Merge branch 'fix/bound-sync-lab-reads' into fix/review-meds-docs

This commit is contained in:
Marc-André Bombeck committed 2026-10-25 03:02:57 +01:00
commit 84064ea045
13 files changed
+170 -65

No files matched your search

@@ -63,6 +63,7 @@ import { resolveIndexProvider } from "@/lib/documents/index-document";
import { loadDocumentChatText } from "@/lib/documents/content-index";
import { runInboundExtraction } from "@/lib/documents/extract";
import { reserveBudget } from "@/lib/ai/coach/budget";
import { SYNC_EXTRACT_MAX_READS } from "@/lib/documents/ocr-plan";
const mockModule = vi.mocked(isModuleEnabled);
const mockProvider = vi.mocked(resolveIndexProvider);
@@ -195,8 +196,8 @@ describe("maybeAutoStageLabFacts", () => {
mockText.mockResolvedValue({ text: longReport, source: "verbatim" });
await maybeAutoStageLabFacts("u1", "d1");
const unbounded = mockExtract.mock.calls[0]![0].chunks.length;
expect(unbounded).toBeGreaterThan(1);
const inRequest = mockExtract.mock.calls[0]![0].chunks.length;
expect(inRequest).toBeGreaterThan(1);
const arm = vi.fn(async () => {});
await maybeAutoStageLabFacts("u1", "d1", {
@@ -207,6 +208,24 @@ describe("maybeAutoStageLabFacts", () => {
expect(arm).toHaveBeenCalledWith(2);
});
it("stays inside one request without a run deadline, and says what it left out", async () => {
// The index route without Prefer: respond-async continues into this
// staging with no schedule. Twelve hundred lines is many reads of text.
const longReport = Array.from(
{ length: 1200 },
(_, i) => `Hämoglobin ${i} g/dl Referenzbereich 13-17 LDL 120 mg/dl`,
).join("\n");
mockText.mockResolvedValue({ text: longReport, source: "verbatim" });
await maybeAutoStageLabFacts("u1", "d1");
const reads = mockExtract.mock.calls[0]![0].chunks.length;
expect(reads).toBeGreaterThan(0);
expect(reads).toBeLessThanOrEqual(SYNC_EXTRACT_MAX_READS);
expect(updateMany.mock.calls[0]![0]).toMatchObject({
data: { extractionIncomplete: true },
});
});
it("skips (manual path only) when a module is off", async () => {
mockModule.mockImplementation(async (_u, key) => key !== "labs");
const res = await maybeAutoStageLabFacts("u1", "d1");
@@ -201,16 +201,21 @@ describe("planInboundTextReads", () => {
const line = (i: number) => `Analyt ${i} ${i},0 mg/dl 1 - 9`;
const text = (lines: number) =>
Array.from({ length: lines }, (_, i) => line(i)).join("\n");
const schedule = (calls: number) => ({
maxModelCalls: () => calls,
arm: async () => undefined,
});
const noDeadline = schedule(Number.POSITIVE_INFINITY);
it("reads a short text as one read", () => {
const plan = planInboundTextReads(text(10), undefined);
const plan = planInboundTextReads(text(10), noDeadline);
expect(plan.args.chunks).toHaveLength(1);
expect(plan.partial).toBe(false);
});
it("splits a long text at line boundaries into reads no longer than one read carries", () => {
const whole = text(400);
const plan = planInboundTextReads(whole, undefined);
const plan = planInboundTextReads(whole, noDeadline);
expect(plan.args.chunks.length).toBeGreaterThan(1);
for (const chunk of plan.args.chunks) {
expect(chunk.text!.length).toBeLessThanOrEqual(INBOUND_TEXT_READ_CHARS);
@@ -221,10 +226,6 @@ describe("planInboundTextReads", () => {
});
it("grows the parts when the deadline allows few reads, and reads from the start past twice the size", () => {
const schedule = (calls: number) => ({
maxModelCalls: () => calls,
arm: async () => undefined,
});
// Two reads' worth of calls: the parts grow to hold the text.
const grown = planInboundTextReads(text(200), schedule(4));
expect(grown.args.chunks).toHaveLength(2);
@@ -239,7 +240,7 @@ describe("planInboundTextReads", () => {
});
it("caps the reads of a run with no deadline", () => {
const plan = planInboundTextReads(text(4000), undefined);
const plan = planInboundTextReads(text(4000), noDeadline);
expect(plan.args.chunks.length).toBe(OCR_MAX_PAGES);
expect(plan.partial).toBe(true);
});
+1 -18
View File
@@ -36,6 +36,7 @@ import {
} from "@/lib/documents/extract";
import {
OCR_EXTRACT_MODEL_CALLS,
SYNCHRONOUS_READ_SCHEDULE,
type OcrSchedule,
} from "@/lib/documents/ocr-plan";
import { OcrExtractError } from "@/lib/documents/ocr-read";
@@ -267,24 +268,6 @@ function extractionFailure(err: OcrExtractError): AiRunOutcome<never> {
}
}
/**
* The most reads an extraction makes inside one HTTP request (the extract
* route without `Prefer: respond-async`, which the iPhone app sends). Each
* read is up to `OCR_EXTRACT_MODEL_CALLS` model calls, so this bounds the
* request at six calls where an unplanned read of a twenty-page report made
* forty. A longer document is read in as many reads by sharing its pages or
* text out between them, and what does not fit is reported as
* `extractionIncomplete` on the document; the background run, which has a
* deadline of its own, reads it a page at a time.
*/
export const SYNC_EXTRACT_MAX_READS = 3;
