FDA Data Query RAG System

Expert guidance for developing a RAG (Retrieval-Augmented Generation) system prototype to query FDA pharmaceutical data.

Core Principles

Code Design

Demonstrate excellent code organization and design patterns

Keep code simple, small, and concise

Only include behavior directly relevant to the current discussion

Always produce minimal code - avoid over-engineering

Prefer factoring into smaller functions over adding explanatory comments

Code Formatting

Use 4 spaces indentation for TypeScript files

Keep configuration files in their generated format (preserve original indentation and semicolons)

Do not use semicolons in TypeScript code

Preserve formatting conventions in config files

Comments

Keep existing comments unless code changes require updates

Never add comments describing what changed or referencing deleted code

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

Domain-Specific Guidelines

FDA Labels and Chunks

Understanding the data structure is critical:

**Chunk Nature:**

Chunks are text segments extracted from individual FDA labels

Multiple chunks from the same label are likely distinct and independently relevant

Do NOT deduplicate chunks within a single label - they represent different sections

**Deduplication Strategy:**

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

Remove or down-rank chunks with similar text from different manufacturers

The challenge: achieving diversity in retrieved chunks across manufacturers

Similar text from different sources may indicate common pharmaceutical information

**Query Results:**

If multiple chunks from the same label match a query, include both - they likely cover different relevant aspects

Prioritize diversity across manufacturers when possible

Balance relevance scoring with diversity to avoid monopolizing results with a single manufacturer

Implementation Approach

1. **Start Minimal**: Implement only what's being discussed in the current context

2. **Organize Well**: Structure code to reflect clear separation of concerns

3. **Refactor Early**: Break into smaller functions at first sign of complexity

4. **Preserve Simplicity**: Resist adding features or behavior beyond immediate needs

5. **Domain-Aware**: Apply FDA-specific deduplication and ranking strategies appropriately

Example Workflow

When building FDA query features:

1. Parse FDA label data into meaningful chunks (by section/subsection)

2. Index chunks with metadata (manufacturer, drug name, label section)

3. For queries, retrieve relevant chunks using semantic similarity

4. Apply cross-manufacturer deduplication to similar chunks

5. Preserve multiple chunks from same label if independently relevant

6. Return diverse results balancing relevance and manufacturer variety

FDA Data Query RAG System

FDA Data Query RAG System

Core Principles

Code Design

Code Formatting

Comments

Domain-Specific Guidelines

FDA Labels and Chunks

Implementation Approach

Example Workflow

Reviews (0)