and approval threshold. The system should show which source supports which part of the answer.
Do not hide contradictions or missing pieces inside one seamless paragraph.
The Source-First Pattern
For high-stakes use, show the retrieved source passages first and make synthesis secondary. This helps users maintain direct contact with authoritative material.
The Answer-First Pattern
For low-risk, repetitive knowledge work, a concise answer followed by citations may reduce friction. The user can drill into evidence when needed.
The Role-Specific Answer Pattern
A frontline user may need a procedure. A manager may need the policy and exception logic. A specialist may need deeper references. Role-specific answer design can reduce overload while preserving common source truth.
The Knowledge Base for Onboarding
New employees repeatedly ask where information lives and how processes work. A strong SI knowledge base can make onboarding interactive while preserving citations to approved sources.
The unanswered questions from new hires are also valuable signals about documentation gaps.
The Knowledge Base for Customer Support
Support teams benefit from product information, policy, troubleshooting, entitlement rules and escalation paths. The system should combine knowledge with current customer state without confusing static documentation with live account facts.
The Knowledge Base for Sales
Sales teams need current product, pricing, case studies, objections, competitor notes and commercial rules. Sensitive pricing and discount authority should remain permission-aware.
The Knowledge Base for Finance
Finance knowledge includes policies, chart-of-account guidance, close procedures, controls and reporting rules. Current financial figures should remain in systems of record rather than in static knowledge documents.
The Knowledge Base for HR
HR knowledge can include policies, benefits, onboarding and manager guidance. Personal employee data should remain in appropriately controlled systems and should not be mixed casually into general knowledge retrieval.
The Knowledge Base for Legal
Legal knowledge requires strong source authority, jurisdiction, version and privilege controls. SI can help locate clauses and internal positions, but legal interpretation remains under appropriate professional processes.
The Knowledge Base for Engineering
Engineering knowledge includes architecture decisions, runbooks, coding standards, repository documentation and incident lessons. Code and production state may require live tool access rather than static documents.
The Knowledge Base for Operations
Operations knowledge often changes quickly. Runbooks, thresholds, incident procedures and escalation contacts need strong currentness and ownership.
The Knowledge Base for Education and Training
Learning organisations can store curriculum, competency standards, examples, teaching resources and assessment rubrics. The knowledge base should distinguish current standards from illustrative material.
The Knowledge Lifecycle
- Create or approve source.
- Assign owner and metadata.
- Publish as current.
- Index for retrieval.
- Use in workflows.
- Collect gaps and corrections.
- Review after change or cadence.
- Revise or supersede.
- Archive or retire.
- Retest important questions.
This lifecycle turns knowledge maintenance into an operating process rather than a one-time migration project.
The Knowledge Review Cadence
Review frequency should reflect risk and change rate. Critical operational content may need event-based review. Stable reference content can use longer intervals.
Do not force every page into the same annual review cycle if the underlying domains behave differently.
The Event-Triggered Review
Material policy changes, product releases, incidents, regulatory changes and system migrations should trigger review of affected knowledge immediately.
The Search-Triggered Review
Repeated failed searches or user corrections can trigger content review even when the scheduled review date is far away.
The Incident-Triggered Review
If an incident involved wrong or stale knowledge, update the source and test affected retrieval patterns as part of recovery.
The Knowledge Gap SLA
For important domains, define how quickly recurring knowledge gaps should be resolved. An unanswered question that blocks customer support may deserve faster ownership than a rare reference query.
The Expert-Interruption Metric
Measure how often specialists are asked routine questions that the knowledge base should handle. A reduction can demonstrate value while preserving experts for exceptions.
The Time-to-Source Metric
Measure how long it takes users to reach the authoritative source, not only how fast the model generates an answer.
The Trust Metric
Ask whether users verify and rely on the knowledge system appropriately. High usage with frequent manual double-checking may indicate weak trust or weak evidence presentation.
The Knowledge Coverage Metric
Estimate what proportion of important recurring questions have maintained answers. Coverage should focus on business relevance, not raw document count.
The Knowledge Debt Backlog
Treat stale, ownerless, conflicting and missing knowledge as debt. Prioritise repair according to workflow impact, frequency and consequence.
This creates a more useful backlog than simply counting old pages.
The Knowledge Owner Review
Owners should see queries, corrections and source usage for their domain. This helps them improve the content based on real demand.
The Knowledge Council
Large organisations may use a cross-functional forum to define common metadata, permissions, retention and search standards while leaving content authority with domain teams.
Central governance should enable consistency without taking ownership away from the people closest to the work.
The Knowledge Base Pilot
Choose one domain with frequent questions and clear sources. Build a small corpus, define ownership, test known answers and gaps, then measure time-to-source and expert interruption.
Do not begin with the whole company.
The Knowledge Base Scale Test
Before expanding, test whether metadata, permissions, ownership and retrieval quality still work as source count and user diversity increase.
The Knowledge Base Continuity Test
If the SI layer is unavailable, users should still be able to find the underlying sources. The knowledge base should not make organisational knowledge inaccessible without the model.
The Knowledge Base Independence Test
A new employee or another team member should be able to use the system without the original builder explaining hidden source rules.
The Knowledge Base Security Test
Test whether restricted content can leak through generated summaries, related results or inferred answers. Permission boundaries must apply to the full retrieval and generation path.
