Anvia

Appearance

Build a basic RAG pattern ​

Level: Pattern · Estimated time: 30 minutes

Outcome ​

Embed application documents, search them for a question, and give only the retrieved text to an agent. This version is intentionally local and rebuilds its in-memory index on every start.

When to use it ​

Use this to learn or evaluate retrieval on a small public corpus. Use a durable adapter and an ingestion job for production. Add authorization before indexing private or tenant-owned content.

Request flow ​

documents → embedDocuments → InMemoryVectorStore → index(model).search(...) → bounded context → agent.prompt(...).send().

Setup ​

sh
pnpm add @anvia/core @anvia/openai @anvia/transformers

Index the corpus ​

ts
import { embedDocuments } from "@anvia/core/embeddings";
import { InMemoryVectorStore } from "@anvia/core/vector-store";
import { createTransformersEmbeddingModel } from "@anvia/transformers";

type Note = { id: string; title: string; text: string; source: string };
const embeddingModel = await createTransformersEmbeddingModel();
const embedded = await embedDocuments(embeddingModel, notes, {
  id: (note) => note.id,
  content: (note) => `${note.title}\n${note.text}`,
  metadata: (note) => ({ source: note.source }),
});
const index = InMemoryVectorStore.fromDocuments(embedded).index(embeddingModel);

Retrieve, then answer ​

ts
const matches = await index.search({ query: question, topK: 4 });
const context = matches.map((match) => [
  `Source: ${match.metadata?.source ?? match.id}`,
  match.document.text,
].join("\n")).join("\n\n---\n\n");

const response = await agent.prompt([
  "Answer only from the context. If it is insufficient, say so.",
  `Question: ${question}`,
  `Context:\n${context}`,
].join("\n\n")).send();

agent is an ordinary server-side AgentBuilder result. Manual retrieval makes the boundary visible; use dynamicContext when automatic prompt-time retrieval better fits your application.

Run and expected behavior ​

Ask a question represented in the notes, then an unrelated question. The first answer should name an actual source; the second should say the context is insufficient. Ranking and prose vary, so assert source eligibility and grounded behavior rather than exact wording.

Failure cases ​

Empty corpora, embedding dimension changes, duplicate IDs, low-quality chunks, provider failures, and irrelevant high-scoring matches all need explicit handling. Retrieval relevance is not proof that a claim is correct.

Security and ownership ​

The application owns corpus provenance, access policy, deletion, freshness, chunking, and source display. Indexed documents are untrusted prompt content and cannot grant permissions or authorize tools.

Production changes and tests ​

Move ingestion out of startup, persist embeddings, record model/version, add thresholds and evaluation sets, and cap context tokens. Test known queries, insufficient context, stale/deleted documents, malicious instructions, embedding failures, and source citation membership.

Runnable references ​

Extensions ​

Add metadata filters, a durable vector store adapter, reranking, and retrieval evaluations.

Built for Anvia.