Product Requirements Document: GMC For-Benefit Economy RAG Agent¶

Version: 1.0 Date: 2026-06-23 Status: Draft

1. Product Vision¶

An AI agent that empowers GMC economic planning team members to design, evaluate, and iterate on economic proposals—ensuring every decision is grounded in GMC's founding documents, aligned with the Fourth Sector / for-benefit framework, and traceable to authoritative sources.

The agent is not a decision-maker. It is a structured reasoning partner that surfaces relevant knowledge, checks alignment across multiple frameworks, and flags gaps before decisions are made.

2. Problem Statement¶

GMC's economic design is unprecedented: a purpose-built city organized around Gross National Happiness, 8 pillar industries, a SAR legal framework, digital asset integration, and a for-benefit economic philosophy. Team members face several challenges:

2.1 Knowledge Fragmentation¶

GMC documents span multiple sources (gmc.bt, Fourth Sector Group, Wikipedia, news, investment guides, speeches) across English and Dzongkha. No single repository exists. Team members must manually search across sources, often missing relevant context.

2.2 Framework Complexity¶

Economic proposals must simultaneously satisfy: - Alignment with one or more of 8 pillar industries - For-benefit organizational characteristics (2 defining + 8 secondary) - GNH impact across 4 pillars and 9 domains - Sustainability mandates (zero-waste, carbon-negative, 100% renewable) - SAR legal/regulatory compatibility - "Mindful Capitalism" philosophical coherence

No manual checklist exists. Trade-offs between frameworks are hard to identify.

2.3 Decision Reproducibility¶

When team members make economic design decisions, the reasoning and source documents used are not systematically captured. This makes future audits, onboarding, and iteration harder.

2.4 Inconsistent Application of Philosophy¶

Without a structured reasoning framework, different team members may interpret GMC's for-benefit mandate differently, leading to inconsistent economic designs.

3. User Personas¶

3.1 Economic Designer (Primary)¶

A team member drafting a specific economic initiative—an enterprise charter, a licensing framework, a tax incentive structure, a cooperative model, or a digital asset protocol.

Needs: Quickly check if their draft aligns with GMC's values and frameworks. Find precedents. Identify gaps.

Frustration: "I know for-benefit principles are important, but I'm not sure if my proposal meets them. I don't have time to re-read every document."

3.2 Policy Reviewer (Secondary)¶

A team member reviewing proposals from the Economic Designer.

Needs: A systematic, consistent evaluation of any proposal against all applicable frameworks. Traceable reasoning.

Frustration: "How do I compare two competing proposals fairly? My evaluation criteria shift depending on who I ask."

3.3 Knowledge Steward (Tertiary)¶

A team member responsible for document management and the knowledge base.

Needs: Track which documents are ingested, identify gaps, update sources.

Frustration: "Documents are everywhere. I don't know what's been read, what's been considered, and what's missing."

3.4 Governance Lead¶

A team member responsible for ensuring GMC's governance and regulatory framework is followed.

Needs: Verify proposals are within SAR legal bounds. Check licensing precedents.

4. Features & Capabilities¶

4.1 Core Feature: Multi-Framework Proposal Evaluation¶

Given a user's economic proposal (text input), the agent:

Routes the query to the appropriate reasoning path(s) via auto-detection
Retrieves the most relevant document chunks from the vector store
Runs the proposal through each selected reasoning path's structured chain-of-thought
Returns a structured evaluation with:
Alignment scores (per framework)
Evidence citations (document ID + exact text)
Gap analysis (what's missing or needs clarification)
Precedent references (similar existing GMC initiatives)

4.2 Capability: Gap Analyzer¶

Given a partially-defined initiative (e.g., "We want a wellness tourism cooperative in GMC"), the agent:

Identifies which for-benefit characteristics are addressed vs. missing
Flags which pillar-industry requirements are satisfied vs. unknown
Suggests what decisions the team needs to make next
Queries the user with targeted questions to fill gaps

4.3 Capability: Precedent & Case Retrieval¶

Search across all embedded documents for similar initiatives, decisions, or models.

"Has GMC already licensed a fintech company? What was the process?"
"Are there examples of for-benefit cooperatives in wellness tourism?"
"What sustainability requirements did Matrixport have to meet?"

