RAG Pattern

Basic RAG

import { RAG, mockRetriever } from 'agentfootprint';

const retriever = mockRetriever([{
  chunks: [
    { content: 'Return policy: 14 days for full refund.', metadata: { source: 'policy' } },
    { content: 'AppleCare+: 3 years, $199/year.', metadata: { source: 'applecare' } },
  ],
}]);

const rag = RAG.create({ provider, retriever })
  .system('Answer based on retrieved documents only. Cite your source.')
  .topK(5)
  .build();

const result = await rag.run('What is the return policy?');

How it works

Query → Retrieve (topK chunks) → Inject into context → CallLLM → Response

The RAG pattern is a single-pass pipeline: retrieve relevant chunks, inject them into the LLM context, and generate an answer. No tool loop — one LLM call per query.

Custom retrievers

Implement the Retriever interface to connect any vector database:

const retriever = {
  retrieve: async (query, options) => {
    const results = await vectorDB.search(query, { limit: options?.topK ?? 5 });
    return {
      chunks: results.map(r => ({
        content: r.text,
        metadata: { source: r.source, score: r.score },
      })),
      query,
    };
  },
};

Observability

const result = await rag.run('What is the return policy?');

// Narrative shows: retrieved chunks, injected context, LLM response
rag.getNarrative();