Map the expertise.
Then capture what it knows.
Peer-validated competency network + AI expert interviews + three-phase extraction pipeline + semantic comparison across experts. Knowledge Mapping is the institutional memory layer that captures who knows what, how they actually do it, and reconciles it across the org, with every decision reversible.
How do we handle escalations from enterprise customers when the on-call engineer is offline?
AI
From the documented runbook merged from Marco K. and Sofia L.'s expert interviews (similarity 0.97, manager-approved 2025-02-18): page the secondary on-call via PagerDuty; if no ack in 5 minutes, escalate to the EM through #incident-eng. For enterprise-tier customers, Sofia's addition opens a dedicated #cust-XYZ-incident channel for comms.
Who else knows this procedure if both Marco and Sofia are unavailable?
AI
From the competency graph: 3 additional validated experts on “incident response · enterprise”: Luca R. (confirmed by 4 peers, Anche Io path), Anna P. (manager-nominated, expert interview scheduled), Federico C. (confirmed, no documented process yet). Suggest paging Luca first; he has the same documented runbook as Marco.
Who knows what, validated before documentation begins.
The system starts with the competency map, not the document. Self-claims and manager nominations flow through a four-verdict peer validation network. Only validated competencies trigger an expert interview.
- Five-verdict validation: Confirmed · Disagree · Not familiar · “Anche Io” (I have it too) · “Shared Others” (nominate a colleague). The shape of the network maps shared expertise without requiring a survey.
- Threshold-based promotion: candidate → validated experience only after configurable confirmations. No expert interview wasted on a claim peers reject.
- Cross-pollination via Anche Io: when a peer flags they also have the competency, the system spawns a parallel expert interview, capturing variants of the same process from multiple sources.
- Scope-aware distribution: validation requests routed by organization element, project, or global, peers are chosen for expertise proximity, not org-chart proximity.
Validation Session · KMC-0042
MK
Marco K. · Engineering · APAC
Self-claim: incident response runbook for enterprise customers
Your verdict
Anche Io = you also have this competency. Choosing it triggers a parallel expert interview for you, and the system will compare your process with Marco's once both are documented.
Tacit expertise extracted by a conversational AI, no form to fill.
Once a competency is validated, the system invites the expert to an adaptive chat. The interview agent decides what to probe and when the process has been captured with all critical details, no fixed length, no fixed script.
- Adaptive questioning: the agent reads each answer and picks the next question, probing specifics (names, channels, timing) where the answer is vague.
- Completion detection: the agent decides when the process has the steps, stakeholders, tools, and timing needed for downstream extraction, and not before.
- Real-time streaming: SSE-based session, full conversation history persisted, recoverable across sessions.
- Context-aware: the agent receives expert role, organization, project, and the competency context, not a blank slate.
- Multilingual capture: expert interviews in the contributor's preferred language; downstream retrieval works across languages.
Three phases. Conversation in, structured knowledge out.
Every completed interview runs through three specialized LLM agents in sequence, orchestrated as Celery tasks so the UI never blocks. Output is a structured Process with steps, stakeholders, tools, timing, and a completeness score.
- Phase 1 · Reordering: the Reordering Agent rewrites the chat into a chronological narrative, stripping chitchat and meta-talk. Output: a clean linear text the next phase can parse deterministically.
- Phase 2 · Extraction: the Extraction Agent emits structured data: process name, description, ordered steps, stakeholders (name, role, contact channel, timing), tools, overall timeline.
- Phase 3 · Validation: the Validation Agent computes a 0–1 completeness score and flags only critical gaps (external stakeholders, critical tools, timing for high-impact steps). Below 0.75, the interview is flagged for follow-up.
- Smart noise filtering: tools like “discuss” or “review” are not flagged. Internal team contacts are not flagged. Only material gaps surface.
- Audit trail: every LLM call is logged (model, prompt, response, tokens) for compliance review.
When two experts document the same process, the system reconciles them.
Once two or more experts have documented processes in the same context, a semantic comparison agent runs automatically. It identifies shared steps, unique contributions, and conflicts, then routes the result through a three-band decision logic.
- Similarity > 0.95: creates a Manager Approval request. A merge is never automatic; the comparison goes to the campaign creator with shared / unique steps and the agent's reasoning. The manager approves into a shared process, or keeps separate.
- Similarity 0.4–0.95: marked as partial match. Both processes coexist; the context is tagged for incremental review.
- Similarity < 0.4: flagged for human review. The two experts likely describe different processes that happened to share a competency label.
- Reasoning with real names: the comparison report uses the actual expert names (“Marco's escalation step”, not “Expert 1”), making decisions readable across the management chain.
- Conflict surfacing: where steps contradict (Marco says “escalate via Slack”, Sofia says “via email”), the agent flags the conflict instead of silently picking one.
Process Comparison · CTX-074
pending approval
Similarity
0.97
MKMarco K.
vs
SLSofia L.
