RAG Chat (Swan)

Swan is MealPal’s AI nutrition chatbot, built with RAG (Retrieval-Augmented Generation) for accurate, source-backed answers.

Architecture

1. User asks a question via the chat interface
2. Query classification — determines if the query is temporal (date-based) or semantic
3. Retrieval:
   • Temporal queries (e.g., “what did I eat yesterday”) → direct database queries
   • Semantic queries (e.g., “how much protein do I need”) → pgvector similarity search
4. Generation — retrieved context is injected into the LLM prompt
5. Streaming response — answer streams back via SSE with citations

Knowledge Base

Swan’s knowledge base includes:

• Health Canada dietary guidelines
• USDA food composition data
• WHO nutrition recommendations

PDFs are ingested, chunked, embedded (OpenAI text-embedding-3-small, 1536 dimensions), and stored in PostgreSQL with pgvector.

Key Design Decisions

• Embeddings for finding, not storing — we use embeddings to find relevant content, not to encode facts
• Temporal queries bypass embeddings — date-based questions go straight to SQL
• Prompt caching — system prompts are cached in Valkey to reduce token costs
• Citation tracking — every RAG response includes source citations