The Knowledge Base Change Test
Update one material source and verify that retrieval reflects the change, stale derived content is identified and important regression questions still pass.
The Knowledge Base Failure Drill
Simulate stale policy, conflicting versions, missing source and permission mismatch. The system should surface uncertainty or deny access rather than fabricate certainty.
The Knowledge Base Migration Rule
When moving repositories, preserve authority, metadata, links and version relationships. A migration that copies text but loses meaning can reduce retrieval quality.
The Knowledge Base Vendor Rule
Do not let a vendor-specific feature become the only place source authority is encoded. Keep ownership and knowledge rules portable at the organisational level.
The Knowledge Base Human-Agency Rule
Users should be able to inspect sources, challenge answers, report gaps and understand why the system returned a result. The interface should not hide organisational knowledge behind an opaque assistant.
The Knowledge Base Standard
Current source + clear owner + explicit authority + permission-aware retrieval + visible evidence + gap feedback is the minimum grammar of a workplace knowledge base built for Super Intelligence.
When those elements are present, SI can help employees reach useful organisational knowledge faster without becoming an unaccountable source of truth.
Production Knowledge Base Layer 1 — Ingestion
Ingestion should be selective and governed. The objective is not to collect as many documents as possible; it is to make the right knowledge eligible for retrieval. Every ingestion path should preserve source identity, owner, version, scope and permissions.
Automated ingestion is valuable only when the upstream repository itself is governed. Otherwise it can import drafts, duplicates and stale files at machine speed.
Production Layer 2 — Normalisation
Documents arrive as PDFs, pages, spreadsheets, slides, code, tickets and database records. Normalisation converts them into forms that retrieval can index while preserving the original source.
Do not destroy structure that carries meaning. Tables, headings, lists, code blocks and metadata should survive when they matter to interpretation.
Production Layer 3 — Chunking
Chunk by semantic unit: section, procedure step, definition, clause, FAQ, decision or table. The chunk should be independently understandable enough for retrieval but still linked to its source context.
Chunk size should follow the document’s meaning rather than one universal character count.
Production Layer 4 — Indexing
Index exact terms, semantic representations and metadata. Exact identifiers, codes and names need lexical support; natural-language paraphrases benefit from semantic retrieval.
Production Layer 5 — Permission Filtering
Filter by authorised scope before the retrieved material reaches the model. User role, team, project membership, confidentiality and geography can all matter.
A secure interface does not rely on the model to hide content it should never have received.
Production Layer 6 — Retrieval
Retrieval should combine the user’s task, role, current state and query with authority and metadata. The highest semantic match is not necessarily the correct source.
Production Layer 7 — Reranking
When several candidate passages are found, reranking can favour sources that are canonical, current, in scope and directly relevant. Authority should be part of the ranking logic.
Production Layer 8 — Context Assembly
The selected knowledge should be assembled into a task-specific context packet. Avoid flooding the model with the top twenty results when three authoritative sections answer the question.
Production Layer 9 — Generation or Action
The model uses the context to answer, draft, classify, compare or plan. If the workflow can act, the context should include the relevant procedure and permission boundaries.
Production Layer 10 — Verification
Material claims should be traceable to the retrieved source. Structured values can be checked deterministically. High-consequence conclusions may require human or professional review.
Production Layer 11 — Feedback
Wrong answers should produce diagnostic feedback: wrong source, stale source, missing source, scope error, interpretation error or model failure. Generic thumbs-down data is less useful than cause.
Production Layer 12 — Maintenance
Source owners review, update, merge, supersede and retire knowledge. Technical owners maintain retrieval and permissions. The knowledge base remains useful only if maintenance is an operating process.
The Knowledge Base Source Lifecycle
- Draft
- Reviewed
- Approved
- Published
- Current
- Superseded
- Archived
- Retired
Not every source needs all states, but high-impact knowledge should distinguish current published content from drafts and history.
The Draft Exclusion Rule
Drafts should not be eligible for authoritative answers unless the workflow explicitly asks for draft material. This prevents recent but unapproved content from outranking policy.
The Superseded Rule
Superseded content can remain searchable for historical questions but should be filtered from current operational retrieval.
The Archive Rule
Archived content is preserved for record or context but is not part of the active knowledge surface.
The Knowledge Base Version Graph
Where versions matter, link current and superseded sources. This allows the system to answer “what changed?” and to use the correct version for historical cases.
The Knowledge Base Decision Log
Some organisational knowledge comes from decisions rather than policy. A decision record should state who decided, when, evidence, rationale, scope, duration and what would trigger review.
Time-limited decisions should not become permanent policy by accident.
The Knowledge Base Exception Register
Repeated exceptions can become durable knowledge if the organisation decides they should. Record exception, decision owner, reason, scope and expiry.
This helps SI distinguish a legitimate exception from an informal workaround.
The Knowledge Base Glossary
Maintain operational definitions and synonyms for terms that affect work. A glossary is one of the highest-leverage knowledge objects because it improves both human coordination and model interpretation.
The Knowledge Base FAQ Layer
FAQs are useful for repeated low-risk questions, but answers should link back to authoritative sources. The FAQ is a navigation and explanation layer, not a separate policy source.
