fix(invoice-inbox): read whole PDFs (last-page slice + truncation retry) (#2014)
* fix(invoice-inbox): read whole PDFs (last-page slice + truncation retry) PDF extraction read only part of well-structured PDFs, two confirmed mechanisms (21-day prod window: 49 sliced docs, 29 silent empties): - The auto-extract page budget was 3 (Bedrock-latency legacy, issue #553) and the slice kept only the first pages, so multi-page invoices lost the final page where totals, OCR and 'Att betala' sit. The budget is now 8 on pdf-native backends (Claude reads PDFs directly); the slice always keeps the last page. Rasterizing self-host backends keep the old budget of 3. - A max_tokens-truncated model answer was parsed as-is, failed, and became an all-null extraction with no trace. extractFromDocument now reports stop_reason max_tokens / finish_reason length as truncated; the extractor retries once at double AI_EXTRACTION_MAX_TOKENS and logs ai_extraction_truncated either way. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012xosyW53HUa9JoFiDayhSk * fix(invoice-inbox): sweep cutoff covers the slower two-call extraction Skeptic finding on #2014: the crash-recovery sweep flipped 'processing' rows to an empty skeleton after 2 minutes, but a deferred extraction can now legitimately run 3-5 minutes (8 native pages plus one truncation retry at a doubled token cap), so the sweep stole the row and the CAS discarded the worker's real result. Cutoff raised to 10 minutes. Also: pages_partial_note made period-agnostic (old rows were extracted from first-pages-only slices, so naming the last page was retroactively wrong for them), and two stale first-pages-only comments updated. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012xosyW53HUa9JoFiDayhSk * fix(invoice-inbox): keep the first extraction response when the retry throws CodeRabbit finding on #2014: a throttled/failed retry call bubbled to the outer catch before rawText was assigned, discarding a first response whose text may parse fine despite the truncation flag. The retry is now caught locally (logged as ai_extraction_retry_failed) and the first result flows on. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012xosyW53HUa9JoFiDayhSk --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Mattsson
co-authored by
Claude Fable 5
Claude Fable 5
parent
e8aa0670ca
commit
a1cafe495f
@@ -1323,6 +1323,7 @@ One line per decision: `[YYYY-MM-DD] <decision>: <why>`. Appended by agents and
|
||||
[2026-08-28] AR-PDF minus fix uses ASCII hyphen formatting, not font embedding: registering a Unicode TTF for react-pdf would change the whole document's typography and bundle size to fix one glyph; formatPdfKronor keeps built-in Helvetica and sidesteps WinAnsi's missing U+2212.
|
||||
[2026-08-28] Same-bank warning limited to observed one-session banks (SEB only): prod shows Handelsbanken tolerates 4 concurrent sessions, and the generic warning made a user abandon a legitimate renewal. Planned sync-death visibility work was dropped: already shipped via #1271 (health probe), #1727 (stale state), #1969 (cron unstarve).
|
||||
[2026-08-28] Same-bank warning revised to three tiers after skeptic refutation: hard warn SEB, silent/calm only for verified multi-session banks (Handelsbanken, 4 distinct session_ids observed), legacy hedged warning for unknown banks (fail closed), shared-session siblings exempt (fan-out carries them).
|
||||
[2026-08-28] PDF auto-extract page budget: raised to 8 on pdf-native backends and slice now keeps the last page: the 3-page cap was Bedrock-latency legacy (issue #553), but kept a cap at all so 100-page statements do not burn tokens; last page kept because totals/OCR sit there. Truncation retry doubles AI_EXTRACTION_MAX_TOKENS once instead of raising the default: keeps steady-state cost flat.
|
||||
[2026-08-28] Per-run salary agent gap: new gnubok_set_run_salary staged tool + v1 PATCH instead of extending gnubok_update_payslip_line: the base-salary value lives on salary_run_employees (draft-gated), not on the payslip line (review-agnostic display copy); overloading the line tool would hide the status semantics and keep the recalc-overwrite trap.
