RAG Knowledge Base Standards

Efficiently connecting LLMs (GPT-4) with private knowledge (Medical Papers, Recipes, Patient History).

1. Vector Database: Supabase pgvector

Since we use Postgres, we use pgvector. No need for specialized DBs.

1. •Enable extension: create extension vector;
2. •Create table:
   sql

   create table documents (
     id bigserial primary key,
     content text,
     metadata jsonb,
     embedding vector(1536) -- OpenAI uses 1536 dims
   );
   

2. Chunking Strategy (Crucial)

• •Do not embed whole books.
• •Sliding Window: Overlap chunks to preserve context.
  • •Recipes: Chunk by "Step" or "Ingredient Group".
  • •Medical Papers: Chunk by "Paragraph" (~500 tokens).

3. Retrieval Strategy

• •Hybrid Search: Combine Vector Search (Semantic) + Keyword Search (BM25).
  • •Why? "Vitamin C" is a keyword. "Good for immunity" is semantic. You need both matches.
• •Reranking: Use a reranker (Cohere/CrossEncoder) to sort the top 10 results from the DB before sending to GPT.

4. Prompt Engineering for RAG

Structure the prompt to force the LLM to use the context.

You are a Clinical Nutrition Assistant. Answer based ONLY on the provided Context.
If the context doesn't have the answer, say "I don't know based on available clinical guidelines".

Context:
{retrieved_chunks}

Question:
{user_query}