The Knowledge Base Template Layer
Templates can help SI produce outputs that fit workplace standards. Label which fields are mandatory, which are examples and which require human judgment.
The Knowledge Base Example Library
Approved examples show style, structure and application. Keep them separate from policy and include why the example is considered good.
The Knowledge Base Negative Example Library
Negative examples can teach boundaries: unsupported customer promise, incomplete incident handoff, wrong document version, weak report or misleading claim.
They are especially useful in evaluation and reviewer training.
The Knowledge Base Case Library
Representative historical cases can support learning and testing, but personal or confidential information should be handled appropriately. Case examples should not silently become universal rules.
The Knowledge Base Structured Data Layer
Exact values, inventories, account status and transactions belong in structured systems of record. The knowledge base should know how to locate and explain them rather than duplicate stale copies.
The Knowledge Base Document Layer
Policies, SOPs, manuals and reports belong in document systems. Retrieval should preserve structure and version.
The Knowledge Base Knowledge-Graph Layer
Where useful, concepts, entities and relationships can be represented explicitly: product governed by policy, system owned by team, procedure uses form, decision affects project.
A full graph database is optional; the conceptual graph can begin with links and metadata.
The Knowledge Base Search Interface
Users should be able to search and ask natural-language questions. The interface should show sources, allow source opening and make gaps or conflicts visible.
The Knowledge Base Admin Interface
Owners need tools to review stale content, duplicates, gaps, feedback, permission issues and source health. A user-facing chatbot alone is not enough.
The Knowledge Base Health Dashboard
- Sources by domain
- Sources without owner
- Sources overdue for review
- Superseded content still retrieved
- Top unanswered questions
- Top correction categories
- Permission failures
- Retrieval success by workflow
- High-risk sources without tests
- Recently changed canonical sources
The dashboard should support maintenance decisions, not become vanity analytics.
Knowledge Coverage
Coverage asks whether the knowledge base contains enough current authority for the workflows it claims to support. It does not mean documenting everything.
Measure coverage against high-value questions and procedures.
Knowledge Precision
Precision asks whether retrieved sources are actually relevant and applicable. A large corpus can have high coverage and poor precision.
Knowledge Freshness
Freshness asks whether time-sensitive sources are current enough for the task. High-risk stale knowledge deserves stronger monitoring.
Knowledge Ownership Health
Track high-value sources without active owners. Ownership gaps are organisational risk, not simply metadata defects.
Knowledge Conflict Rate
Track cases where sources disagree. High conflict can indicate duplicated authority or unresolved policy.
Knowledge Gap Rate
Track questions that correctly return no answer. The goal is not zero gaps; the goal is to know which gaps deserve documentation.
Knowledge Gap Closure Time
Measure how long important repeated gaps take to reach an owner and become maintained knowledge.
Knowledge Correction Time
When users identify wrong or stale knowledge, measure how quickly the source is corrected and how dependent workflows are retested.
Knowledge Base Evaluation Set Design
- Common known-answer questions
- Rare known-answer questions
- Known gaps
- Wrong-scope questions
- Permission-restricted questions
- Conflicting-source questions
- Version-history questions
- Procedure questions
- Definition questions
- Current-state questions requiring external systems
This mix evaluates the whole knowledge architecture rather than only average retrieval accuracy.
Evaluation: Source Selection
Did the system retrieve the right canonical source? A correct final answer from the wrong source can still be operationally fragile.
Evaluation: Scope
Did the system apply the correct geography, product, role and time period?
Evaluation: Citation Fidelity
Does the cited passage actually support the generated claim? Citation presence alone is not enough.
Evaluation: Completeness
Did the answer preserve critical qualifiers, exceptions and thresholds from the source?
Evaluation: Gap Handling
Did the system abstain when the knowledge base lacked authority?
Evaluation: Conflict Handling
Did the system surface conflict instead of synthesising incompatible rules?
Evaluation: Permission Safety
Could an unauthorised user retrieve or infer restricted knowledge?
Evaluation: Currentness
Did the system use the source version applicable to the case date and current workflow?
Evaluation: Workflow Usefulness
Did the retrieved knowledge reduce search, rework or decision preparation in the actual workflow?
The Knowledge Base Permission Model
Use the organisation’s existing identity and access model where possible. Knowledge-base permissions should not become a parallel unmanaged security system.
Role, group, project and document-level permissions can all matter.
Permission Inheritance
If a folder or repository grants access, the SI system should respect that inheritance. When content is moved or ownership changes, retrieval access should update accordingly.
Permission Revocation
Access removal should propagate quickly. An employee who loses access to a source should not retain indirect access through the assistant.
Permission Testing
Include adversarial questions that attempt to retrieve restricted content through paraphrase, summary or inference.
The Least-Knowledge Principle
An SI workflow should receive only the knowledge needed for the task. Broad access can increase both privacy risk and retrieval noise.
The Need-to-Know Context Packet
Build task-specific context after applying permissions and scope. This is the knowledge equivalent of least privilege.
The Knowledge Base and Privacy
Durable knowledge should generally avoid unnecessary personal information. When a workflow needs personal data, retrieve it from the appropriate system for the authorised task.
The Knowledge Base and Confidential Business Information
Pricing strategy, legal advice, security architecture and executive material may require restricted domains. SI access should mirror the organisation’s legitimate access model.
