Qatar Post AI

Solution Docs · powered by Eloquent

Qatar Post Internal AI Assistant

Solution & Requirements Document

Version 1.1 — May 12, 2026

Prepared by Eloquent for Qatar Post

1. Executive Summary

This document specifies the Internal AI Assistant solution for Qatar Post employees, built on the Eloquent platform. The assistant enables employees to ask natural-language questions about Qatar Post’s HR, IT, Finance, and operational policies in Arabic or English; receive answers grounded in the actual policy documents with inline citations; attach files for in-context analysis; and complete service-desk transactions — such as opening an IT ticket — directly from the conversation.

Compared with the v1.0 scope captured in May 2025, the Eloquent platform has matured significantly. v1.1 of this document reflects current capabilities including:

  • A full Agent Builder that lets Qatar Post administrators create role-specific assistants — HR Helper, IT Support, Finance Q&A — each with its own knowledge sources, skills, and tools, without writing code.
  • A pre-built library of specialist agents (Financial Analyst, Legal Advisor, HR Advisor, Account Executive, Content Writer, Customer Success Manager) that Qatar Post can adopt out-of-the-box or use as templates for tailored personas.
  • Document-grounded citations that open the source PDF, Word, or PowerPoint file inline with the cited region highlighted on the correct page.
  • An AI Brain panel that shows employees exactly which knowledge-base searches the agent ran while preparing each answer — full transparency, not a black box.
  • A Library of shared knowledge with nested folders, organised per business area (HR, Finance, Legal, IT, etc.).
  • A consolidated Analytics Dashboard with chat volume, active users, token consumption, model cost, usage trends, and per-model breakdowns.
  • A Projects & Tasks workspace with Kanban-style task tracking (To-Do / In Progress / Review / Done / Blocked) — useful for service-desk follow-up workflows.
  • Multi-tenant isolation with per-organisation database schemas and JWT-secured API access.

The remainder of this document is organised as a solution specification: each functional area describes what the user sees, the relevant admin controls, and the supporting capability on the Eloquent platform.


2. Solution Overview

The Qatar Post Internal AI Assistant runs on the Eloquent platform, a multi-tenant system designed for enterprise AI assistants grounded in organisational documents. It operates across three personas:

PersonaDescriptionPrimary Surface
EmployeeAny Qatar Post staff member. Asks policy questions, attaches documents for advice, opens IT/HR tickets.Chat interface (web).
Department AdminDepartment lead (IT, HR, Facilities) responsible for ticket triage and knowledge curation in their area.Library + Service Desk admin tabs.
Platform AdminQatar Post IT/Information Management role. Owns agents, knowledge sources, users, and analytics.Eloquent Admin Console (full).

3. End-User Experience

3.1 Conversational Q&A

The assistant answers employee questions about Qatar Post policies and regulations — HR Policy, IT Policy, Finance Manual, Information Security Framework, Acceptable Use Policy, and any other documents Qatar Post chooses to ingest.

Employee asks “How to request a new laptop?” The agent answers using the IT Service Catalogue, shows source chips below the answer, and opens the cited PDF in-place. The right-side AI Brain panel lists the knowledge-base searches the agent ran to compose the response.

Features visible in the screenshot:

  • Streaming answers: text appears word-by-word.
  • Multi-turn: full conversational context preserved within a chat session.
  • Bilingual: Arabic and English supported in the same conversation, including mixed-language input.
  • Persistent history: the left sidebar lists prior chats; users can return to any of them.
  • Source chips below the answer: each answer carries a strip of source-document references; clicking a chip opens the document.
  • Inline source viewer: the cited PDF opens beside the conversation in a read-only viewer — no download needed.
  • AI Brain panel (right): shows which knowledge-base searches the agent ran in real time. See §3.7.

3.2 Document-Grounded Citations

Every answer that uses an internal source includes inline citations. Clicking a citation opens the source viewer beside the conversation, with the cited region highlighted on the correct page.

Citation drill-down — the agent answers a Qatar corporate-tax question using the Simplified Guide on Capital Gains Tax (2026).pdf. Each cited paragraph is highlighted in green on the page where it occurs, and the right-hand panel lists every passage the agent retrieved.

Supported behaviours:

  • PDF citations: render the original PDF with passage-level highlights on the cited page. Multiple page references are listed (e.g. pp. 1–3, 5, 7–9).
  • Word and PowerPoint: rendered inline using an embedded Office viewer in read-only mode.
  • Markdown / web articles: rendered with the original URL available to open in a new tab.
  • Source cards: each answer displays a strip of source-document chips below the message; each chip shows file icon, title, page ranges, and reference count.
  • Verifiable: an employee can always click through to the exact passage the assistant used — no opaque answers.

