Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.teeem-ai.com/llms.txt

Use this file to discover all available pages before exploring further.

Teeem AI never just executes. Every response goes through the following six steps. Where ChatGPT-style generative AI replies with text, Teeem AI verifies permissions, gathers context, surfaces insight, and follows through to approvable actions.

​
The six steps

1

Question — natural-language request from a teammate

The request arrives via channel mention, DM, or web chat input — depending on the channel policy.
@Teeem Update the company KPI dashboard with this month's data
and build a 5-slide PPT for leadership. Post it to #c-leadership.
At this point Teeem AI identifies intent, channel, sender, and origin (which channel, which thread).
2

Permission check — user, department, access verification

Before doing anything, the agent checks whether the sender has:
CheckExample
RBAC roleIs the sender at least power_user to access KPIs?
ABAC data classThe KPI is confidential — only their team can see it
Channel ACLCan they post to #c-leadership?
Tool policyIs a high-impact tool like manage_website enabled?
External-send gateDoes this require a human approver?
If permission is missing, the work stops there and the user gets a clear, actionable explanation.
3

Context integration — connect related information

Information is collected from:
  • Knowledge base — KPI definitions, report templates
  • App Packs — current month’s revenue, orders, inventory
  • ERP connectors — additional data from Douzone or Wehago
  • Channel history — past reporting patterns in the same channel
  • Prior results — last year’s KPI for the same period
BM25 (Korean tokenizer) plus vector search are combined via Reciprocal Rank Fusion to pull only the most relevant snippets into context.
Long conversations stay cheap and accurate via automatic compression: tool outputs older than three turns are replaced with summaries, and the context window auto-compacts at 70%.
4

Insight — find the change points

Not a list of data — meaningful signal:
  • Change points — revenue −15%, inventory turnover +20%
  • Anomalies — values outside the normal range
  • Pattern shifts — account concentration up vs last quarter
  • Recommendations — “investigate shipping-delay impact at account A”
Tool calls happen frequently here: forecast, alert_monitor, analyze_data, diff_documents.
5

Execute — reports · approvals · queries · cleanup

Beyond the answer, the agent does the actual work — when permissions allow.
ActionTools
Build a 5-slide PPT (this month’s KPIs + charts)generate_pptx, generate_chart
Apply company fonts and colour scheme(organisation template)
Attach to #c-leadership with a summary messageupload_to_channel
DM the sender on completionslack_mention
For sensitive external sends, an approval gate triggers here and waits for a human OK before delivery.
6

Record — every step into the tamper-evident log

Everything this single response did is written into the hash-chain audit log:
  • Sender, message, channel
  • Permission-check results
  • Tool calls with inputs and outputs (PII-masked)
  • External system calls (Douzone API, etc.)
  • Final response content
  • Cited KB document IDs
Logs are retained for 5 years; the integrity-verification API confirms tamper-evidence at any chosen point.

​
A real example — one message, six steps

How the same request unfolds across the steps: 
@Teeem Summarise last week's revenue by account and build a PPT.
Flag any anomalies separately.
  • Intent: report generation
  • Output: PPTX
  • Time range: last week
  • Group by: account
  • Highlight: anomalies
  • Sender: kim@company.com (Slack ID mapped)
  • kim@company.com is power_user ✓
  • Revenue data is confidential — power_user allowed ✓
  • PPTX generation enabled ✓
  • No external send → no extra gate ✓
  • Query App Pack sales_orders for 2026-04-21 to 2026-04-27
  • Aggregate by account + compare to prior week
  • Pull standard report PPTX template from KB
  • Inherit colour scheme from previous similar requests
  • Total revenue: ₩120M (−15% vs prior week)
  • Account A: −₩30M (largest impact)
  • Accounts B, C: within normal range
  • Account D: new revenue +₩15M
  • Anomaly candidate: A (likely shipping delay)
  • generate_pptx — cover + summary + per-account charts + anomaly slide
  • generate_chart ×4 — bars by account, trend line, comparison doughnut, anomaly highlight
  • Attach PPTX to the Slack thread
  • Reply: “Last week’s revenue report. Biggest change: account A (−₩30M, likely shipping delay).”
  • Citation card (App Pack model + KB template) posted as a separate persistent message
The hash-chain audit log gets these entries:
{"ts":"2026-05-02T10:23:11Z","actor":"kim@company.com","action":"agent.request","msg_hash":"...","channel":"slack:#sales"}
{"ts":"...","actor":"system","action":"rbac.check","result":"allow","prev_hash":"..."}
{"ts":"...","actor":"system","action":"app_query","model":"sales_orders","filter":"...","prev_hash":"..."}
{"ts":"...","actor":"system","action":"tool.generate_pptx","output_hash":"...","prev_hash":"..."}
{"ts":"...","actor":"system","action":"agent.response","msg_hash":"...","prev_hash":"..."}

​
What’s different — vs generative AI

ChatGPT-style generative AITeeem AI
Output shapeText responseText + tool execution + file generation + system calls
Data connectionLimited to model training dataCompany ERP, groupware, apps, mail, calendar, e-commerce
PermissionsNot bound to user identityPer-user RBAC + ABAC + channel ACL
AuditConversation history (manual backup)Tamper-evident hash chain (5 years)
CitationsGenerally noneEvery grounded answer comes with sources
DeploymentSaaS onlySaaS + on-prem (negotiated)

​
Next

Permission model

Which checks run, in detail.

Citations and sources

What the source cards next to every answer mean.