The Knowledge Base and External Content
External websites, vendor documents and customer submissions can be useful references but should not automatically receive internal authority.
Label external provenance and verify material claims before they influence decisions.
The Knowledge Base and Web Search
Internet search can complement internal knowledge, but the system should distinguish public external information from internal policy. One should not silently override the other.
The Knowledge Base and Model Knowledge
Generic model knowledge can explain concepts, but local workplace rules should come from the organisational knowledge base. The interface should make that distinction clear when consequence matters.
The Knowledge Base and Memory
Memory can preserve personal preferences and continuity, but it should not become hidden policy. Current canonical sources should override stale remembered facts.
The Knowledge Base and Context Windows
Models have finite context. Retrieval should select the few most relevant sources rather than assume more tokens always improve performance.
The Knowledge Base and Prompt Design
Prompts should specify the task and how to use sources, not duplicate whole policies. This keeps organisational knowledge maintainable outside model-specific instructions.
The Knowledge Base and Tool Use
Tool-using agents should retrieve the procedure and current state before acting. Action permissions should depend on both source guidance and user authority.
The Knowledge Base and World Return
Knowledge can guide action, but completion should be confirmed by the destination system. Do not let the knowledge layer become a fictional record of what happened.
The Knowledge Base Incident Review
When an SI workflow fails, determine whether the wrong source was indexed, retrieved, interpreted or applied. Preserve source IDs and versions for important cases.
The Knowledge Base Change Impact Map
When a canonical source changes, identify dependent workflows, evaluation cases and agents. High-impact knowledge changes can require retesting before production autonomy continues.
The Knowledge Base Dependency Map
Track which systems, repositories and owners feed the knowledge base. External provider or repository failures can affect retrieval across many workflows.
The Knowledge Base Continuity Plan
Critical workflows should know how to access authoritative sources when the SI layer is unavailable. Humans should not lose the ability to find policy because the chatbot is down.
The Knowledge Base Build Team
- Domain owners
- Knowledge managers or documentation owners
- Frontline users
- Security and identity owners
- Technical retrieval team
- Governance or risk roles where needed
The knowledge base is cross-functional because organisational truth does not belong to the AI team alone.
The Knowledge Base 30-Day Build
Week 1 — Inventory
Choose one domain, inventory sources, owners, duplicates and recurring questions.
Week 2 — Govern
Select canonical sources, add metadata, resolve critical conflicts and define access.
Week 3 — Retrieve
Build search and source-grounded answers for a limited user group. Use known-answer and known-gap tests.
Week 4 — Learn
Review gaps, corrections and user search friction. Update sources and decide which adjacent domain to add next.
The Knowledge Base 90-Day Build
Month one proves one domain. Month two connects a second domain and shared metadata. Month three adds stronger evaluation, gap routing and one production workflow that consumes the knowledge.
The goal is to prove governance and learning before broad ingestion.
The Knowledge Base Migration Rule
Do not require teams to rewrite every document before adoption. Start by governing the highest-value sources and gradually improve structure as retrieval evidence reveals the need.
The Knowledge Base Simplification Rule
If two documents serve the same purpose, consolidate or clarify precedence. Better retrieval is not a substitute for simpler knowledge architecture.
The Knowledge Base Human-Agency Rule
Users should be able to inspect sources, report wrong knowledge and challenge an answer. SI should make organisational knowledge more legible, not more opaque.
The Knowledge Base No-Answer Rule
The system must be allowed to say “the organisation has not documented this.” That sentence is often more valuable than a plausible answer because it reveals a real knowledge gap.
The Knowledge Base No-Blend Rule
Do not blend conflicting policies into a synthetic compromise. Surface the conflict and route it to the owner.
The Knowledge Base No-Authority-by-Similarity Rule
A semantically similar document is not automatically applicable. Authority, scope, version and permission must participate in retrieval.
The Knowledge Base No-Duplicate-Truth Rule
Avoid maintaining different copies of the same policy for each assistant. Build once, govern once and retrieve across workflows.
The Knowledge Base No-Shadow-Policy Rule
Prompts, examples and model memory should not quietly become alternate policy. Durable organisational truth belongs in governed sources.
The Knowledge Base Promotion Gate
Expand to more users or autonomous workflows only when canonical source selection, permissions, citation and gap handling are stable.
The Knowledge Base Demotion Gate
Reduce autonomy when source quality falls, a conflict emerges, permission controls fail or dependent knowledge changes materially.
The Knowledge Base Retirement Gate
Retire unused indexes, connectors and duplicate repositories when they no longer support live workflows. Knowledge infrastructure needs lifecycle control too.
A Copyable Knowledge-Source Card
- Source ID
- Title
- Knowledge class
- Owner
- Scope
- Version
- Effective date
- Review date
- Confidentiality
- Canonical status
- Related sources
- Supersedes
- Keywords
- Known gaps
- Dependent workflows
The source card makes governance and retrieval visible without requiring every team to understand the indexing implementation.
Frequently Asked Questions
Do we need a vector database to build an SI knowledge base?
Not necessarily. The architectural need is governed retrieval. The technical implementation can use existing enterprise search, databases, document systems, semantic indexes or a combination.
Should we ingest every company file?