“Cites authoritative sources. Answers grounded in publications from the same authorities your auditors, lawyers, and HR managers cite every day.”

3.3 File Attachments in Chat

Employees can attach files directly to a message to get answers about their specific situation. Supported attachment classes:

  • Documents (PDF, DOCX, PPTX) — analysed and referenced like any internal source for the duration of the chat.
  • Spreadsheets (XLSX, CSV) — read row-by-row; the assistant can answer questions about the data and compute simple aggregations.
  • Data files — generic structured-data uploads.

Attachments appear as removable chips below the input. The assistant indicates when a file has finished analysis so the user knows when they can ask follow-up questions about it.

3.4 Voice Mode

Employees can switch the chat into Voice mode to speak rather than type. Voice mode uses the browser’s Web Speech API with both Arabic and English recognition. Recognised speech is streamed into the conversation as the user speaks.

3.5 Chat History & Resumption

A left sidebar lists the user’s recent chats, paginated 15 at a time with on-demand “Load more”. Clicking a chat re-opens the full transcript so the user can resume where they left off. The current chat title auto-updates from the first user message.

3.6 In-line Helpdesk Ticket Creation

When an employee describes a problem that the assistant cannot resolve through Q&A — for example “my VPN is not connecting” — the assistant offers to open a Helpdesk ticket on the employee’s behalf. The assistant:

  1. Asks a small number of clarifying questions configured for the relevant department.
  2. Confirms a summary with the user.
  3. Calls the configured ticketing tool (the built-in Service Desk, Jira, ServiceNow, or a custom webhook to Qatar Post’s existing ticketing system).
  4. Returns the ticket reference number in-chat so the employee has immediate confirmation.

The same flow supports HR requests, facilities issues, and any other department a Platform Admin chooses to enable.

3.7 AI Brain — Retrieval Transparency Panel

Every chat surface includes an AI Brain panel to the right of the conversation. As the agent prepares each answer, the panel streams a live view of the actions the agent took:

  • Which knowledge-base searches it ran (and the queries).
  • Which documents came back from each search.
  • Which tools it called and with what arguments.
  • Any intermediate reasoning steps the agent chose to surface.

This turns the assistant from a black box into an auditable workflow. An employee — or an internal auditor reviewing a transcript — can always see exactly where each fact in an answer came from. For Qatar Post this materially raises trust and lowers the bar for adopting AI in regulated areas.


4. Agent Builder

The Agent Builder is the central administrative surface for defining what the assistant is, what it knows, and what it can do. Qatar Post Platform Admins can create one general-purpose assistant or several role-specific ones (for example: IT Helper, HR Q&A, Finance Reference).

4.1 Agent Management & Specialist Library

Eloquent ships with a library of pre-built specialist agents. Qatar Post can adopt them as-is, clone them as starting points, or build new agents from scratch.

Agent picker — SAIE the generalist on top, with six specialist agents below: Financial Analyst, Legal Advisor, HR Advisor, Account Executive, Content Writer, Customer Success Manager. The employee picks the right specialist for the task, or stays with the generalist for anything else.

For the Qatar Post rollout, the catalogue will typically include:

  • SAIE (Generalist) — entry point for any employee question; routes to specialists when needed.
  • HR Helper — Qatar Post HR policy, leave, salary increase requests, EOSB.
  • IT Helper — Qatar Post IT policy, service catalogue, ticket creation.
  • Finance Reference — Finance Manual, expense procedures, audit trail.
  • InfoSec Advisor — Information Security Framework, Acceptable Use Policy, Information Classification.

Agent Management capabilities:

  • List view: card-style listing of all agents in the tenant with status and last-modified information.
  • Create new: launches the Agent Builder with sensible defaults.
  • Import / Export: full agent configuration is portable as a JSON document — useful for promoting from staging to production or for backups.
  • Versioning: each agent carries a semantic version label so admins can track changes to instructions and configuration over time.

4.2 Designing an Agent

The builder exposes the following per-agent fields:

FieldPurpose
NameDisplay name shown to employees.
AvatarUploaded image or AI-generated SVG avatar.
GoalA one-paragraph description of the agent’s job. Used to auto-generate the system prompt.
System promptThe full instruction set sent to the model. Auto-generated from the Goal, fully editable.
SkillsThe capabilities attached to this agent (see §4.3).
ModelSelection from the available foundation models.

The builder uses a visual layout so admins can see the agent, its skills, and the tools inside each skill as a single diagram rather than a long form.

