RAG over private wiki with citations

create extension if not exists vector;
create table chunks (
  id bigserial primary key,
  notion_page text not null,
  body text not null,
  embedding vector(1536) not null
);
create index on chunks using ivfflat (embedding vector_cosine_ops) with (lists = 100);

for (const page of await listNotionPages()) {
  const chunks = splitByHeading(page.body, 600);
  const embeddings = await openai.embeddings.create({ model: "text-embedding-3-small", input: chunks });
  await db.query("insert into chunks (notion_page, body, embedding) select $1, unnest($2::text[]), unnest($3::vector[])", [page.id, chunks, embeddings.data.map((d) => d.embedding)]);
}

server.tool("retrieve_chunks", { query: z.string(), k: z.number().default(6) }, async ({ query, k }) => {
  const [e] = (await openai.embeddings.create({ model: "text-embedding-3-small", input: [query] })).data;
  const { rows } = await db.query("select id, body from chunks order by embedding <=> $1::vector limit $2", [e.embedding, k]);
  return { content: [{ type: "text", text: rows.map((r) => `[chunk:${r.id}] ${r.body}`).join("\n\n") }] };
});

You answer ONLY from the chunks retrieved.
Every sentence MUST end with [chunk:N] where N is the source chunk id.
If no chunk supports a claim, say "I don't know."