No. Begin with governed high-value knowledge. Ingesting everything can worsen retrieval and increase security risk.
How do we handle outdated documents?
Mark or retire them from current retrieval while preserving historical access where needed. Version and effective date should remain visible.
Who should own the knowledge base?
Technical teams can own the platform, but domain owners should own meaning and source authority. Knowledge governance is distributed.
How do we handle unanswered questions?
Return a gap state and route repeated important gaps to knowledge owners. Do not force the model to invent local policy.
Can model memory replace the knowledge base?
No. Memory can support continuity and preference, but workplace authority should live in maintained organisational sources.
How do we know the knowledge base is good?
Users find authoritative current sources faster, wrong-source retrieval is low, gaps are visible, permissions hold and corrections improve the source layer.
What comes next?
Continue to How to Connect Super Intelligence to Workplace Documents and Files, which focuses on the technical and operational bridge from governed knowledge architecture to real repositories.
The Core Knowledge-Base Rule
Build the knowledge base so the workplace can tell the difference between what it knows, what it used to know, what it does not know and what a particular user is allowed to know.
That distinction is the foundation for reliable Super Intelligence at work.
Worked Knowledge Domain: Internal IT Help
An internal IT knowledge base can begin with common device, software, access and troubleshooting procedures. The canonical source set should identify supported platforms, current versions, approved tools and escalation rules. SI can interpret the employee’s natural-language problem and retrieve the right runbook while account and device state come from the relevant systems.
The knowledge loop improves when repeated unsupported questions become documentation tasks. A recurring issue that only one technician can answer is evidence of hidden knowledge, not a reason to give the model permission to improvise.
Worked Knowledge Domain: Product and Customer Support
Product knowledge should separate current specifications, troubleshooting, warranty or eligibility rules, approved claims and examples. Customer-specific state remains in CRM or service systems. SI combines the product source with the authorised case context at the moment of work.
This architecture reduces the temptation to copy customer records into a static knowledge repository and reduces the risk that a general product document is mistaken for an individual customer entitlement.
Worked Knowledge Domain: Employee Onboarding
Onboarding knowledge can include role expectations, current policies, systems, glossaries, standard processes and learning material. The assistant can answer routine questions with source references and route undocumented questions to the appropriate owner.
The system should teach new employees where authority lives. It should not become a substitute for manager judgment, sensitive HR decisions or relationship-building.
Worked Knowledge Domain: Engineering Operations
Engineering knowledge can combine architecture decisions, runbooks, repository documentation, incident learnings and operational standards. Retrieval should be version-aware and permission-aware. Current code and service state come from engineering systems rather than static copies.
A coding or operations agent becomes safer when it retrieves the current runbook and environment state before tool use, rather than relying on remembered patterns from previous conversations.
Worked Knowledge Domain: Finance Operations
Finance knowledge should distinguish policy, procedure, definitions and live financial state. The knowledge base can explain reporting rules, close processes and approval thresholds while authoritative amounts remain in financial systems and deterministic calculations.
This separation lets SI help with interpretation and documentation without becoming a parallel ledger.
The Knowledge Base Transfer Standard
A strong workplace knowledge base should survive changes in model provider, chat interface or agent framework. Source ownership, canonicality, permissions, metadata and gap handling belong to the organisation. The model is one consumer of that infrastructure.
This portability is strategically important. It prevents organisational knowledge from becoming trapped inside a particular product’s memory or prompt format and allows the same governed sources to support human search, assistants, agents and future systems.
The Knowledge Base Maintenance Contract
Every high-value knowledge domain should have a maintenance contract: owner, source lifecycle, review trigger, permission owner, gap route and change-impact process. The contract can be lightweight, but it should make clear who is responsible when knowledge becomes wrong.
A knowledge base without maintenance may perform well on launch day and degrade quietly afterwards. Maintenance is not separate from the system; it is part of the system.
The Final Knowledge-Base Standard
A workplace knowledge base is ready for Super Intelligence when current authority is easier to retrieve than stale or informal knowledge, gaps remain visible, permissions survive retrieval and every important source has an owner.
When those conditions hold, SI can become a powerful interface to organisational memory without becoming the owner of organisational truth.
The Knowledge Base Production Gate
Before the knowledge base becomes a dependency for live workplace decisions, verify four things: current authoritative sources are easier to retrieve than drafts or history; access controls are enforced before context reaches the model; known gaps produce abstention rather than invented local policy; and important answers remain traceable to source and version.
This gate distinguishes a useful internal search experiment from a production knowledge layer. Production does not require perfect knowledge, but it does require known boundaries around what the system can and cannot establish.
The Knowledge Base Expansion Gate
Expand to another department only after the first domain has demonstrated source ownership, retrieval quality, permission safety and a working gap loop. New domains introduce new terminology, confidentiality and authority. Reuse the architecture, then reassess local meaning.
A company-wide corpus created before those controls exist can make search look powerful while making organisational truth less legible.
The Knowledge Base Agent Gate
When agents will act on retrieved knowledge, raise the evidence standard. Test wrong-scope retrieval, stale sources, permission boundaries, conflicts and should-stop cases. An answer that is merely useful to a human can become consequential when it automatically triggers a tool.
