Mindcraft Book audit
Mindcraft / Capabilities / Knowledge Agents
Module 07 · Capability

Knowledge that
answers itself.

Train an agent on your docs, transcripts, contracts, history. People ask in plain English; it answers with sources. Tribal knowledge un-trapped. Senior engineers stop being the human Stack Overflow.

What this fixes

Three places
institutional knowledge goes to die.

If you've ever had a senior engineer go on parental leave and watched the team struggle for three weeks, you're who we built this for.

Problem 01

Knowledge lives in heads.

Your senior engineer knows everything about the deployment pipeline. They're going on parental leave. The Notion is two years out of date. You see the problem.

Problem 02

Onboarding is a Slack archaeology dig.

New hire spends three weeks asking the same questions across four channels. Old hires repeat the same answers. Nobody likes any part of it. Productivity is delayed by weeks.

Problem 03

The right answer exists somewhere.

It's in a Loom from eight months ago, or a contract clause in box 47 of Drive, or the back of an email thread. Someone in your company knows it; nobody knows who. So nobody asks.

Sample query

What it
actually does.

A real question asked in Slack two minutes ago, lightly anonymised. The agent pulls from five sources across Notion, Drive, Slack, and a contract — synthesises the answer, cites every claim, and respects the asker's permission scope.

Answered · with sources 5 sources used Confidence: 94% Asked · today, 14:42 EST · answered in 1.8s
knowledge-agent Mira K.· Engineering · 2m ago Engineering scope
M
Mira Karboasked at 14:42
What's the refund process for orders over €1,000 in EU? And does it work the same for UK after Brexit?
Searched 12,470 docs Retrieved 5 sources Synthesised answer
1.8s
M
Mindcraft Knowledge· today, 14:42 · 94% confidence

For EU orders over €1,000, the refund process requires manager approval [1] and must comply with the GDPR-aligned 14-day statutory cooling-off period [2]. There are three steps:

  1. Customer submits a refund request via the customer portal [3], attaching the order ID and reason.
  2. Manager review within 48 hours. Approver must be the regional ops lead — usually Lukas (DACH) or Sofía (Iberia) [1].
  3. Refund processed via Stripe in the original currency [4]. Customer gets confirmation by email; finance gets a Slack notification.

For UK orders post-Brexit, the process is similar but with two extra steps: VAT reclamation paperwork (HMRC form C285) and customs declaration handling. The detailed playbook lives in the legal channel [5] — Rachel from finance owns the UK-specific piece.

⚠ Note Source [4] (Finance Ops Playbook) is 6 months old. The Stripe integration changed in August — recommend verifying with Rachel before quoting timelines for the customer.
Logged · index #q-44801
Sources used 5 of 12,470 indexed
1
EU Refund Policy v3
📓 Notion 2 days old Owner: Lukas H.
"Refunds over €1,000 require regional ops manager approval. Approval window: 48h business days…"
Relevance
96%
2
GDPR Compliance Handbook
📁 Drive 14 days old PDF · 24 pages
"…statutory 14-day cooling-off period applies to all EU customers, regardless of order value…"
Relevance
88%
3
Customer Portal Setup
📘 Confluence 1 month old Owner: Web team
"Refund request form: /portal/refunds → routes to ops queue with 48h SLA tag…"
Relevance
84%
4
Finance Ops Playbook
📓 Notion 6 months old ⚠ may be stale
"Refunds processed via Stripe API in the original transaction currency. Confirmation email auto-sent…"
Relevance
78%
5
UK / Brexit Trade Note
💬 Slack #legal 3 months old From: Rachel B.
"For UK refunds we now need HMRC form C285 and customs declaration. I own this — ping me directly…"
Relevance
72%
What ships

What's in the box.

Six things you get with every Knowledge Agent engagement. Real ingestion, real citations, real permissions.

01

Ingest everything you've got.

Notion, Confluence, Google Drive, Slack history, Loom transcripts, contracts, emails, PDFs, internal wikis. We index it all into one searchable layer — including legacy systems.

02

Ask in plain English.

"What's the refund process for orders over $1k in EU?" · "Has anyone tried integrating Tool X before?" · "What did we promise the partner in the August 2024 deal?" Real-language questions, real answers.

03

Answers with citations.

Every answer comes with sources — actual links to the actual document, page, line, or transcript timestamp. No more "I think I remember reading…" Ground truth always reachable.

04

Fresh, automatically.

