The most expensive answer is a confident wrong one

Teams adopt AI assistants, then discover a new failure mode: plausible hallucination about their own work. "We decided to defer payments"—never logged. "Release is May 3"—board says May 17. "Legal approved"—approval task still open.

AI project knowledge base Q&A fixes the retrieval layer. WKFGo combines project wiki, search_docs via MCP, smart_search, and the chatbot connected to PostgreSQL so answers trace to pages, tasks, and decisions your organization actually recorded.

Problem: Q&A without sources

Ungrounded assistants optimize for fluency, not audit trails. Symptoms:

New hires act on chat answers without checking wiki
Executives quote AI in customer calls—contradicting finance
Repeated questions because nobody publish_meeting_wiki after decisions

Grounding is a product and habit problem, not a bigger model alone.

Framework: publish → index → query → cite → fix gaps

Publish durable knowledge — create_wiki_page, publish_meeting_wiki, decision pages
Index — Vector search over wiki; document module for PDFs
Query — Chatbot or MCP: search_docs, get_wiki_page, smart_search
Cite — Answers should link task IDs and wiki titles humans can open
Fix gaps — When Q&A fails, update_wiki_page—do not train people to trust silence

Good questions for KB Q&A

"What is our rollback procedure for service X?"
"Who approved the API rate limit change?"
"Summarize open risks for Project Atlas this week."
"Find wiki pages about SSO last updated this quarter."

Bad expectations

Legal interpretation without counsel review
Real-time secrets or credentials in chat logs
Automatic external customer replies without human approval

Solution: Q&A stack in WKFGo

Layer 1 — Wiki and meeting publish

After steering meetings, publish_meeting_wiki creates searchable decision text. Architecture and runbooks live in create_wiki_page with owners and last-reviewed dates.

Without publish discipline, Q&A returns "not found"—which is honest.

Layer 2 — MCP search_docs

IDE and executive assistants call search_docs with scoped credentials. Results respect feature keys. Use read-only API keys for broad teams; reserve write tools for facilitators.

Prompt pattern: "Search docs for 'data retention policy'; cite page titles; if missing, say data gap."

Layer 3 — Chatbot with live DB

The Node chatbot service queries tasks, finance summaries (within permission), and wiki vectors. Connect frontend widget via REACT_APP_CHATBOT_URL for in-app Q&A.

Treat chatbot output as draft intelligence—PM validates before policy actions.

Executive Q&A without leakage

Executives often ask sensitive questions in chat—headcount plans, deal timing, litigation-adjacent work. Route executive assistants through scoped MCP keys tied to their user, not shared admin tokens. Summaries for staff meetings should redact names and amounts when wiki pages are restricted; search_docs returns only what their role may read.

For board prep, chain Q&A: search decisions → verify with decision_log → attach citations to brief talking points. Fluency without citations is rejected in review.

Layer 5 — Weekly knowledge hygiene

When Q&A cites a decision, verify approval tasks and decision_log entries. Status questions should chain list_tasks, get_task, not prose alone.

Reducing hallucination risk

Require citation links in assistant prompts
Block demo read-only users from write MCP tools mid-answer
Log failed queries; quarterly wiki gap review
Bilingual teams: index content in the language you search

Layer 5 — Weekly knowledge hygiene

Assign a 15-minute wiki hygiene slot: merge duplicate pages found via search, archive superseded decisions, and update_wiki_page last-reviewed stamps. Q&A quality tracks hygiene—when assistants fail, fix the page, not the prompt alone.

Run ask_human when retrieval returns conflicting pages (two SSO decisions, different dates). Ambiguity should surface to facilitators, not merge silently in chat.

Anti-patterns

Connecting ChatGPT to email exports instead of live API
Superuser tokens in shared Slack bots
Using Q&A instead of updating the board after changes
Ignoring ask_human when retrieval confidence is low

Knowledge base Q&A fails when wiki pages contradict the board. Establish a rule: decisions in meetings become wiki within twenty-four hours; task status changes reference the decision page. get_wiki_page and search_docs ground chatbot answers—hallucination drops when citations are required.

Run a monthly audit: do AI answers cite task IDs from MCP? If not, tighten prompts to require citations before any stakeholder-facing draft leaves the team.

Permission-aware tools mean the same question gets different grounded answers for PM vs contractor—that is a feature, not a bug. Train teams to expect scoped truth.

FAQ

Chatbot vs MCP search_docs?
Chatbot is in-app conversational; MCP serves IDE and automation—both should ground in same data.

Does Q&A see finance data?
Only if user's access allows; same as UI.

Can we restrict sensitive projects?
Yes via feature access and project scoping in queries.

How to handle "I don't know"?
Preferred over fabrication—then file update_wiki_page task.

Ask questions that cite your project truth

Publish decisions to wiki, search with permissions, and demand citations—not vibes.

Try it now

Put these patterns on live project data—not slide decks.