4.4 Capability: Enterprise Architect¶

Step-by-step guided workflow for designing a for-benefit organization within GMC:

Choose pillar industry → 2. Define social purpose → 3. Design earned-income model → 4. Select ownership structure → 5. Set governance rules → 6. Define compensation → 7. Set return limits → 8. Establish transparency protocols → 9. Plan dissolution/protected assets → 10. Check SAR legal compatibility

Each step retrieves relevant documents and asks the user to make specific decisions.

4.5 Capability: Comparative Analysis¶

Compare two proposals side-by-side against all 7 reasoning paths with scored output.

4.6 Capability: Decision Log¶

Every query and response is automatically logged with: - Full user input - Document chunks retrieved (with source IDs) - Reasoning path(s) used - Agent output - Timestamp and session ID

Exportable as markdown or JSON for audit trails.

4.7 Capability: Session Workspaces¶

Named workspaces for different economic workstreams: - "Digital Assets Framework" - "Agri-Tech Co-op Design" - "Wellness Tourism License" - "Youth Employment Plan"

Each workspace maintains its own retrieval context and conversation history.

5. Reasoning Path Architecture¶

5.1 Path Auto-Detection¶

When a user submits a query, the agent first classifies it into one or more reasoning paths:

Query Pattern	Detected Path(s)
"Design a [type] enterprise in [pillar]"	Enterprise Design (Path 2), Pillar Alignment (Path 1)
"Evaluate this proposal for [initiative]"	All relevant paths based on content
"Does this meet GMC's sustainability rules?"	Sustainability (Path 4)
"Is this compatible with the SAR framework?"	Legal/Regulatory (Path 5)
"Compare these two approaches"	Comparative mode
"What's missing from our [sector] approach?"	Ecosystem Mapping (Path 6)

5.2 The 7 Reasoning Paths¶

Path 1: Pillar Alignment Check - Map proposal to each of 8 pillar industries - Score: full alignment / partial / misaligned / conflicting - Identify cross-pillar synergies - Retrieve pillar-specific GMC documents

Path 2: For-Benefit Design Assessment - Check 2 defining characteristics: Social Purpose embedded? Earned Income model? - Check 8 secondary characteristics: Inclusive Ownership, Stakeholder Governance, Fair Compensation, Reasonable Returns, Social/Environmental Responsibility, Transparency, Protected Assets - Flag missing or weak elements - Retrieve Fourth Sector Group documents

Path 3: GNH Impact Analysis - 4 Pillars: spiritual well-being, cultural preservation, good governance, environmental conservation - 9 Domains: psychological well-being, health, education, time use, cultural diversity, community vitality, ecological diversity, living standards, good governance - Score positive/neutral/negative impact per domain - Retrieve GNH framework documents

Path 4: Sustainability & Regenerative Compliance - Zero-waste compatibility - Carbon-negative status - 100% renewable energy plan - Biodiversity corridor impact - Organic food system alignment - Plastic-free mandate - Retrieve Arup/technical sustainability documents

Path 5: Legal / Regulatory / SAR Compatibility - "One Country, Two Systems" governance framework - SAR Basic Law principles - Financial services licensing rules - Digital asset/crypto regulations - Tax incentive eligibility - Business registration requirements - Retrieve legal framework documents and licensing precedents

Path 6: Stakeholder & Ecosystem Mapping - Identify all stakeholders affected - Map required Fourth Sector ecosystem elements: - Financial markets - Public policy - Education & training - Marketing & comms channels - Assessment & reporting standards - Ratings & certification - Technical assistance - Research & understanding - Culture - Connection & representation - Flag which ecosystem elements exist vs. need creation

Path 7: Mindful Capitalism Coherence Check - Balance of prosperity vs. mindfulness - Does the proposal respect Bhutanese cultural identity? - Does it serve Bhutanese youth and future generations? - "Happiness and well-being of people must be the purpose of capitalism" (PM Tobgay) - "Development should not come at all costs, but neither can it be denied" (King) - Check for: long-term orientation, community benefit, cultural preservation, spiritual awareness - Retrieve King's speeches, TIME article, Bitcoin Pledge philosophy

