Llm wiki faq

LLM-Wiki FAQ

Goal

Give users concrete, evidence-aware answers to common LLM-Wiki adoption questions without overclaiming. Prefer practical operating models over hype.

When to use

Use when the user asks:

• "Why do I need this?"
• "What benefits will I get?"
• "Is there evidence this works?"
• "Is this better than RAG?"
• "How do I keep it alive?"
• "Do I need Obsidian?"
• "Is this just another note-taking system?"
• "Will this become slop?"
• "What should I measure?"
• "I am not a developer; how would I use this?"
• "Why should I deal with PRs/MRs when Confluence only needs a browser?"
• "Will the wiki fit in the model context as it grows?"
• "Is this knowledge for agents or humans?"
• "What content can I upload, and what about sensitive data?"
• "Where should I start?"
• "When should I not use it?"
• "Who owns or reviews the wiki?"
• "How much time or money will this take?"
• "How do I prove ROI?"
• "Does this replace Confluence, Notion, SharePoint, RAG or search?"
• "How do I migrate safely?"
• "What if the agent is wrong?"
• "How do permissions, vendor lock-in, link rot or multi-agent access work?"
• "What are the strongest arguments against this?"
• "Will this hurt my own understanding?"
• "Will this become a write-only archive?"

Inputs

• The stakeholder's question or objection, in their own words (e.g. "Is this better than RAG?", "Do I need Obsidian?").
• Context about the target wiki: greenfield vs already migrating, team size, whether git/PR workflows already exist.
• Whether sensitive or regulated data is involved, so the answer can route to llm-wiki-privacy-redactor or llm-wiki-model-policy when needed.
• Which bundled reference files are available under references/ (evidence-pack, adoption-objections, additional-adoption-q-and-a, criticism-pack) to ground the answer.

Required references

Read these when available:

1. references/evidence-pack.md for evidence-backed adoption arguments.
2. references/adoption-objections.md for non-developer, Confluence/browser, token, human-readable, slop and sensitive-data objections.
3. references/additional-adoption-q-and-a.md for start-small, ownership, cost, ROI, migration, RAG/search, lock-in and multi-agent questions.
4. references/criticism-pack.md for serious criticism and mitigation answers.
5. Repository-level docs:
   • references/docs/adoption-objections.md
   • references/docs/adoption-q-and-a.md
   • references/docs/criticism-and-mitigations.md

Evidence policy

Use three evidence levels:

1. Direct evidence: LLM-Wiki-specific papers, benchmarks or implementations.
2. Adjacent evidence: GraphRAG, global RAG, long-term memory, context engineering and long-context studies that support the mechanism.
3. Local evidence: the user's own metrics after adoption: retrieval hit rate, answer reuse, review backlog, stale pages and output beyond the vault.

Always distinguish these levels. Do not claim that adjacent evidence proves every LLM-Wiki implementation will work.

For criticism, also distinguish:

solvable with schema/discipline
partially mitigated with residual risk
not solvable by tooling

Procedure

1. Identify the user's objection

Map the question to one of:

User question	Core answer
Why is this needed?	It reduces repeated context reconstruction and wiki bookkeeping cost.
What benefits?	Faster recall, reusable research, decision memory, better agent context and auditable knowledge.
Evidence?	Direct evidence is early; adjacent evidence strongly supports structured persistent memory and precomputed corpus structure.
Better than RAG?	Not universally; use LLM-Wiki for durable synthesis and RAG/search for retrieval.
Keep alive?	Use capture -> triage -> ingest -> query -> file-back -> lint -> review -> refresh.
Need Obsidian?	No; Obsidian is a good UI, not the core pattern.
Not a developer?	Use browser/Obsidian/form/chat capture; git/PRs can be hidden behind agent and maintainer workflows.
Confluence is easier?	Capture should be as easy as Confluence; PR/MR review is only for trusted/published/verified knowledge.
Tokens enough?	Do not paste the whole wiki; load index.md, search, and open relevant pages progressively.
Agents or humans?	Both; Markdown must stay human-readable, source-backed and reviewable, not an opaque agent cache.
Sensitive content?	Classify data first; route public/internal/confidential/restricted/secrets through different storage/model policies.
Where to start?	Run a small reversible pilot before migrating or building infrastructure.
When not to use?	Avoid it when the corpus is tiny, unreviewed, not reused or ownerless.
Cost/time?	Start with files/search; add model/retrieval/CI cost only after metrics justify it.
ROI?	Compare with-wiki vs without-wiki on real questions and output beyond the vault.
Ownership/review?	Agents draft and maintain bookkeeping; humans promote reviewed/verified knowledge.
Replace tools?	Usually no; Confluence/Notion can be capture surfaces, RAG/search can remain retrieval.
Migration?	Inventory first, dry-run plan second, staged reversible migration third.
Agent mistakes?	Assume errors; use raw sources, draft status, provenance, lint, review gates and dry-run.
Vendor lock-in?	Keep Markdown/JSON/YAML/git/raw files as source of truth; indexes are caches.
Multi-agent use?	Use explicit schemas, write boundaries, least privilege and review gates.
Serious criticism?	Run a criticism-first screen: smoothing, poisoning, cognitive debt, drift, cost, injection, benchmark vacuum and write-only archive risk.

