Skip to main content
PDGRH

AI Advisor (Digital Chief of Staff) CEO / Admin

Intelligent advisor powered by Claude, covering all 27 Kaltiv modules. 44 AI tools in 4 layers, tenant-specific knowledge base, proactive recommendations, conversational memory and scheduled reports.

Access

Sidebar → AgricultureAI Assistant (chatbot)

Knowledge base management: SettingsAdvisor Knowledge

Or directly via: /settings/advisor-knowledge

The 6 Tabs

The Advisor Knowledge page consolidates all management features in 6 tabs:

TabIconDescription
DocumentsFileUpload documents (PDF, DOCX, XLSX), track embedding status, delete
FactsLightbulbAdd, edit, delete manually taught facts
RecommendationsStarView proactive recommendations generated by AI
MemoryBrainVisualise the assistant's long-term conversational memory
SchedulerClockConfigure automatic reports (daily, weekly, custom)
SourcesLinkConnect external sources (Google Drive, RSS feeds)

Document Management

Tab: Documents

The assistant uses your internal documents to enrich its responses. The processing pipeline is fully automated.

Processing Pipeline

Upload (PDF/DOCX/XLSX)
→ Text extraction
→ Chunking
→ Vector embedding generation
→ Tenant-isolated storage

Available Actions

ActionDescription
UploadDrag and drop or select a file (PDF, DOCX, XLSX)
StatusTrack processing: pending → extraction → embedding → ready
SearchThe assistant automatically queries your documents via vector search
DeleteRemove a document and its associated embeddings
Recommended Documents

Start with your strategic documents: business plan, standard operating procedures (SOPs), annual objectives, audit reports. The richer the knowledge base, the more relevant the recommendations.

Teaching Mode (Facts)

Tab: Facts

Teaching mode lets you inject business knowledge directly into the assistant's memory without uploading documents.

Two Methods

  1. In conversation: say "Remember that..." and the assistant records the fact automatically
    • Example: "Remember that our palm yield target is 18 tonnes/hectare for 2026."
  2. Via the interface: Settings → Advisor Knowledge → Facts tab → Add a Fact button

Fact Management

ActionDescription
AddEnter a fact with a category (objective, process, policy, contact, etc.)
EditUpdate an existing fact when information changes
DeleteRemove an outdated fact
SearchFacts are automatically queried in the triple-source RAG pipeline

Facts are integrated into the hybrid triple RAG search: shared base + tenant documents + taught facts.

Proactive Recommendations

Tab: Recommendations

The assistant generates recommendations based on cross-module data analysis. Recommendations are powered by 6 machine learning models deployed in production.

Trigger Types

TriggerExample
Financial anomalyAbnormally high spending on a budget line
Yield dropProduction below expected threshold (ML yield model)
HR riskRising absenteeism, potential departures (ML turnover model)
Commercial opportunityFavourable price trend, underutilised stock
Lean gapBlocked PDCA cases, declining 5S score, SPC violations
ML predictionYield, quality, price forecasts (oil/nuts/papaya), demand
Preventive maintenanceEquipment at risk of failure (ML maintenance model)

Recommendation Lifecycle

Automatic detection (cross-module analysis)
→ Recommendation generated
→ Notification (sidebar badge)
→ View (Recommendations tab)
→ Action or archive

Recommendations also appear via a notification badge in the sidebar.

Personalised Memory

Tab: Memory

The assistant learns from each conversation to personalise its responses over time.

What the Assistant Memorises

ElementDescription
User preferencesPreferred report format, modules of interest, response language
Accumulated contextRecurring topics, past decisions, usage patterns
Confidence decayOlder memories gradually lose weight — recent information takes priority

Actions

  • View: see the full conversational memory
  • Purge: delete outdated or incorrect memory entries

Memory is isolated by tenant and by user. No sharing between tenants.

Scheduled Reports

Tab: Scheduler

Configure automatic reports generated by the AI assistant on a defined schedule. Reports are executed via Vercel Cron and stored in Supabase Storage.

Report Types

TypeFrequencyContent
Daily briefingEvery morningKPI summary, active alerts, pending tasks
Weekly reviewEvery MondayWeek's performance, trends, recommendations
Custom trackingConfigurableReport on a specific topic at chosen intervals

Configuration

  1. Go to Settings → Advisor Knowledge → Scheduler tab
  2. Click New Scheduled Report
  3. Define: subject, frequency (daily, weekly, monthly, custom), format
  4. The report is generated automatically and accessible in the history

Technical Infrastructure

Scheduled reports use the Vercel Cron endpoint /api/cron/advisor-schedules, which queries the advisor_schedules table and triggers generation via the Claude API.

External Connectors

Tab: Sources

Connect external data sources to automatically feed the knowledge base.

