Embeddings

better-supabase/ai-sdk/embeddings connects the knowledge and memory blocks to the AI SDK. The blocks take any Embedder; this subpath builds one from an AI SDK model and adds the tool the model searches with.

pnpm add ai

EmbeddersCopy Anchor Link

import { embedWith, supabaseEmbed } from "better-supabase/ai-sdk/embeddings";

const embedder = embedWith("openai/text-embedding-3-small");

embedWith(model, options) calls the SDK's embedMany with maxParallelCalls (4 by default) and providerOptions. A model string goes through the AI Gateway. The embedder's model name (openai/text-embedding-3-small, or name when you pass one) is stored with each chunk, so knowledge.reembed can find what another model embedded.

supabaseEmbed() runs Supabase's built-in gte-small model in an Edge Function, through Supabase.ai.Session, with no API key. It returns 384 numbers, so install the modules with options: { dimensions: 384 }.

Search toolCopy Anchor Link

import { searchTool, toSourceParts } from "better-supabase/ai-sdk/embeddings";

const result = streamText({
  model,
  messages,
  tools: {
    searchKnowledge: searchTool(knowledge, organizationId, {
      scopes: [{ scope: "organization" }, { scope: "chat", id: chatId }],
      onHits: (hits) => sources.push(...toSourceParts(hits)),
    }),
  },
});

searchTool gives the model a query input and returns up to k (8) chunks as { results: [{ id, title, content }] }, with ids in the form documentId#index the model can cite. The search runs as the caller, so row level security decides what the model sees. toSourceParts(hits) returns one source-document part per document, for the assistant message.

RerankingCopy Anchor Link

import { rerankWith } from "better-supabase/ai-sdk/embeddings";

searchTool(knowledge, organizationId, {
  k: 20,
  rerank: rerankWith("cohere/rerank-v3.5", { topN: 5 }),
});

rerankWith reorders the hits with the SDK's rerank and keeps the topN best, with the reranker's score.