/** The read plan of a request: a fixed number of calls, nothing to arm. */
export const SYNCHRONOUS_READ_SCHEDULE: OcrSchedule = {
maxModelCalls: () => SYNC_EXTRACT_MAX_READS * OCR_EXTRACT_MODEL_CALLS,
arm: async () => undefined,
};
/** Extract and stage facts from one owned document. The caller authorised and charged it. */
export async function executeDocumentExtract(args: {
userId: string;
+2 -1
View File
@@ -59,7 +59,8 @@ export interface IndexOrigin {
worker: boolean;
/**
* A background run's deadline, which the lab staging after the index plans
* its reads against. A request has none: it reads as before.
* its reads against. A request has none: the staging then makes at most
* `SYNC_EXTRACT_MAX_READS` reads and reports what it left out.
*/
schedule?: OcrSchedule;
}
+8 -5
View File
@@ -43,6 +43,7 @@ import {
} from "@/lib/documents/extract";
import {
OCR_EXTRACT_MODEL_CALLS,
SYNCHRONOUS_READ_SCHEDULE,
type OcrSchedule,
} from "@/lib/documents/ocr-plan";
import { OcrExtractError } from "@/lib/documents/ocr-read";
@@ -261,7 +262,10 @@ async function stageFactsIfStored(
* `schedule` is the run's deadline when a job reads (the auto-index job, the
* document run): the report's text is then split into as many reads as fit
* before it, instead of up to `OCR_MAX_PAGES` reads of two calls each that a
* slow model could stretch past the job's expiry (#1274).
* slow model could stretch past the job's expiry (#1274). Absent, the staging
* runs inside a request (the index route without `Prefer: respond-async`)
* and is planned as one ({@link SYNCHRONOUS_READ_SCHEDULE}); what does not
* fit is reported as `extractionIncomplete` on the document.
*/
export async function maybeAutoStageLabFacts(
userId: string,
@@ -333,14 +337,13 @@ export async function maybeAutoStageLabFacts(
{ cap, surface: "job" },
);
const plan = planInboundTextReads(chat.text, options.schedule);
const schedule = options.schedule ?? SYNCHRONOUS_READ_SCHEDULE;
const plan = planInboundTextReads(chat.text, schedule);
let result: InboundExtractionResult;
try {
// A run that cannot fit even one read before its deadline refuses here;
// the reservation is settled below like any other failed read.
await options.schedule?.arm(
plan.args.chunks.length * OCR_EXTRACT_MODEL_CALLS,
);
await schedule.arm(plan.args.chunks.length * OCR_EXTRACT_MODEL_CALLS);
result = await runInboundExtraction({
userId,
provider: provider.pick.entry.instance,
+2 -2
View File
@@ -462,7 +462,7 @@ export interface InboundReads {
/** A document's text in parts at line boundaries, as many as the run's deadline allows. */
export function planInboundTextReads(
text: string,
schedule: OcrSchedule | undefined,
schedule: OcrSchedule,
): InboundReads {
const { chunks, unreadParts } = planTextReads(
text,
@@ -483,7 +483,7 @@ export function planInboundTextReads(
*/
export async function planInboundReads(
source: InboundReadSource,
schedule: OcrSchedule | undefined,
schedule: OcrSchedule,
): Promise<({ ok: true } & InboundReads) | { ok: false }> {
if (source.mode === "text") {
return { ok: true, ...planInboundTextReads(source.text, schedule) };
+22 -4
View File
@@ -46,9 +46,27 @@ export interface OcrSchedule {
arm(modelCalls: number): Promise<void>;
}
/**
* The most reads one HTTP request makes: the extract route, and the lab
* staging the index route continues into, both without `Prefer:
* respond-async` (the iPhone app sends neither). Each read is up to
* `OCR_EXTRACT_MODEL_CALLS` model calls, so this bounds the request at six
* calls where an unplanned read of a twenty-page report made forty. A longer
* document is read in as many reads by sharing its pages or text out between
* them, and what does not fit is reported as `extractionIncomplete` on the
* document; a background run, which has a deadline of its own, reads it a
* page at a time.
*/
export const SYNC_EXTRACT_MAX_READS = 3;
/** The read plan of a request: a fixed number of calls, nothing to arm. */
export const SYNCHRONOUS_READ_SCHEDULE: OcrSchedule = {
maxModelCalls: () => SYNC_EXTRACT_MAX_READS * OCR_EXTRACT_MODEL_CALLS,
arm: async () => undefined,
};
/** How many reads the deadline leaves room for. */
export function maxReadsFor(schedule: OcrSchedule | undefined): number {
if (!schedule) return Number.POSITIVE_INFINITY;
export function maxReadsFor(schedule: OcrSchedule): number {
return Math.max(
1,
Math.floor(schedule.maxModelCalls() / OCR_EXTRACT_MODEL_CALLS),
@@ -111,7 +129,7 @@ export function splitOcrText(
*/
export function planTextReads(
text: string,
schedule: OcrSchedule | undefined,
schedule: OcrSchedule,
target: number,
): { chunks: OcrChunk[]; unreadParts: number } {
const maxReads = Math.min(maxReadsFor(schedule), OCR_MAX_PAGES);
@@ -167,7 +185,7 @@ export async function planVisionReads(
bytes: Buffer,
mime: OcrVisionMime,
pdfSupported: boolean,
schedule: OcrSchedule | undefined,
schedule: OcrSchedule,
): Promise<OcrVisionPlan> {
if (mime !== "application/pdf") {
return {