Skip to main content

Data Sources and RAG

Data sources let a Custodian Agent answer from the content of your own files. When you attach a file, Custodian Labs extracts its text, splits it into chunks, embeds them, and retrieves the most relevant chunks for each chat message (retrieval-augmented generation).

Supported file types​

PDF, DOCX, XLSX, XLS, and plain-text formats (.txt, .md, .csv, .json, .yaml). Text files must be UTF-8. See Reference → Supported file types.

Attach a file​

From a path (Node.js)​

import { Custodian } from "@custodianlabs/sdk";

const custodian = new Custodian({
  model: "gpt-4o",
  systemPrompt: "Answer using the provided documents.",
}).addDataSourceFile("./docs/product_manual.pdf");

const app = await custodian.deploy();

With the builder:

const app = await new CustodianBuilder()
  .withModel("gpt-4o")
  .withPrompt("Answer using the provided documents.")
  .withDataSourceFile("./docs/product_manual.pdf")
  .deploy();

Call the method more than once to attach multiple files.

From in-memory bytes (browser, edge, serverless)​

Runtimes without a filesystem pass the bytes directly instead of a path:

custodian.addDataSourceFile({
  data: fileBytes,          // Uint8Array | ArrayBuffer | Blob
  filename: "manual.pdf",
  contentType: "application/pdf", // optional
});

This is the FileInput type, shared by every file method in the SDK (including the Guardian Layer client).

Files upload at deploy time​

addDataSourceFile() stages the file locally. The upload happens inside .deploy(), after the assistant is created. See Custodian Agent Class.

Inspect retrieved context​

Each ChatResponse includes retrievedContexts — the chunks used for that answer:

const reply = await app.chat("How do I configure webhooks?");
console.log(reply?.response);
console.log(reply?.retrievedContexts);

Use it to check whether the agent is grounding its answer in the expected source material.

Good source-file practices​

  • Prefer files with extractable text over scanned images.
  • Use clear section headings.
  • Remove unrelated content before uploading.
  • Split large knowledge bases into focused files.

Next steps​