2. Answer with the right level of confidence

Use language like:

• "The strongest direct evidence I found is..."
• "This is adjacent evidence, not a direct proof of personal LLM-Wiki."
• "The local decision should be based on a 2-4 week pilot and metrics."
• "For your case, the key benefit is..."
• "This objection is valid; the workflow should be designed to avoid that friction."
• "The safe default is a reversible pilot, not a full migration."
• "This mitigation reduces risk but does not remove it."

Avoid:

• "LLM-Wiki is proven better than RAG."
• "Obsidian is required."
• "The wiki will stay correct automatically."
• "More notes means more value."
• "Everyone must use git and PRs."
• "Token limits do not matter."
• "Generated summaries are documentation quality by default."
• "It is safe to upload everything."
• "This replaces your existing tools immediately."
• "No one needs to own review."
• "The tool solves cognitive ownership."

3. Use the core argument

The central argument:

LLM-Wiki is useful when raw sources are already plentiful but reusable context is scarce. It turns raw material into an inspectable, versioned, agent-readable knowledge layer and makes maintenance cheap enough to keep running.

Add the criticism-aware qualifier:

It is not useful when it becomes a write-only archive, mirrors greppable sources, replaces human synthesis where internalization matters, or lacks provenance/review owners.

4. Handle serious criticisms

Use references/criticism-pack.md.

Default pattern:

## Direct answer

Yes, that criticism is real.

## Where it applies

## Mitigation

## Honest residual risk

## What I would do for your case

Key answers:

• Consensus smoothing: use extraction/antagonist/friction/linking rules; residual risk is schema quality.
• Context poisoning: use raw immutability, claim provenance and lint; residual risk exists between lint passes.
• Cognitive debt: automate bookkeeping, not ## My synthesis; residual risk is not solvable by tooling.
• Drift: schedule lint/reconciliation separate from ingest; residual cost remains.
• Greppable corpora: avoid LLM-Wiki unless it compresses many sources.
• Token burn: measure cost/query against baseline and use model tiering/content hashes.
• Prompt injection: permissions are the boundary, not prompt rules.
• Benchmark vacuum: use local pilot metrics and do not overstate public evidence.
• Novelty decay: build output loop before capture expansion.

5. Handle non-developer and browser-first objections

For non-developers, recommend a low-friction surface first:

browser/form/chat/Obsidian/Confluence capture -> agent triage -> review queue -> wiki pages

Do not make git literacy a prerequisite. Explain that git gives history, rollback and review, while the daily UI can be browser-first.

For Confluence comparisons:

• agree that browser editing is easier;
• separate capture from governance;
• suggest append-only inbox or browser editing for ordinary contributions;
• reserve PR/MR gates for verified, public, regulated or CODEOWNERS-owned pages;
• suggest batch review PRs instead of one PR per note.

6. Handle token-growth concerns

Explain progressive context loading:

small instruction pointer -> concise wiki index -> search -> selected pages -> raw evidence only when needed

If the proposed workflow requires loading the entire wiki into every prompt, call it a context-dump anti-pattern and route to llm-wiki-retrieval-architect.

7. Handle human-readable / slop concerns

State that the wiki is for both humans and agents. Trust comes from:

• preserved raw sources;
• provenance links;
• status: draft|reviewed|verified|stale|archived;
• review_required and confidence fields;
• protected human synthesis sections;
• lint for unsupported, stale and contradictory claims;
• human promotion from draft to reviewed/verified.

Do not treat generated text as official documentation until it is source-anchored and reviewed.

