ChatFDA RAG System Developer

An expert assistant for developing and maintaining the ChatFDA RAG (Retrieval-Augmented Generation) system prototype designed to query FDA pharmaceutical label data.

What This Skill Does

This skill guides AI agents to work on the ChatFDA codebase with proper understanding of:

Instructions

Code Design Principles

1. **Maintain Good Code Organization**

- Demonstrate clean separation of concerns

- Keep code simple, small, and concise

- Only include behavior directly relevant to the current discussion

- Always produce minimal code — no speculative features or abstractions

2. **Avoid Over-Engineering**

- Don't add functionality beyond what is being discussed

- Prefer simple solutions over complex ones

- Factor complex logic into smaller functions rather than adding explanatory comments

Formatting Standards

1. **TypeScript Code**

- Use 4 spaces for indentation (not 2)

- Do NOT use semicolons in TypeScript code

2. **Configuration Files**

- Keep formatting as originally generated

- Preserve semicolons if they exist in config files

- Preserve original indentation in config files

Comment Guidelines

1. **Preserve Existing Comments**

- Keep existing comments unless they need updates to reflect code changes

- Update comments when the code they describe changes

2. **Never Add These Types of Comments**

- Comments describing what changed in a diff

- References to deleted code

- Change logs or version history

3. **When to Add Comments**

- Use judiciously — prefer refactoring to smaller, self-documenting functions

- Only add when the domain logic is genuinely non-obvious

Domain Knowledge: FDA Labels and Chunks

1. **Understanding Chunks**

- Chunks are segments of text extracted from a single FDA label

- Two chunks from the same label are typically different pieces of information

- Both chunks from a label may be relevant to a query — don't deduplicate by label

2. **Deduplication Strategy**

- Deduplication means removing or ranking lower chunks with similar text from DIFFERENT manufacturers

- The challenge is achieving diversity in retrieved chunks

- Focus on cross-manufacturer deduplication, not intra-label deduplication

3. **Query Results**

- Multiple chunks from the same label in results is expected and valid

- Prioritize semantic diversity across manufacturers over uniqueness by source label

Usage Examples

**Good:**

```typescript

// Query: "Add a function to rank chunks by manufacturer diversity"

// AI produces a small, focused function with 4-space indentation, no semicolons

function rankByManufacturerDiversity(chunks: Chunk[]): Chunk[] {

const seen = new Set<string>()

return chunks.filter(chunk => {

if (seen.has(chunk.manufacturer)) return false

seen.add(chunk.manufacturer)

return true

})

}

```

**Bad:**

```typescript

// Query: "Add a function to rank chunks"

// AI adds unnecessary features, uses 2-space indentation, adds semicolons

function rankChunks(chunks: Chunk[], options?: RankingOptions): Chunk[] {

// Added: 2024-01-15 - Initial implementation

const seen = new Set<string>();

// Support multiple ranking strategies (not requested)

const strategy = options?.strategy || 'default';

// ... 50+ lines of unnecessary abstraction

}

```

Constraints