Agentic use should therefore depend on a governed knowledge base plus explicit workflow context. The knowledge base tells the agent what the organisation knows; the workflow tells it what it may do with that knowledge.
The Knowledge Base Review Meeting
Periodically bring domain owner, knowledge owner, frontline user and technical owner together to review the top gaps, stale-source incidents, retrieval failures and permission issues. The meeting should change sources or retrieval rules, not merely review usage statistics.
This keeps the knowledge system attached to operating reality. Frontline questions reveal where documentation no longer matches work; owners decide what becomes canonical; technical teams implement the change.
The Knowledge Base Final Checklist
- High-value knowledge domains have owners.
- Canonical current sources are identifiable.
- Draft, example, historical and live-state classes are distinct.
- Scope and permissions are enforceable.
- Metadata supports retrieval rather than decoration.
- Known-answer, known-gap and wrong-scope tests exist.
- Users can inspect sources and report errors.
- Repeated gaps reach the right owner.
- Superseded content is excluded from normal use.
- Dependent workflows can be identified after material source changes.
The Final Knowledge Principle
The goal is not to make every company document searchable. The goal is to make organisational authority, memory and uncertainty reliably retrievable.
When the knowledge base can distinguish source from example, current from historical, authorised from restricted and documented from unknown, Super Intelligence becomes a far more dependable workplace interface.
The Knowledge Governance Charter
A knowledge base becomes dependable when the organisation writes down who owns authority, who may publish, who can retire content and how conflicts are resolved. The charter should be short and operational rather than a generic knowledge-management policy.
- Who can declare a source authoritative?
- Who can approve material changes?
- Who can retire or archive old versions?
- Who manages metadata and taxonomy?
- Who reviews permissions?
- Who receives unresolved conflicts?
- Who owns search and retrieval quality?
- Who is responsible for incidents involving wrong knowledge?
The Source Hierarchy
Where several sources cover the same topic, define which outranks which. A current policy may outrank a training deck. A signed contract may outrank CRM notes for legal commitment. A live customer record may outrank a historical summary.
The hierarchy should be encoded into retrieval rules where possible so users do not need to remember every precedence manually.
The Conflict Resolution Workflow
- Detect conflicting sources.
- Identify owners and authority levels.
- Pause authoritative synthesis when necessary.
- Resolve the conflict at the source.
- Update metadata and supersession links.
- Retest affected queries.
- Notify users or workflows if the conflict materially affected prior output.
This turns contradiction into a repair process rather than a model-selection problem.
The Knowledge Incident
A knowledge incident occurs when wrong, stale, restricted or misleading information materially affects work. Examples include a superseded policy appearing current, a restricted source leaking through an answer or an incorrect procedure driving an external action.
Treat material incidents with the same seriousness as other operational failures: contain, correct the source, inspect affected workflows, update controls and capture the lesson.
The Knowledge Incident Log
Record incident date, affected source, affected workflows, detection method, impact, repair and prevention. This history helps identify recurring failure patterns such as weak retirement or poor metadata.
The Knowledge Risk Classes
- Low risk: internal reference material with easy correction.
- Moderate risk: operational guidance that affects workflow quality.
- High risk: knowledge affecting money, rights, security, safety or regulated activity.
- Critical: knowledge whose wrong application can create serious or systemic harm.
Risk class should influence review cadence, permission, retrieval strictness and how much human verification is required.
The Knowledge Review Calendar
Not every page needs the same review schedule. Create domain-specific review cadences and event triggers. High-change operational material may be reviewed monthly or after each release; stable reference definitions may need only periodic review.
The Knowledge Expiry Pattern
Temporary guidance should have an expiry date where appropriate. Incident procedures, campaign rules and transition guidance can become dangerous when they remain searchable after the context ends.
The Knowledge Sunset Pattern
When a process or product is retired, schedule removal or archival of related knowledge. Sunset links should point users to the current alternative rather than leaving dead-end documentation.
The Knowledge Merge Pattern
When duplicate documents are discovered, merge them into one canonical source or clearly separate their purpose. Merging should preserve useful history, ownership and references.
The Knowledge Split Pattern
A large document may contain several unrelated workflows or authority levels. Split it into clearer knowledge objects while preserving the canonical relationship. This often improves retrieval precision and review ownership.
The Knowledge Reference Graph
Useful knowledge rarely exists in isolation. Policies link to procedures, procedures link to forms, forms link to systems, decisions link to evidence and runbooks link to escalation contacts.
A reference graph allows users and SI to move from answer to surrounding context without flattening everything into one document.
The Related-Knowledge Pattern
Show related procedures, templates and exceptions around a retrieved source. This helps the user understand the operating environment while keeping the authoritative answer central.
The Knowledge Dependency Map
Some documents depend on other sources. A regional procedure may depend on global policy; a sales playbook may depend on product specifications. Track these dependencies so upstream changes trigger downstream review.
The Knowledge Change Blast Radius
Before publishing a major update, ask which derived documents, workflows, training materials and agents depend on the source. The broader the blast radius, the stronger the need for staged change and regression testing.
The Knowledge Regression Suite
Maintain representative questions for critical domains. After material source or retrieval changes, rerun them and compare the answers, citations and abstention behaviour.
Regression testing turns knowledge quality into an observable operating control rather than an assumption.