4.3 Skills & Tools

Eloquent uses a two-level model:

  • A Skill is a coherent capability — for example Knowledge Base, Open IT Ticket, HR Request, Document Drafting.
  • Each skill has its own skill instructions (a system-prompt fragment that activates when the skill is in use) and a small set of Tools the agent may invoke while the skill is active.
Skill configuration — the Knowledge Base skill is configured with custom instructions (“Use knowledge_search to find internal policies; only answer based on search results, not general knowledge”) and one tool: Search Knowledge Base, pointed at the internal assistant library with a max of 10 results.

The configuration shown in the screenshot is the recommended starting point for the Qatar Post generalist: a Knowledge Base skill bound to the internal-assistant library, with explicit instructions forcing the agent to ground every answer in retrieved content rather than its general training data.

Additional skill types relevant to Qatar Post:

  • Open IT Ticket — calls into the Service Desk or an external ITSM system (see §4.4).
  • Open HR Request — same pattern, routed to HR.
  • Document Drafting — drafts memos, emails, request templates inline.
  • Custom skill — defined by admins for QP-specific workflows.

4.4 Helpdesk / ITSM Integration

Helpdesk ticketing is delivered as a Skill whose tools call out to the configured backend system. Four backends are supported out of the box:

  1. Eloquent Service Desk (built in) — tickets are routed to department-scoped administrators within the platform.
  2. Jira Service Management — via API tool with credentials configured in the admin console.
  3. ServiceNow — via API tool.
  4. Custom / in-house — via Webhook tool to any internal endpoint Qatar Post exposes.

This flexibility means the assistant can integrate with whatever ticketing system Qatar Post already operates without changing the chat experience for the employee.

4.5 Embeddable Chat Widget

Any agent can be deployed as a chat widget that drops into an existing internal portal with a single script tag. The Embed Customiser exposes four tabs:

TabControls
AppearancePrimary/secondary colours, button style, position, fonts, dark/light mode.
BehaviourAuto-open on page load, persistence across navigation, retry behaviour, default greeting.
SecurityAllowed-domain whitelist, CORS origins, JWT-based identity passthrough so the widget knows the logged-in user.
Embed CodeAuto-generated <script> snippet ready to paste into the host page.

A live preview panel shows the widget as the admin tunes it.


5. Knowledge & Library Management

Qatar Post Platform Admins are responsible for keeping the assistant’s knowledge base current. Ingestion is fully self-service.

5.1 Document Ingestion Pipeline

Admins drag-and-drop one or many files into a library. Each file becomes an ingestion job tracked through five states:

Pending → Processing → Completed, with branches to Failed (with diagnostic message) and Cancelled.

The pipeline performs automatically:

  • Format normalisation — DOCX, PDF, PPTX, XLSX, MD, TXT, HTML.
  • Chunking — documents are split into semantically meaningful chunks for retrieval.
  • Embedding — each chunk is converted to a vector representation for similarity search at query time.
  • Indexing — chunks become available to agents within minutes of ingestion completing.

Each job shows total files, total size, chunk count, and timing information so admins can confirm coverage.

5.2 Knowledge Management

The Knowledge Management section of the Admin Console is the home for all things document-related: libraries, individual documents, ingestion jobs, and the underlying knowledge graph that supports relationship-aware retrieval.

Capabilities:

  • Multiple knowledge bases per organisation — e.g. HR Policy, IT Policy, Finance Manuals — assigned to specific agents.
  • Schema designer — define node and edge types for richer, graph-based retrieval where simple flat search is not enough.
  • Explorer — visualise what the assistant can see, run queries, and audit coverage.
  • Automated ingestion — schedule recurring bulk ingestion (e.g. a folder synced from SharePoint).
  • Website ingestion — crawl an internal portal with depth and same-domain controls.

5.3 Bulk & Website Ingestion

Where policies are published on Qatar Post’s internal web portal, the Website ingestion flow can crawl a set of URLs and ingest the pages into a library. Bulk ingestion supports folder-level batch processing for one-shot migrations from a SharePoint share or a network drive.

5.4 Library

The Library is the day-to-day surface for organising ingested documents. It groups them into named, shareable collections — each agent can be assigned one or more libraries, so an HR agent only sees HR libraries while an IT agent only sees IT libraries.

Library landing — eight shared libraries are visible at the top level: internal assistant, CX, Marketing, Sales, HR, Legal, Finance, and Org Documents. Each library is owned by a business area and can be assigned to one or more agents.

Within a library, documents can be organised into nested folders so large policy sets stay navigable.

Inside the Finance Library — the QA subsection groups Banking Regulations, Income Tax, QFC Tax, and Transfer Pricing folders. Folder structure scales to arbitrary depth.

