Docufin SDK

Docufin SDK is the workflow surface for building document pipelines.

Install

bun add @docufin/sdk

Quick start

import { pipeline } from "@docufin/sdk/workflow";

export const OcrExtractReview = pipeline("ocr-extract-review", {
  concurrency: 2,
})
  .each(async (doc, { steps }) => {
    const text = await steps.ocr(doc);
    const schema = await steps.readSchema("invoice_schema");
    return await steps.extract(text, { schema });
  })
  .review({
    title: "Approve extraction",
    timeout: "72h",
  });

Concepts

• Doc: A document reference (uri, optional mimeType, filename, meta).
• Item<T>: A value paired with its source doc ({ doc, value }). .review() runs on items.
• Streams: .each() starts with docs. Returning non-doc values switches the stream to items.

Pipeline DSL

• .each(fn): Runs per doc. Return a Doc to keep a doc stream, or any other value to produce items.
• .all(fn): Runs once on the full set. Return a Doc/Doc[] to keep docs, or a value to finish with a scalar.
• .review(fn | config): Runs on items only. Pass a handler or a config with title/timeout.

Concurrency is applied only to .each() steps.

Steps API

Use steps inside handlers to access runtime integrations:

• OCR and ML: steps.ocr, steps.classify, steps.extract, steps.llm
• Schemas and prompts: steps.readSchema, steps.readPrompt
• Review: steps.review
• Artifacts: steps.readArtifact, steps.writeArtifact, steps.readText, steps.writeText, steps.readJson, steps.writeJson
• Docs: steps.mergePdf, steps.writeTextAsDoc

Example artifact usage:

import { pipeline } from "@docufin/sdk/workflow";

export const ArtifactExample = pipeline("artifact-example").each(
  async (doc, { steps }) => {
    const text = await steps.readText(doc);

    const { uri } = await steps.writeText({
      text,
      filename: "raw.txt",
    });

    return await steps.readJson(uri);
  }
);

Params schema

Pass a schema with parse or safeParse to validate job params.

import { z } from "zod";
import { pipeline } from "@docufin/sdk/workflow";

const ParamsSchema = z.object({
  topic: z.string().optional(),
});

export const Haiku = pipeline("haiku", {
  schema: ParamsSchema,
}).all(async (_docs, { params, steps }) => {
  const haiku = await steps.llm(`Write a haiku about ${params.topic}.`);

  return await steps.writeTextAsDoc(`${haiku.trimEnd()}\n`, {
    filename: "haiku.txt",
    mimeType: "text/plain",
  });
});

Runtime rules

• The runtime is the only source of durable tasks and waits.
• Do not wrap runtime primitives in Promise.all; use .each() / .all() fan-out instead.
• The pipeline runtime hides task handles and queue details from pipeline authors.

More examples

See .memory/sdk/cookbook.md for more recipes.