|
||||
[2026-08-28] Inbox underlag divergence (#1548) is a separate underlag_status field, not a nulled matched_transaction_journal_entry_id: the book-direct and bulk-book routes 409 on an already-booked transaction, so hiding the verifikat would make the rail re-offer a booking that always fails; the UI keeps divergent items in Att göra with an explanation and a link instead. Anchored-elsewhere conflicts are counted and logged by the daily reconcile cron, never auto-resolved (moving a document between verifikat is a human decision; never-steal is the 2026-08-13 invariant), and not escalated to processing_history from the inline booking path (no dedupe key; it would fire on every booking). The reconcile's 'InboxUnderlagReconciled' event needed a processing_event_types row (FK): the script's old 'InboxUnderlagBackfilled' type was never registered, so its appends had always failed silently.
|
||||
[2026-08-28] Inbox underlag reconcile (#1548) bounds link work, not the read: the candidate set (matched, unconsumed) holds permanent residents (samlingsverifikat siblings, anchored-elsewhere items) that never leave it, so a read cap ordered by uuid would revisit the same window nightly and starve the tail. The full scan is four columns per row; maxItems now caps unlinked items linked per run, the rest are counted as deferred; transactions whose items already read anchored are still propagated outside that budget, because only the propagation anchors the transaction's pinned document and stamps settled items out of the scan (idempotent, self-shrinking). A verifikat in a locked/closed period is its own status (unlinked_locked): the period-lock trigger rejects the link every time, so it is neither retried nor promised to the user as automatic. An unreadable document row is reported as 'unknown' on the wire and kept out of the booked bucket, matching the helper's absence-is-never-anchored contract.
|
||||
|
||||
@@ -273,6 +273,80 @@ describe('extractInvoiceFields', () => {
|
||||
expect(extractJsonObject(oversized)).toBe(oversized)
|
||||
})
|
||||
|
||||
// ── max_tokens truncation retry (2026-08) ──────────────────
|
||||
// Line-item-heavy documents can blow the output cap; the truncated JSON
|
||||
// used to parse to nothing and look like an unreadable document.
|
||||
|
||||
it('retries once with a doubled cap when the output was truncated at max_tokens', async () => {
|
||||
mockCreate
|
||||
.mockResolvedValueOnce({
|
||||
content: [{ type: 'text', text: JSON.stringify(VALID_RESULT).slice(0, 40) }],
|
||||
stop_reason: 'max_tokens',
|
||||
})
|
||||
.mockReturnValueOnce(aiResponse(VALID_RESULT))
|
||||
const { data } = await extractInvoiceFields({
|
||||
buffer: Buffer.from('%PDF'),
|
||||
mimeType: 'application/pdf',
|
||||
fileName: 'many-line-items.pdf',
|
||||
})
|
||||
expect(mockCreate).toHaveBeenCalledTimes(2)
|
||||
const firstMax = mockCreate.mock.calls[0][0].max_tokens
|
||||
expect(mockCreate.mock.calls[1][0].max_tokens).toBe(firstMax * 2)
|
||||
expect(data.totals.total).toBe(6.25)
|
||||
expect(data.confidence).toBe(1)
|
||||
})
|
||||
|
||||
it('falls back to the empty result when the retry is truncated too', async () => {
|
||||
mockCreate
|
||||
.mockResolvedValueOnce({
|
||||
content: [{ type: 'text', text: '{"lineItems":[{"desc' }],
|
||||
stop_reason: 'max_tokens',
|
||||
})
|
||||
.mockResolvedValueOnce({
|
||||
content: [{ type: 'text', text: '{"lineItems":[{"description":"still cut' }],
|
||||
stop_reason: 'max_tokens',
|
||||
})
|
||||
const { data, rawText } = await extractInvoiceFields({
|
||||
buffer: Buffer.from('%PDF'),
|
||||
mimeType: 'application/pdf',
|
||||
fileName: 'many-line-items.pdf',
|
||||
})
|
||||
expect(mockCreate).toHaveBeenCalledTimes(2)
|
||||
expect(rawText).toBe('{"lineItems":[{"description":"still cut')
|
||||
expect(data.totals.total).toBeNull()
|
||||
expect(data.confidence).toBe(0)
|
||||
})
|
||||
|
||||
it('keeps the first response when the retry call itself throws', async () => {
|
||||
// The truncated first answer happens to be complete valid JSON (the flag
|
||||
// can fire on the last token); a throttled retry must not discard it.