Library features:

  • Create / rename / delete libraries from the landing page.
  • Visibility scopePublic (All Organization) or restricted to specific roles.
  • Nested folder hierarchy with breadcrumb navigation.
  • Document operations — upload, preview, replace (versioning), download, delete.
  • Assign to agents — a single library can be shared across several agents (e.g. HR and Finance both reference Org Documents).
  • Seed libraries — Eloquent ships sample libraries that can be cloned and adapted, accelerating initial rollout.

6. Admin Dashboards

6.1 Chat Log Dashboard

Platform Admins can browse every conversation between employees and the assistant. The Chat Log lists chats with metadata and provides a transcript view for detail.

Columns and controls:

  • Chat ID — internal reference.
  • User — Qatar Post email address.
  • Agent — which agent handled the conversation.
  • Category (classification) — HR, IT, Facilities, Policy, Document, General.
  • Start time / End time — full timestamp.
  • Duration — wall-clock chat length.
  • Messages — message count.
  • Statuscompleted / active / ended / archived.
  • CSAT score — captured from the in-chat satisfaction prompt.

Filters: date range, agent selector, user email, free-text search by chat ID. Clicking a row opens the Chat Details drawer with the full transcript and the session metadata above.

6.2 Analytics Dashboard

The Analytics page aggregates conversation data across agents and time ranges.

Agent Analytics — top-of-page KPIs (Total Chats, Total Messages, Active Users, Total Tokens, Total Cost, Avg Session). The Usage Trends chart plots chats, messages, and users over the selected period. Below, Token Distribution by Model and Cost by Model break down model spend.

Available views:

Metric classSpecific measures
VolumeTotal chats, total messages, active users.
EngagementAverage session duration, usage trends over time.
CostTotal tokens consumed, total API cost (USD), per-model breakdown.
Model mixToken distribution by model, cost by model — useful for cost optimisation.
ClassificationBreakdown of chats by category (HR, IT, Facilities, Policy, Document, General).
OutcomesTickets created, documents generated, task-completion rate.
Tool usageFrequency each agent tool was invoked.
CSATAverage satisfaction score, response count, word-cloud of free-text feedback.
Answered vs UnansweredDerived from classification + outcome data — surfaces topics where the assistant fell back to “I don’t know”.

Filters: date range, agent selector. Data refreshes on demand.

6.3 Service Desk Admin

The Service Desk surface lets Qatar Post manage how tickets opened by the assistant are routed and structured. Two tabs:

Admin Accounts

  • Define administrators per department (IT, HR, Facilities, or any Custom department such as Procurement or Legal).
  • Each department gets a colour code (IT blue, HR emerald, Facilities amber, Custom purple) for quick visual scanning.
  • Actions: create, update, revoke permissions, delete.

Field Managers

  • Define which custom fields appear on a ticket created in each department. For example IT tickets might require Asset ID and Severity, while HR tickets require Subject Type and Confidentiality.
  • Fields can be required or optional, text/select/date.
  • The assistant uses these field definitions at runtime to ask the employee the right clarifying questions before submitting the ticket.

6.4 Team & Role Management

Found under Organisation Settings → Team. Functions:

  • Invite team members by email.
  • Assign role: Admin (full access to Admin Console) or User (chat only).
  • Resend or revoke pending invitations.
  • Update an existing member’s role.
  • Remove a member.

Combined with the Service Desk admin tab, this gives Qatar Post complete control over who can access the Admin Console and what they see inside it.


7. Projects & Tasks Workspace

Beyond chat and tickets, Eloquent provides a lightweight Projects surface where the assistant — or an employee — can capture follow-up work. This is useful for service-desk workflows where a ticket is only the first step of a multi-stage task.

Projects workspace — Kanban board with five status columns (TO DO, IN PROGRESS, REVIEW, DONE, BLOCKED), tabs for Tasks / Docs / Releases, and filters by Status, Priority, and Type. Multiple view modes (Column, Table, Tree) make the same data navigable for different roles.

Highlights:

  • Tasks — Kanban or table view; status flow To Do → In Progress → Review → Done with a Blocked sideline.
  • Docs — collaborative documents attached to a project.
  • Releases — group tasks into release milestones for larger initiatives.
  • Views — Column (Kanban), Table, and Tree views over the same task data.
  • Filters — by Status, Priority, Type, and free-text search.
  • Created from chat — the assistant can open a task from a conversation (e.g. “Track my laptop request to completion”), giving Qatar Post a single audit trail per request.

8. Multi-Tenancy, Security & Compliance

The Eloquent platform is built for organisations that cannot tolerate data co-mingling.