- SHARED Page secondary on-call via PagerDuty if primary doesn't ack in 5 min
- MARCO Escalate to EM directly via #incident-eng Slack, not email
- SHARED Acknowledge in StatusPage within 10 min of detection
- SOFIA Open #cust-XYZ-incident channel for enterprise-tier customers
Recommendation: Merge into shared process: Sofia's customer-comms step extends the runbook for enterprise-tier customers, with no conflict against Marco's escalation path.
A searchable knowledge base, with an undo button on every decision.
The output is a knowledge base of documented processes, searchable, filterable, drill-down to chronological narrative + stakeholders + tools + timing. Every operation is soft-delete: the audit trail survives every merge, every undo, every refresh.
- Server-side DataTables: fast search and filter even at thousands of documented processes. Drill from row to the full process detail.
- Completeness score visible per process: 0–1 score from Phase 3 validation, with the missing-info list. Knowledge gaps are explicit, not hidden.
- Reversible merges: a manager-approved merge can be undone in one click. Original processes restored, shared process archived, audit trail preserved on both sides.
- Soft delete on everything: archived data stays queryable for audit (GDPR Article 30, ISO 27001 evidence). Restore is a state change, not a backup recovery.
- Three-tier access model: Owner / App Manager / Campaign Creator see the full pipeline; regular users see only their own overview and pending actions (validations, expert interviews, approvals).
The captured knowledge feeds every agent on the platform.
Knowledge Mapping is the source, not the destination. The validated, documented, deduplicated knowledge base becomes the retrieval substrate for the rest of the platform: chatbots, onboarding agents, strategic analysts.
- Chatbot Assistants cite back to the original expert interview that captured the answer.
- Onboarding agents serve new-hire questions from validated processes instead of generic FAQs.
- Internal handovers hand a structured brief to the incoming owner, not a Loom link.
- Need-to-know scope on every asset. Each captured asset carries a confidentiality scope (organization unit, project, or private to the author), so retrieval can be limited to who is cleared to see it.
- Right-to-delete honored at expert level: an erasure request soft-deletes the contributions; the merged process remains with the expert's identifiable contributions stripped from the trail, and the erasure event is logged.
Where knowledge mapping pays off.
Four operating contexts where the cost of not capturing structured knowledge is measurable, and the system pays back in the same cycle.
- Critical Knowledge Preservation
- Retirement & Succession Risk
- Founder-Era Knowledge Capture
- Single-Point-of-Failure Audit
- Operational Resilience
- Process Standardization (Multi-Site)
- Best-Practice Identification
- Cross-Team Variant Reconciliation
- M&A Knowledge Integration
- Customer-Escalation Playbook
- New-Hire Onboarding KB
- Incident-Response Runbook
- Structured Mentoring Programs
- Failed-Project Lessons
- Quarterly Refresh Cycle
Start Now
For any request, please contact us
Knowledge Mapping FAQ
What HR Ops, Knowledge Managers, and Heads of Operations ask first.
What happens when multiple people claim the same competency?
Each person goes through their own validation: peers respond independently for each candidate. When peers respond with “Anche Io”, the system creates parallel candidates for them. After expert interviews, the semantic comparison agent detects similarity and routes the result to manager approval: merge into a shared process, keep separate as variants, or flag for review.
Does an expert have to fill in forms?
No. The expert has a chat with the AI interviewer. The agent adapts questions to what's been said, decides when the process is complete, and the three-phase pipeline extracts the structured data afterward. Average session is 30–60 minutes; experts can pause and resume.
What if a merge was wrong?
Undo is one click. The shared process is archived, the original processes are restored, experiences re-point to the originals, and the metadata records the undo with timestamp and manager identity. Reversibility is a system property, not an opt-in feature.
Who sees what?
Three-tier model. Owner / App Manager / Campaign Creator see the full pipeline (candidates, sessions, interviews, comparisons, knowledge base). Regular users see only their own overview and pending actions: the validations, expert interviews, and approvals on their plate.
GDPR right-to-delete on captured knowledge?
Honored. Soft delete at expert level removes the contributions from active retrieval; merged processes remain but the expert's identifiable contributions are stripped from the trail. Audit log retains the event of erasure for compliance review without retaining the erased content itself.
How does it handle multilingual orgs?
Expert interviews run in the contributor's preferred language; the extraction pipeline preserves the linguistic context; downstream retrieval works across languages. A peer validating from a different language sees the claim translated, not the raw text.
Does the knowledge stay current?
Configurable refresh campaigns. The system can re-validate existing competencies on a quarterly cadence; re-validation interviews supersede outdated facts and the audit trail preserves the prior state. Stale-flagging marks chunks not referenced or revalidated within the threshold.
What's the architecture?
Specialized LLM agents orchestrated via the Agno framework, each run as a Celery task: a conversational interview agent, a three-phase extraction pipeline (reordering, extraction, validation) with a loss-audit pass, and a semantic process-comparison agent, alongside synthesis and assessment agents. Competency normalization, validation, and context-matching run as LLM-backed steps in the same framework. Vector-ready data layer for future semantic-search integration.
Want to scope a Knowledge Mapping campaign with us?
Contact Sales