The Zero-Result Workflow
When search returns nothing, the system should offer a useful next state: clarify the query, suggest related terms, route to an owner or create a knowledge-gap request.
Zero results are not merely a search problem; they are a signal about corpus coverage and user language.
The Wrong-Result Workflow
Allow users to report that a result is wrong or irrelevant. Capture the query, returned source, expected source if known and reason. These examples are valuable for ranking and metadata improvement.
The Untrusted-Source Workflow
External material, old attachments and user-uploaded files may be useful but should not automatically gain the same authority as maintained internal sources. Label and isolate them until validated.
The Knowledge Intake Workflow
- Receive new source.
- Identify owner.
- Classify source type.
- Assign authority level.
- Set audience and confidentiality.
- Add version and currentness metadata.
- Review for duplication or conflict.
- Publish and index.
- Add regression questions if material.
The Knowledge Authoring Standard
Documents become more SI-readable when they use clear headings, explicit definitions, stable terminology, visible exceptions, unambiguous step order and concrete ownership. Good documentation helps humans first and retrieval systems second.
Definition-First Knowledge
Critical terms should have explicit definitions rather than relying on local intuition. Stable definitions reduce ambiguity across departments and improve semantic search.
Procedure-First Knowledge
Operational procedures should state preconditions, steps, decisions, exceptions, required evidence, owner and completion state. This makes the procedure easier to retrieve and safer to use in automation.
Decision-First Knowledge
Decision records should capture the question, evidence, owner, rationale, outcome and reassessment trigger. This prevents future users from treating the outcome as context-free policy.
Exception-First Knowledge
Where exceptions recur, document them deliberately. Explain when the standard rule does not apply, who can approve the exception and whether the exception is temporary or reusable.
The Knowledge Base and Templates
Templates are high-value knowledge objects because they turn abstract standards into repeatable output. However, a template should link back to the policy or procedure that justifies its fields.
The Knowledge Base and Forms
Forms can encode required information and reduce missing context. SI can help prefill or explain forms, but required fields and validation should remain deterministic where possible.
The Knowledge Base and Runbooks
Runbooks need current system state, clear triggers, step order, stop conditions and escalation. For high-impact operations, the runbook should remain directly inspectable even when SI helps execute or explain it.
The Knowledge Base and FAQs
FAQs are useful for recurring simple questions, but they can become stale duplicates. Link them to canonical sources and review them when the source changes.
The Knowledge Base and Playbooks
Playbooks often contain strategy, examples and recommended practice rather than hard policy. Mark the authority level so SI does not present a sales or marketing playbook as mandatory rule.
The Knowledge Base and Historical Knowledge
Historical decisions and retired procedures can remain valuable for audit and learning. Keep them accessible through explicit historical filters rather than ordinary current retrieval.
The Knowledge Base and Tacit Knowledge
Interview experts about recurring judgment cues, exceptions and workarounds. Some tacit knowledge can become explicit documentation; some remains human judgment and should be labelled as such.
The goal is not to force all expertise into text. It is to distinguish what can be shared from what still requires experience.
The Knowledge Capture Interview
- What questions do people repeatedly ask you?
- What do you know that is not written down?
- What exceptions do newcomers miss?
- What sources do you trust most?
- What signals tell you a case is unusual?
- What decision can only your role make?
- What information becomes stale fastest?
- What mistake recurs because documentation is weak?
The Expert-Review Workflow
Experts should review high-impact derived knowledge before publication. SI can draft the article or FAQ from interviews and sources, but the expert confirms accuracy, scope and authority.
The Knowledge Base and New-Hire Questions
New hires are excellent sensors for missing context because they cannot rely on institutional memory. Capture their repeated questions and use them to improve structure, terminology and navigation.
The Knowledge Base and Exit Knowledge
When employees leave, capture project state, key decisions, important relationships and unresolved risks where appropriate. Do not treat every personal note as organisational property; focus on knowledge needed for continuity.
The Knowledge Base and Projects
Projects generate temporary but valuable knowledge: decisions, assumptions, lessons and artefacts. At project close, decide what should become durable reference, what should remain archived and what should be deleted.
The Knowledge Base and Incidents
Incidents generate lessons, runbook updates and changed controls. Postmortems should feed maintained knowledge rather than remain isolated reports.
The Knowledge Base and Product Releases
A release may change specifications, support guidance, sales messaging and training. Link release state to dependent knowledge so updates happen consistently.
The Knowledge Base and Regulation
Regulatory or policy changes can have wide knowledge impact. Record effective date, jurisdiction and source authority so SI does not apply old guidance after the transition.
The Knowledge Base and Third-Party Knowledge
Vendor manuals, public standards and external research can be useful reference sources. Label origin and authority, and distinguish them from internal policy or organisation-specific interpretation.
The Knowledge Base and Generated Content
Machine-generated summaries should be treated as derived content unless they have been reviewed and approved as maintained knowledge. Generation alone does not create institutional authority.
The Knowledge Base and User Notes
Personal notes can improve local productivity but should not automatically enter the enterprise corpus. Promote only the notes that have durable value and suitable review.
The Knowledge Base and Collaborative Editing
Collaborative editing can accelerate maintenance but should preserve review status, ownership and version history. A draft being edited by many people is not automatically ready for retrieval.
