ConusAI vendor v0.1.0-dev.20261027+9b1131fb

modules/documents/adapters/ocr.ts

@@ -0,0 +1,55 @@
+// The OCR port over the platform's document reader: the one file of this module that knows the SDK's
+// error type.
+//
+// `DocumentReader` (kit-platform) is `@conusai-sdk/ocr` wired to the deployment contract: `submit`
+// returns when the platform has accepted the document and `wait` polls the job to its document. What
+// this adapter adds is the one distinction core needs and the SDK only encodes in a string: "the
+// platform says this job failed" (permanent, the document is marked failed) against everything else
+// (a network error, a timeout, a 5xx: the job throws and the queue retries it). A page image the platform
+// has none of is not a failure either: the page is read from its text.
+
+import type { DocumentReader } from "@conusai-kit/platform"
+import { OcrError } from "@conusai-sdk/ocr"
+import { type DocumentOcr, OcrJobFailed } from "../core/ports.js"
+
+export interface SdkOcrOptions {
+  /** Poll interval of `wait`, in milliseconds. */
+  readonly intervalMs?: number
+  /** Give up on one job after this long; the document then retries from its stored job id. */
+  readonly timeoutMs?: number
+}
+
+export function sdkDocumentOcr(
+  reader: Pick<DocumentReader, "submit" | "wait" | "page">,
+  options: SdkOcrOptions = {},
+): DocumentOcr {
+  return {
+    async submit(bytes, { filename, idempotencyKey }) {
+      const job = await reader.submit(bytes, { filename, idempotencyKey })
+      return { jobId: job.id }
+    },
+    async wait(jobId, { signal } = {}) {
+      try {
+        return await reader.wait(jobId, {
+          ...(options.intervalMs === undefined ? {} : { intervalMs: options.intervalMs }),
+          ...(options.timeoutMs === undefined ? {} : { timeoutMs: options.timeoutMs }),
+          ...(signal ? { signal } : {}),
+        })
+      } catch (error) {
+        if (error instanceof OcrError && error.type === "job-failed") throw new OcrJobFailed()
+        throw error
+      }
+    },
+    async page(jobId, page, { signal } = {}) {
+      try {
+        const image = await reader.page(jobId, page, signal ? { signal } : {})
+        return { bytes: image.bytes, mediaType: image.contentType }
+      } catch (error) {
+        // The platform keeps a job's input for a day and draws a page on demand: a job past that, or a
+        // page it has none of, is a page read from its text. Anything else is transient and thrown.
+        if (error instanceof OcrError && (error.status === 404 || error.status === 422)) return undefined
+        throw error
+      }
+    },
+  }
+}