8. Handle content and sensitive-data concerns

Classify before ingesting:

Data class	Default stance
Public	Normal wiki flow; cloud models allowed if policy permits.
Internal	Private repo/storage and model policy.
Confidential	Prefer local or approved private provider; redact before cloud.
Restricted / regulated	Do not ingest by default; explicit approval, minimization and retention rules.
Secrets	Never ingest; rotate if captured.

Route to llm-wiki-privacy-redactor, llm-wiki-model-policy, llm-wiki-threat-model or llm-wiki-security-review when the user mentions PII, secrets, regulated records, customer data, credentials or provider boundaries.

9. Handle broader adoption questions

For start-small, cost, ROI, ownership, review, migration, replacement, lock-in and multi-agent questions, use references/additional-adoption-q-and-a.md.

Key defaults:

• start with a reversible pilot;
• assign human owners for truth and review;
• keep capture cheap and review risk-based;
• keep existing tools until the wiki proves value;
• use search/RAG as complements, not enemies;
• keep Markdown/raw files as source of truth;
• measure outputs, not page count.

10. Cite or name evidence

For current facts, browse and cite. For static bundled evidence, summarize the source map from references/evidence-pack.md and criticism map from references/criticism-pack.md.

Source map (verify-before-use: re-check each source before citing its result as current, and label the evidence level):

• LangChain wiki-memory writeups: position wiki memory as a compact, persistent, agent-readable knowledge layer (adjacent evidence).
• Anthropic context-engineering guidance: context is finite; agents benefit from just-in-time file/context navigation (adjacent evidence).
• "Lost in the Middle": long context alone is unreliable (adjacent evidence).
• GraphRAG / GlobalRAG papers: flat chunk retrieval struggles on global/corpus-level questions (adjacent evidence).
• Mem0 / MemMachine reports: vendor-adjacent benchmarks where persistent structured memory beat full-context or naive RAG baselines; do not present as proof for LLM-Wiki.
• 2026 LLM-Wiki retrieval paper: reported early direct evidence on multi-hop and structured queries; re-verify its scope and status before citing.

11. Offer a pilot instead of a leap of faith

Recommend a bounded pilot:

20-50 sources
10-20 real questions
weekly lint
file back reusable answers
measure retrieval hit rate and answer reuse
measure cost/query versus grep/search baseline
run critique audit before scaling

12. Route to the next skill

• Needs solution choice: llm-wiki-choose.
• Needs setup: llm-wiki-setup.
• Has existing docs: llm-wiki-doctor then llm-wiki-migration-planner.
• Needs evidence measurement: llm-wiki-eval (including bounded pilot benchmarks).
• Worries about slop or serious objections: llm-wiki-critique-audit, llm-wiki-trust-audit, llm-wiki-provenance.
• Asks about Obsidian: llm-wiki-obsidian-hardening if they already use it; otherwise llm-wiki-local-first-stack.
• Asks about token growth or RAG/search boundaries: llm-wiki-retrieval-architect.
• Asks about team/browser workflow, ownership or review: llm-wiki-team-rollout and llm-wiki-github-action.
• Asks about sensitive data, permissions or model boundaries: llm-wiki-privacy-redactor, llm-wiki-model-policy, llm-wiki-threat-model, llm-wiki-security-review.
• Asks about archive/lock-in/publishing: llm-wiki-archive and llm-wiki-export-publish.
• Asks about instruction files or multiple agents: llm-wiki-agent-memory-bridge.

Output

Use this shape:

## Direct answer

## Why this matters for your case

## Evidence level / operating assumption

## Trade-offs and failure modes

## Practical workflow / next step

Safety gates

• Do not overstate evidence.
• Do not treat generated summaries as proof.
• Do not recommend cloud processing for sensitive material without explicit approval.
• Do not recommend Obsidian by default.
• Do not say the wiki stays alive automatically; explain the maintenance loop.
• Do not make git/PR/MR knowledge a prerequisite for non-developer users.
• Do not solve token concerns by suggesting larger context windows alone.
• Do not call generated pages "documentation" unless they are readable, source-backed and reviewable.
• Do not ingest secrets.
• Do not claim existing tools must be replaced before a pilot.
• Do not recommend a team rollout without explicit owners and review gates.
• Do not use page count as a success metric.
• Do not hide residual risks after mitigations.
• Do not automate human synthesis when the user is trying to understand or write original work.