Google Drive

StepAction
1Go to the Sources tab
2Select type Google Drive
3Configure the folder ID to sync
4Synchronisation runs automatically at regular intervals
5New files are automatically indexed and embedded

RSS Feeds

StepAction
1Go to the Sources tab
2Select type RSS
3Enter the RSS feed URL (sector news, market prices, etc.)
4Articles are automatically fetched and integrated into the knowledge base
Multi-source

You can configure multiple sources simultaneously. Each source is synced independently and content is isolated by tenant.

AI Chatbot — 44 Tools

The chatbot is accessible via Agriculture → AI Assistant in the sidebar. It responds in natural language and queries data in real time.

Overview of the 4 Layers

LayerToolsScope
L1 — Core12 toolsHR, Payroll, Leave, Attendance, Daily work, Sales, Stock, CMMS, Accounting, Agriculture
L2 — Lean15 toolsPDCA, 8D, QRQC, Kanban, BSC, SPC, FMEA, 5S, Gemba, OPL, VSM, Takt Time, SMED, TPM, OKR
L3 — Knowledge7 toolsTenant documents, embeddings, taught facts, hybrid triple RAG search, external connectors
L4 — Advisory10 toolsML predictions, proactive recommendations, cross-module correlation, memory, scheduled reports

Example Questions by Layer

L1 — Core:

  • "How many active employees do we have?"
  • "What was last month's revenue?"
  • "Which leave requests are pending approval?"
  • "What's the current carton stock level?"

L2 — Lean:

  • "Which PDCA cases are overdue this week?"
  • "Show the BSC score by perspective."
  • "Are there any Nelson rule violations on the SPC charts?"
  • "Summarise this month's Gemba observations."

L3 — Knowledge:

  • "What do our SOPs say about the harvest procedure?"
  • "Remember that supplier X raised prices by 12%."
  • "What information do we have on OHADA standards?"

L4 — Advisory:

  • "What is the yield forecast for plot A?"
  • "Generate a briefing for the management committee."
  • "What are your recommendations for this week?"
  • "What is the palm oil price trend?"

Access Control (RBAC)

The assistant automatically checks RBAC permissions before each query. A user can only query the modules they have access to.

Use Cases

CEO Daily Briefing

The CEO opens the chatbot each morning and asks:

"Morning briefing."

The assistant responds with:

  • Key KPIs (production, finance, HR)
  • Active alerts (low stock, broken equipment, unvalidated leave)
  • Proactive recommendations for the day
  • Tasks pending approval

This briefing can be automated via the Scheduler to be generated every morning at 7am.

Agricultural Intelligence

The production director asks:

"What is the yield forecast for next quarter and what risk factors should I watch?"

The assistant combines:

  • ML yield and quality prediction models
  • Weather data and production history
  • Uploaded technical documents (SOPs, agronomic guides)
  • Taught facts (targets, field constraints)

Financial Alerts

The finance manager receives a proactive recommendation:

"Alert: fuel expenses are 35% above the 3-month average. Recommendation: check recent purchase orders and audit vehicle usage."

This recommendation is generated automatically by the cross-module correlation engine (L4).

Technical Architecture

Database (7 tables)

TablePurpose
tenant_documentsDocuments uploaded per tenant (PDF, DOCX, XLSX)
tenant_embeddingsEmbedding vectors per document chunk
advisor_factsFacts taught by users
advisor_recommendationsProactive recommendations generated by the AI engine
advisor_memoryLong-term conversational memory
advisor_schedulesScheduled report configuration (Vercel Cron)
tenant_sourcesExternal connectors (Google Drive, RSS)

All tables are isolated by tenant_id with RLS (Row Level Security).

Triple-Source RAG Pipeline

User question
→ Hybrid search:
1. Shared base (general Kaltiv embeddings)
2. Tenant documents (tenant embeddings)
3. Taught facts (tenant facts)
→ Enriched context passed to Claude
→ Response with citations and sources

Permissions

RoleAccess
CEO / AdminAll features: chat, documents, facts, recommendations, memory, sources, scheduled reports
HR AdministratorChat (RBAC-authorised domains), view recommendations
ManagerChat (RBAC-authorised domains)
SupervisorChat (Agriculture, Lean, Operations domains)
EmployeeChat (HR domain: profile, leave, payslips)
/agriculture/chatbot → AI Chatbot (conversation)
/settings/advisor-knowledge → Knowledge base (6 tabs)
Integration

The AI advisor is connected to all Kaltiv modules. Recommendations cross HR, production, finance, stock and Lean data for strategic insights. The 6 deployed ML models (yield, quality, price x3, demand) power real-time predictions.

Screenshots

Screenshots of this module will be added in a future documentation update.