Source documents update? The agent updates. Notion edited? Reflected within minutes. No stale answers because nobody re-trained the model. The index is alive.

05

Permission-aware.

Engineers see engineering docs. Sales sees sales playbooks. Finance sees finance. We respect your existing permission structure — the agent never leaks across boundaries it shouldn't cross.

06

Slack + web + API.

Ask in Slack. Use the web UI. Hit it via API from your internal tools. Same agent, multiple surfaces, one source of truth.

How it works

Four steps,
end to end.

Step 1

Ingest your sources.

Notion, Drive, Slack, Loom, contracts. We index it all into one searchable layer.

Step 2

Index + embed.

Semantic search over everything. The agent learns the shape of your knowledge.

Step 3

Question asked.

Plain English, in Slack or web. Permission scope automatic.

Step 4

Cited answer.

With links to ground truth. Confidence flagged. Stale sources marked.

Mission control

What the agent
actually knows.

The dashboard your knowledge owner pulls up to see what's covered, what's stale, what people are asking, and where the blind spots are.

MINDCRAFT · KNOWLEDGE COVERAGE Live · re-indexed every 15 min
Documents indexed
12,470
↑ 142 this week
Questions · 24h
238
↑ 12% wow
Confidence avg
94%
"answered" rate
Median answer time
1.4s
incl. retrieval
Sources by platform
N
Notion
4,281· 34%
G
Google Drive
3,342· 27%
#
Slack history
2,789· 22%
C
Confluence
1,284· 10%
Loom transcripts
578· 5%
§
Contracts (PDF)
196· 2%
Top questions · last 7 days
01
How do I provision a staging environment for <client>?
42×
02
Where's the latest version of the brand-guidelines deck?
31×
03
Refund process for orders over $1k in EU?
28×
04
Who owns the <X> tool in finance ops?
24×
05
What's our SLA for enterprise support tickets?
19×
06
How do I add a new vendor to the AP system?
17×
07
What did we agree with <Partner> on the Aug 2024 deal?
14×
Built on

A stack you can audit, port, and own.

Permissions inherit from your existing systems. Nothing trains on your data; nothing leaves your tenant. Optional VPC deployment for regulated industries.

Claude (RAG + answer) pgvector / Pinecone Slack bot Notion / Confluence API Google Drive / Workspace Custom permission layer
−47%

Time-to-productivity cut for new hires at a 60-person services brand. Onboarding-related Slack messages dropped by half. Senior engineers stopped being the human Stack Overflow — and started doing the engineering they were hired for.

Services · 2025
Frequently asked

Questions we get a lot.

Will it hallucinate?
We use retrieval-grounded generation — the agent only answers from your documents, with citations. If it can't find an answer, it says so rather than making one up. We also do regular accuracy checks on a sample of answers each week, with human review.
How do you handle confidential information?
Permissions inherit from your existing systems (Notion, Drive, Slack). The agent only sees what the asker has access to. Nothing trains on your data; nothing leaves your tenant. Optional VPC deployment for regulated industries.
What about really old or scattered information?
We can ingest from almost any source — including legacy systems, email archives, PDF dumps. Quality of answers depends on quality of sources, but the agent surfaces what's there far better than search. And it flags when sources are stale, so you're never quoting old data without knowing it.
Can the agent suggest who to ask?
Yes. If a question has no good answer in the docs, the agent can route to the historically-relevant person ("Mira shipped this feature, she'd know") based on git/Slack/Notion ownership data. Half the value is knowing who knows.
How fresh is the index?
By default, sources are re-indexed every 15 minutes. Critical sources (like the customer-facing FAQ) can be set to real-time webhook updates. Stale-source warnings appear in answers when relevant docs haven't been updated in 90+ days.
What if the answer is wrong?
Every answer has feedback buttons. "Not quite" responses are routed to a human reviewer (configurable — could be the doc owner, or a knowledge-ops person). Patterns of failure get fixed in the source documents, not in prompts — so improvements compound.
What's it cost?
Knowledge agents are an add-on at +$600/mo on top of any tier — or included with Embedded. Pass-through Claude API + vector storage costs typically $80–400/mo depending on document volume and query rate.
Often paired with

Related capabilities.

Stop losing
tribal knowledge.

The free audit includes a sample knowledge agent indexed against three of your real sources. You ask three questions; we show you what answers come back. No commitment, no slides — just a live demo with your data.

Book the audit