6. Retrieval Strategy¶

6.1 Multi-Stage Pipeline¶

User Query
    │
    ├─ 1. Query Expansion (3 variants)
    │    • Original query
    │    • For-benefit framed rewrite
    │    • Keyword-focused extract
    │
    ├─ 2. Metadata Pre-Filtering
    │    • By pillar industry tag (if detected)
    │    • By document type (if specified)
    │    • By date range (if relevant)
    │
    ├─ 3. Hybrid Search
    │    • Dense (embedding similarity) — top 50
    │    • Sparse (BM25 keyword) — top 50
    │    • Merge & deduplicate → top 75
    │
    ├─ 4. Cross-Encoder Reranking
    │    • Re-score top 75 with cross-encoder
    │    • Keep top 15-25 for context window
    │
    ├─ 5. Context Assembly
    │    • Inject into system prompt with source metadata
    │    • Format: [Source: GMC-004 | Pillar: Finance | Tier: A]
    │
    └─ 6. Response Generation

6.2 Retrieval Modes¶

Mode	k retrieved	Paths run	Use case
Quick	5	1-2 (auto-detected)	Rapid check, informal question
Standard	15	All relevant paths	Proposal evaluation
Deep	30	All paths	Comparative analysis, complex design
Source-find	10	None (pure retrieval)	"Show me where X is mentioned"

6.3 Chunking Strategy¶

Hierarchical approach: - Document-level summary (1 per document): Captures overall purpose, key claims, pillar tags. Stored as separate embedding. - Semantic chunks (~512 tokens, 10% overlap): Paragraph-level splits at natural boundaries (headings, paragraphs, lists). - Sentence-level (~128 tokens): For precise citation of specific claims.

Retrieval: Always retrieve at chunk level, but cite with document-level context.

7. Team Collaboration Features¶

7.1 Session Workspaces¶

/workspace create "digital-assets-framework"
/workspace list
/workspace switch "agri-tech-coops"

Each workspace has: - Independent conversation history - Persistent retrieval context - Workspace-specific metadata filters (e.g., "only retrieve Finance/Digital Assets and legal documents") - Named output exports

7.2 Decision Logging¶

Each session automatically logs to docs/decision-log/{workspace}/{timestamp}.md:

## Decision: [Title]
**Date:** 2026-06-23
**Participants:** [names]
**Proposal:** [text]
**Retrieved Sources:** GMC-002, GMC-004, FSG-002, WIKI-001
**Paths Used:** Pillar Alignment, For-Benefit Assessment, Legal Check
**Agent Output:** [full response]
**Team Decision:** [final choice]
**Rationale:** [notes]

7.3 Disagreement Resolution¶

When two team members interpret a document differently, the agent can: 1. Retrieve the exact source text 2. Show surrounding context 3. Suggest a clarifying question for the governance lead

7.4 Export¶

Export any evaluation as: - Markdown report (for team reading) - JSON (for programmatic use) - Decision log entry (auto-saved)

8. Technical Requirements¶

8.1 Stack¶

Component	Choice	Rationale
LLM	Ollama: Qwen 2.5 (7B or 14B)	Strong multilingual (English + Dzongkha), local, open-source
Embedding	`intfloat/multilingual-e5-large`	Multilingual, state-of-the-art for retrieval
Reranker	`BAAI/bge-reranker-v2-m3`	Cross-encoder, multilingual
Vector Store	ChromaDB (persistent)	Local-first, simple API, good metadata filtering
Full-text Index	SQLite FTS5	BM25 for hybrid search, local
Orchestration	Custom Python (no heavy framework)	Minimal dependencies, full control
Document Store	JSONL + SQLite	Simple, queryable, easy to update
Container	Docker Compose	Ollama + ChromaDB + agent in one stack

8.2 Hardware Requirements (minimum)¶

Component	Minimum	Recommended
RAM	16 GB	32 GB
Storage	10 GB (docs + store)	50 GB
GPU	Not required	8GB+ VRAM for faster generation
Network	Internet for initial document scrape	—

8.3 Performance Requirements¶

Query-to-response time: < 15 seconds (Standard mode)
Embedding ingestion: < 1 second per document
Vector store query: < 500ms for top-50
Concurrent users: 3-5 (team scale)
Context window: 128K tokens (Qwen 2.5 capability)

8.4 Security & Privacy¶

All processing local. No data leaves the machine.
No API keys required (local models only)
Session workspaces are filesystem-isolated
Decision logs are plain markdown (can be version-controlled)

9. Non-Functional Requirements¶

9.1 Transparency¶

Every response must include source citations. The agent never makes a claim without showing the supporting document text.