  • Per-organisation isolation: each tenant — Qatar Post is one tenant — has its own PostgreSQL schema (org_<uuid>) and its own ClickHouse database. There is no shared table where rows from multiple tenants live side-by-side.
  • JWT authentication: every API call carries a signed token. The API Gateway validates the token, extracts orgId and userId, and forwards them to backend services as headers. Backend services use those headers, not anything in the request body, to scope queries.
  • No direct backend exposure: the only externally reachable component is the API Gateway. NATS, internal services, and databases live on a private network.
  • Audit trail: every conversation is persisted in the Chat Log, giving Qatar Post a complete record of what the assistant told whom and when.
  • Right-to-erasure: a user’s chats can be archived or deleted on request; tenant-wide export is supported for compliance reviews.
  • Data residency: deployments can be pinned to a specific region; on-premise and air-gapped installations are available.

9. Mapping: v1.0 Requirements → v1.1 Solution

The table below shows how each requirement captured in the May 2025 document is satisfied — or extended — in the current Eloquent platform.

v1.0 Sectionv1.0 Requirement (summary)v1.1 Solution SectionStatus
2.1Assistant Q/A — answer employee questions from policy documents.§3.1 Conversational Q&ADelivered, extended with streaming, bilingual support, and the AI Brain transparency panel.
2.2Document Citation — direct reference to source documents.§3.2 Document-Grounded CitationsDelivered, extended with inline PDF/Office viewers and passage-level highlights.
2.3Chat Log Dashboard — admin views transcripts & metadata.§6.1 Chat Log DashboardDelivered, columns and filters refreshed.
2.4Helpdesk Integration — assistant opens tickets.§3.6 + §4.4 + §6.3Delivered, extended with multiple ITSM connectors and the built-in Service Desk.
2.6Analytics Dashboard — most-asked topics, answered vs unanswered, user feedback (CSAT).§6.2 Analytics DashboardDelivered, extended with classification, tool usage, token cost, per-model spend, and word-cloud CSAT.
2.7Document Ingestion — admins upload new documents.§5.1 + §5.2 + §5.3 + §5.4Delivered, extended with the Library (nested folders, per-agent assignment), Knowledge Management, website crawl, and automated batch ingestion.
(Not in v1.0) Agent Builder, multi-agent configuration.§4 Agent BuilderNew capability added since v1.0.
(Not in v1.0) Specialist agent library.§4.1New capability.
(Not in v1.0) AI Brain retrieval transparency panel.§3.7New capability.
(Not in v1.0) Voice mode (Arabic / English).§3.4 Voice ModeNew capability.
(Not in v1.0) Embeddable widget.§4.5 Embeddable Chat WidgetNew capability.
(Not in v1.0) File attachments in chat.§3.3 File AttachmentsNew capability.
(Not in v1.0) Service Desk admin (department-scoped admins, custom fields).§6.3 Service Desk AdminNew capability.
(Not in v1.0) Team & role management.§6.4 Team & Role ManagementNew capability.
(Not in v1.0) Projects & Tasks workspace.§7 Projects & Tasks WorkspaceNew capability.

10. Roadmap & Out-of-Scope

This section is intentionally left for Qatar Post to populate with their phased rollout plan and explicit out-of-scope items. Suggested headings:

  • Phase 1 (Pilot): scope, departments, success criteria.
  • Phase 2 (General Availability): expansion plan.
  • Phase 3 (Optional): additional connectors, language expansion, mobile app.
  • Explicit out-of-scope items.

11. Glossary

TermMeaning
AgentA configured assistant persona inside Eloquent — has a goal, system prompt, skills, and knowledge sources.
SkillA coherent capability attached to an agent (e.g. Knowledge Base, Open IT Ticket). Each skill has its own instructions and tools.
ToolA specific action the agent can take inside a skill (lookup, ticket creation, API call).
AI BrainThe right-hand panel in the chat UI that streams the agent’s retrieval steps and tool calls in real time.
LibraryA named, shareable collection of ingested documents. Agents are assigned one or more libraries.
IngestionThe process of loading documents into the assistant’s knowledge base.
EmbeddingA vector representation of a chunk of text used for similarity search.
Knowledge ManagementThe admin section that owns libraries, documents, ingestion jobs, and graph schema.
CSATCustomer Satisfaction score, captured by an in-chat survey.
Tenant / OrganisationAn isolated Eloquent customer environment. Qatar Post is one tenant.
Service DeskEloquent’s built-in ticketing surface with department-scoped admins.
Embed widgetA chat widget that can be dropped into an existing internal web page via a <script> tag.
ProjectA Kanban-style workspace for follow-up tasks, optionally created from chat.