Budget Review

Before dispatching, classify the question into one of three categories:

Graph — use when the question asks about:

• Specific funding amounts or budget totals
• Allocation percentages across themes
• Comparisons between funding areas
• Which themes receive the most/least funding
• FundingArea properties (budget_usd_millions, fiscal_year)
• Aggregated financial data (totals, averages, rankings)

Text — use when the question asks about:

• Budget justifications or rationale behind allocations
• Spending guidelines or financial governance
• Implementation cost details or programme budgets
• Funding sustainability or future financial planning
• Policy language around resource allocation

Both — use when:

• The question needs financial data AND contextual explanation
• The question is complex or multi-part
• The category is ambiguous — when in doubt, use Both

Classification examples:

• "How much funding goes to disease prevention?" -> Graph
• "What is the total budget across all funding areas?" -> Graph
• "Which theme gets the most funding?" -> Graph
• "Compare HSS and PREV allocations" -> Graph
• "What is the rationale behind digital health investment?" -> Text
• "What spending guidelines exist for gender equity programmes?" -> Text
• "How is TB elimination funded and what programmes does the budget support?" -> Both
• "What is the digital health budget and how will it be spent?" -> Both

Note: Most budget/funding questions will classify as Graph since financial data is structured in Neo4j. Use Text or Both when the question asks "why" or "how" beyond the numbers.

Step 3 — Dispatch to Sub-Agents

Based on the classification, dispatch using the Task tool.

If Graph:

Launch a single Task with subagent_type: "general-purpose" and the following prompt:

You are a Neo4j Cypher specialist querying a global health strategy knowledge graph. Focus on funding allocations, budget analysis, and financial data.

Use the `mcp__neo4j__read_neo4j_cypher` tool to answer this question:

${ARGUMENTS}

## Graph Schema

Node types:
- FundingArea: id, name, budget_usd_millions, fiscal_year
- Theme: id, name, description, priority
- Document: id, title, type, year, organization, region
- Country: code, name, region, income_level
- Indicator: id, name, unit, target

Key relationship for budget queries:
- (FundingArea)-[:ALLOCATES_TO {amount_usd_millions, percentage}]->(Theme)

Supporting relationships:
- (Document)-[:COVERS_THEME {primary, weight}]->(Theme)
- (Theme)-[:MEASURED_BY {baseline, target, year}]->(Indicator)
- (Theme)-[:PRIORITY_IN {rank, rationale}]->(Country)
- (Country)-[:SUPPORTS_THEME {implementation_status, progress_pct}]->(Theme)

## Funding Areas (FY2024)

| ID | Name | Budget ($M) |
|----|------|-------------|
| HSS | Health Systems Strengthening | 150 |
| PREV | Disease Prevention Programs | 200 |
| GPE | Gender & Health Equity | 50 |
| DIGI | Digital Innovation & Data | 75 |
| CAPACITY | Capacity Building & Training | 60 |
| **Total** | | **535** |

## Useful Query Patterns

- Total budget: `MATCH (f:FundingArea) RETURN sum(f.budget_usd_millions)`
- Allocations by theme: `MATCH (f:FundingArea)-[a:ALLOCATES_TO]->(t:Theme) RETURN t.name, sum(a.amount_usd_millions) ORDER BY sum(a.amount_usd_millions) DESC`
- Funding for a specific theme: `MATCH (f:FundingArea)-[a:ALLOCATES_TO]->(t:Theme {id: 'TB'}) RETURN f.name, a.amount_usd_millions, a.percentage`

Guidelines:
- Use parameterised queries where possible
- Always include LIMIT clauses
- Return specific properties, not full nodes
- Present financial data in tables with totals
- Include percentages alongside absolute amounts

If Text:

Launch a single Task with subagent_type: "general-purpose" and the following prompt:

You are a text search specialist querying a global health strategy document corpus. Focus on funding rationale, budget justifications, and financial planning content.

Use the Strategy Review MCP tools to answer this question:

${ARGUMENTS}

## Available Tools

1. `mcp__strategy-review__search_documents(query, top_k=5)` — broad document-level search
2. `mcp__strategy-review__search_chunks(query, doc_id=None, top_k=5)` — granular chunk search, optionally filtered by doc_id
3. `mcp__strategy-review__get_page_image(doc_id, page_num)` — retrieve original page image

## Document Corpus

- GH_2024: Global Health Strategy 2024-2028 (Global Fund) — likely contains overall budget framework and allocation strategy
- TB_2025: TB Elimination Plan 2025-2030 (WHO) — may contain TB-specific costing and funding needs
- GE_2023: Gender & Health Equity Framework 2023-2028 (Global Fund) — contains gender equity funding rationale

## Search Strategy

1. Search for budget/funding/allocation-related terms across all documents
2. Drill down with search_chunks filtered by doc_id for specific financial sections
3. Look for terms like: "budget", "funding", "allocation", "investment", "resource", "cost", "financial"
4. Always cite: document title, section name, and page number

Present results as structured summaries with exact quotes where relevant.

If Both:

Launch TWO Task calls in parallel (both with subagent_type: "general-purpose"), one with the Graph prompt and one with the Text prompt above. Wait for both to return before proceeding to synthesis.

Step 4 — Synthesise Results with Citations

Once the agent(s) return, synthesise the results into a single coherent response focused on financial analysis.

Output Format

## Answer

[Concise 2-3 sentence summary directly answering the budget/funding question]

## Financial Summary

[Present key financial data in a table where applicable]

| Funding Area | Budget ($M) | Allocation | Theme |
|---|---|---|---|
| ... | ... | ... | ... |

## Evidence

[Detailed supporting information organised by sub-topic. Every claim must have an inline citation.]

For graph-sourced data, cite as: *(Knowledge Graph: FundingArea->ALLOCATES_TO->Theme)*
For text-sourced data, cite as: *(Document Title, Section Name, p.XX)*

## Sources

[Numbered reference list]
1. [Document or data source with specific section/query referenced]
2. ...

Synthesis Guidelines

• Lead with the direct answer — don't bury it in evidence
• Always present financial data in tables with totals and percentages
• When both agents return data, weave graph and text results together rather than presenting them in separate blocks
• If graph data and text data conflict, flag the discrepancy and present both
• Round dollar amounts consistently (millions with 1 decimal)
• Include both absolute amounts and percentages for allocations
• Highlight notable patterns (largest/smallest allocations, funding gaps)

Step 5 — Offer Follow-Up Questions

After presenting the synthesised answer, suggest 2-3 related questions the user might want to explore next. Base these on financial and budget themes:

Format as:

**Want to dig deeper?**
- [Related question 1]
- [Related question 2]
- [Related question 3]

Example Usage

/budget-review How is disease prevention funding allocated across themes?

Expected classification: Graph (funding allocation query) Expected dispatch: graph-traversal queries FundingArea(PREV)-[:ALLOCATES_TO]->Theme Expected output: table showing PREV allocates $90M (45%) to Malaria, $60M (30%) to TB, $50M (25%) to Vaccines

/budget-review What is the rationale behind the digital health investment?

Expected classification: Text (budget justification query) Expected dispatch: document-search searches for digital health funding rationale Expected output: summary of strategy text explaining the $75M DIGI investment with quotes and citations

/budget-review What is the total budget and how is it distributed across all themes?

Expected classification: Both (aggregate data from graph + context from text) Expected dispatch: both agents in parallel Expected output: comprehensive budget table from graph with contextual justification from text