9.2 Humility¶

When the agent does not have a document that answers the question, it says so. It can ask clarifying questions rather than generating plausible-sounding but unsupported answers.

9.3 Override¶

Team members can override any reasoning path's conclusion with a manual note. The override is logged with the original agent output preserved.

9.4 Extensibility¶

New documents can be added at any time. Re-embedding is incremental. New reasoning paths can be added as Python modules.

9.5 Auditability¶

Every query, every retrieved chunk, every response is logged. Full traceability from user question to source document.

10. Implementation Roadmap¶

Phase 0: Document Ingestion Foundation (Week 1)¶

Scrape all Wave 1 documents from information-directory
Build document storage (JSONL + SQLite)
Implement hierarchical chunking
Generate embeddings for Wave 1
Populate ChromaDB
Verify retrieval quality on sample queries

Phase 1: Core RAG Pipeline (Week 2)¶

Implement hybrid search (dense + BM25)
Integrate cross-encoder reranker
Build context assembly with source citations
Set up Ollama with Qwen 2.5
CLI chat interface (basic Q&A)
Test: "What are GMC's 8 pillar industries?"

Phase 2: Reasoning Path Engine (Week 3)¶

Implement Path 1: Pillar Alignment Check
Implement Path 2: For-Benefit Design Assessment
Implement Path 3: GNH Impact Analysis
Implement Path 4: Sustainability Compliance
Implement Path 5: Legal/Regulatory Check
Implement Path 6: Stakeholder & Ecosystem Mapping
Implement Path 7: Mindful Capitalism Coherence
Path auto-detection classifier
Structured evaluation output format

Phase 3: Agent Workflows (Week 4)¶

Gap Analyzer capability
Precedent & Case Retrieval
Enterprise Architect step-by-step mode
Comparative Analysis

Phase 4: Team Features (Week 5)¶

Phase 5: Expansion & Polish (Week 6)¶

11. Success Metrics¶

Metric	Target	Measurement
Retrieval recall	>85% relevant chunks in top-15	Manual eval on 50 test queries
Citation accuracy	100% sources exist and are correctly attributed	Automated cross-check
User satisfaction	>4/5 rating	Post-session survey
Time saved per proposal	>60% reduction in research time	Measure before/after
Decision log coverage	100% of sessions auto-logged	System check
Reasoning path coverage	>90% of queries routed to at least one path	Classifier accuracy

12. Dependencies & Risks¶

Dependencies¶

Ollama compatibility with target LLMs
Qwen 2.5 Dzongkha capability (preliminary testing needed)
Accessibility of GMC documents (scraping allowed by gmc.bt robots.txt)
Team availability for dogfooding in Phase 5

Risks¶

Risk	Likelihood	Mitigation
GMC documents change frequently	Medium	Version document snapshots; tag with snapshot date
Dzongkha embedding quality poor for specialized legal terms	Low-Medium	Test with sample Dzongkha text; supplement with dictionary
Team adopts agent as oracle (over-reliance)	Low	Design prompt to express uncertainty and ask questions
Scraping blocked by robots.txt	Low	Manual document collection as fallback
LLM hallucinates source documents	Low	Strict prompt constraint: only cite from retrieved chunks

13. Appendix: Key Quotes for System Prompt Injection¶

These core quotes from research should be embedded in the system prompt as grounding principles:

"Happiness and well-being of people must be the purpose of capitalism." — PM Tshering Tobgay

"Development should not come at all costs, but neither can it be denied. It should be carried out in balance where both Mindfulness and Prosperity blend modernity, development and economic prosperity with Bhutan's traditional values." — HM King Jigme Khesar Namgyel Wangchuck

"This is not an experiment. It is a commitment." — Bitcoin Development Pledge

"The project is not just an infrastructure development—it is a People's Project – a nation-building effort, a lifeline for Bhutan, for our economy, security, sovereignty, and ultimately, our future." — HM King, National Day Address 2023

"For-benefits reject the false dichotomy between private interest and public benefit." — Fourth Sector Group