The Knowledge Base and Search Analytics
Use search analytics to discover frequent topics, zero-result queries, repeated reformulations and sources that users abandon quickly. These signals reveal both knowledge gaps and interface problems.
The Knowledge Base and Answer Analytics
Track unsupported answers, low-confidence responses, citation failures and human corrections. The aim is to improve the knowledge system, not to score employees.
The Knowledge Base and Cost
Knowledge maintenance requires authoring, review, indexing, storage and governance. Compare these costs with reduced search time, fewer expert interruptions, faster onboarding and better workflow quality.
The Knowledge Base and Trust
Trust grows when users can see current sources, ownership and evidence. It falls when the system produces polished answers that cannot be traced or when known stale content keeps appearing.
The Knowledge Base and Human Override
Users should be able to challenge a result and access the source directly. An SI interface should not make the underlying organisational knowledge inaccessible.
The Knowledge Base and Explainability
Explain why a source was selected when useful: current policy, applies to this region, owned by Finance, effective this quarter. This gives users operational context rather than only semantic similarity.
The Knowledge Base and Personalisation
Personalisation can improve relevance by considering role, project or location, but it should not hide important general policy or create inconsistent truth. Personalise the view, not the underlying authority.
The Knowledge Base and Memory
Persistent SI memory should not override current knowledge-base state. Memory can help continuity, but canonical sources remain the place where organisational truth is maintained.
The Knowledge Base and Context Engineering
A knowledge base provides the pool of maintained evidence; context engineering selects the subset needed for one task. Good context engineering depends on strong metadata, source authority and currentness.
The Knowledge Base Build Sequence
- Choose one high-value domain.
- Inventory recurring questions.
- Identify canonical sources.
- Assign owners and authority.
- Clean obvious duplicates and stale content.
- Add metadata and permissions.
- Connect search or retrieval.
- Test known answers, gaps and conflicts.
- Launch to a small user group.
- Capture corrections and zero results.
- Repair the source and retrieval layer.
- Expand only after quality is stable.
The Knowledge Base Go-Live Checklist
- Canonical sources identified
- Owners named
- Draft and archived states visible
- Permissions enforced
- Metadata present
- Known-answer tests pass
- Known-gap questions abstain
- Conflicts handled
- Citations inspectable
- Gap queue has an owner
- Review cadence defined
- Fallback source access remains available
The Knowledge Base Scale Checklist
- Taxonomy still works across teams
- Permissions scale across roles
- Source ownership remains local and clear
- Search analytics are monitored
- Shared infrastructure has an owner
- Regional or product variation is modelled
- High-risk domains have stronger controls
- Incident process exists
- Change blast radius can be assessed
Frequently Asked Questions: Knowledge Governance
Who should own the knowledge base?
The platform may have a central technical owner, but domain knowledge should remain owned by the business or professional functions responsible for its accuracy.
How much metadata is enough?
Enough to support authority, currentness, permissions and useful retrieval. Start with a small required set and add fields only when they improve real workflows.
Should every answer be generated?
No. Sometimes the best response is the source itself, a search result list or a structured excerpt. Generation is optional.
How do we avoid stale AI answers?
Maintain source ownership, versioning, retirement, currentness rules and regression tests. Stale knowledge is primarily a source-lifecycle problem.
Can users add knowledge directly?
They can submit contributions, but publication into authoritative knowledge should follow the appropriate review and ownership process.
What if employees keep using old documents?
Improve discoverability of the canonical source, remove or mark obsolete copies and explain the source hierarchy. Search behaviour often reveals where the old path remains easier.
The Final Knowledge Standard
A workplace knowledge base is ready for Super Intelligence when the organisation can explain not only what the answer is, but where it came from, why it is current, who owns it, who may use it and what happens when the answer is missing.
That is the difference between a document collection and organisational memory that can safely support intelligent work.
The Knowledge Base Independence Test
A strong knowledge base should remain understandable if the current AI product disappears. Owners can still identify the canonical source, employees can still open it, and the organisation can still explain what is policy, procedure, reference, example and live state. The SI layer improves access to knowledge; it should not become the only place where knowledge exists.
This test also protects against model lock-in. A new approved model or search interface should be able to consume the same governed sources without recreating organisational truth from scratch.
The Knowledge Base Receiver Test
Ask employees whether the system reduces the time required to find authoritative information and whether the cited source is useful when they need to verify the answer. If users still ask colleagues to confirm every result, retrieval may be technically impressive but organisationally weak.
Receiver trust should grow from source visibility, currentness and correction mechanisms rather than from the assistant’s tone.
The Knowledge Base Learning Flywheel
Every workplace question can improve the system. Supported questions validate coverage. Unsupported questions reveal gaps. Wrong answers reveal source or retrieval problems. Repeated exceptions reveal ambiguous rules. Source changes trigger regression tests. This feedback turns daily work into knowledge maintenance.
The flywheel is what allows the knowledge base to become more useful over time without depending on one expert to remember every correction.
The Knowledge Base Final Standard
A workplace knowledge base is production-ready for Super Intelligence when source authority, ownership, scope, permission, currentness, gap handling and verification remain visible from question to answer.
That is the floor for dependable organisational retrieval. The next layer is connecting those governed sources to the actual workplace documents and files where knowledge lives.
