Rag

RAG Pipeline

RAG (Retrieval-Augmented Generation) gives an AI model access to a custom knowledge base by retrieving relevant chunks at inference and injecting them into the prompt.

Concepts

Embedding model — a small, specialized model (separate from the chat LLM) that converts text into a list of numbers (a vector) capturing semantic meaning. Similar text produces similar vectors. Called via OpenAI, Gemini, etc. embedding APIs — not the chat model.

Vector DB — a search index, not an answering engine. Stores vectors + original text as metadata. At query time, returns the top-K most semantically similar text chunks. The chat LLM never reads the vector DB directly.

Chat LLM — generates the final answer. It only sees plain text: the system prompt, injected chunks, and the user question. RAG just gives it better context.

Ingest (run once, re-run when data changes)

Step	Action	API token?
1	Read files (CSV, MD) and split into text chunks with metadata	No
2	Call embedding model — convert each chunk to a vector	Yes (embedding API)
3	Upsert vectors + original text into vector DB	No

Query (every user message)

Step	Action	API token?
1	Receive user question	No
2	Call embedding model — convert question to a vector	Yes (embedding API)
3	Search vector DB — return top-K matching text chunks (not answers)	No
4	Inject chunks into system prompt as plain text	No
5	Call chat LLM — read system prompt + question, generate answer	Yes (chat API)
6	Stream response to UI	No

Embedding and chat use separate API calls — often different providers. Re-ingesting is expensive at scale (step 2 x every chunk). Query costs 2 API calls per message minimum.

Key decision: inject RAG context at prompt level, not as a tool call. Keeps retrieval invisible to the model's reasoning, avoids tool overhead.

Chunk type

type Chunk = {
  id: string;       // stable, unique — used as vector DB record ID
  text: string;     // content to embed and return at query time
  metadata: {
    source: string;
    [key: string]: string | string[] | number;
  };
};

Ingest script pattern

for (const file of dataFiles) {
  const chunks = file.endsWith(".csv")
    ? chunkCSV(filePath, source)
    : chunkMarkdown(filePath, source);
  const embeddings = await generateEmbeddings(chunks.map(c => c.text));
  const vectors = chunks.map((chunk, i) => ({
    id: chunk.id,
    values: embeddings[i],
    metadata: chunk.metadata,
  }));
  await upsertVectors(vectors, NAMESPACE); // batch in groups of 100
}

Switchable embedding provider

// EMBEDDING_PROVIDER=gemini (default) | openai
async function generateEmbeddings(texts: string[]): Promise<number[][]> {
  const provider = process.env.EMBEDDING_PROVIDER || "gemini";
  if (provider === "openai") return embedWithOpenAI(texts);
  return embedWithGemini(texts);
}

Inject into system prompt

const ragContext = await queryVectors(embed(userMessage), topK);
const systemPrompt = ragContext.length > 0
  ? `${basePrompt}\n\n## Relevant Information\n${ragContext.map(r => r.metadata.text).join("\n\n")}`
  : basePrompt;

RAG_TOP_K tuning

• RAG_TOP_K=5 — default, pure semantic search
• RAG_TOP_K_FILTERED=10-30 — use higher K when metadata filters active (filtering reduces recall before ranking)

Auto-ingest via GitHub Actions

on:
  push:
    branches: [main]
    paths: ["data/**"]

Triggers ingestion automatically when data files change — no manual pnpm ingest needed on deploy.

References

• food-tour-sg scripts/ingest.ts: https://github.com/phucbm/food-tour-sg/blob/main/scripts/ingest.ts
• food-tour-sg lib/rag/embeddings.ts: https://github.com/phucbm/food-tour-sg/blob/main/lib/rag/embeddings.ts
• food-tour-sg lib/ai/prompts.ts: https://github.com/phucbm/food-tour-sg/blob/main/lib/ai/prompts.ts
• Vector DB implementation: see pinecone skill