|
||||
mockCreate
|
||||
.mockResolvedValueOnce({
|
||||
content: [{ type: 'text', text: JSON.stringify(VALID_RESULT) }],
|
||||
stop_reason: 'max_tokens',
|
||||
})
|
||||
.mockRejectedValueOnce(new Error('throttled'))
|
||||
const { data, rawText } = await extractInvoiceFields({
|
||||
buffer: Buffer.from('%PDF'),
|
||||
mimeType: 'application/pdf',
|
||||
fileName: 'many-line-items.pdf',
|
||||
})
|
||||
expect(mockCreate).toHaveBeenCalledTimes(2)
|
||||
expect(rawText).toBe(JSON.stringify(VALID_RESULT))
|
||||
expect(data.totals.total).toBe(6.25)
|
||||
expect(data.confidence).toBe(1)
|
||||
})
|
||||
|
||||
it('does not retry when the answer completed under the cap', async () => {
|
||||
mockCreate.mockReturnValueOnce(aiResponse(VALID_RESULT))
|
||||
await extractInvoiceFields({
|
||||
buffer: Buffer.from('%PDF'),
|
||||
mimeType: 'application/pdf',
|
||||
fileName: 'kvitto.pdf',
|
||||
})
|
||||
expect(mockCreate).toHaveBeenCalledTimes(1)
|
||||
})
|
||||
|
||||
it('returns empty result when AI response fails schema validation', async () => {
|
||||
mockCreate.mockReturnValueOnce(
|
||||
aiResponse({ supplier: { name: 'X' } /* missing required keys */ })
|
||||
|
||||
@@ -0,0 +1,98 @@
|
||||
import { describe, it, expect, beforeEach, afterEach } from 'vitest'
|
||||
import { PDFDocument } from 'pdf-lib'
|
||||
import {
|
||||
slicePdfForExtraction,
|
||||
maxPagesForAutoExtract,
|
||||
MAX_PAGES_FOR_AUTO_EXTRACT,
|
||||
MAX_PAGES_FOR_AUTO_EXTRACT_NATIVE,
|
||||
} from '@/extensions/general/invoice-inbox/lib/upload-and-extract'
|
||||
|
||||
// Pages get index-encoded widths (500 + i) so tests can assert exactly WHICH
|
||||
// source pages survived the slice, not just how many.
|
||||
async function makePdf(pageCount: number): Promise<ArrayBuffer> {
|
||||
const pdf = await PDFDocument.create()
|
||||
for (let i = 0; i < pageCount; i++) pdf.addPage([500 + i, 700])
|
||||
const bytes = await pdf.save()
|
||||
const out = new ArrayBuffer(bytes.byteLength)
|
||||
new Uint8Array(out).set(bytes)
|
||||
return out
|
||||
}
|
||||
|
||||
async function pageWidths(buffer: ArrayBuffer): Promise<number[]> {
|
||||
const pdf = await PDFDocument.load(buffer)
|
||||
return Array.from({ length: pdf.getPageCount() }, (_, i) =>
|
||||
Math.round(pdf.getPage(i).getSize().width)
|
||||
)
|
||||
}
|
||||
|
||||
describe('slicePdfForExtraction', () => {
|
||||
it('keeps the first maxPages-1 pages plus the LAST page (where totals sit)', async () => {
|
||||
const sliced = await slicePdfForExtraction(await makePdf(10), 8)
|
||||
expect(sliced).not.toBeNull()
|
||||
// Pages 0..6 plus page 9: widths 500..506 and 509.
|
||||
expect(await pageWidths(sliced!)).toEqual([500, 501, 502, 503, 504, 505, 506, 509])
|
||||
})
|
||||
|
||||
it('keeps the last page also on the old 3-page budget', async () => {
|
||||
const sliced = await slicePdfForExtraction(await makePdf(5), 3)
|
||||
expect(await pageWidths(sliced!)).toEqual([500, 501, 504])
|
||||
})
|
||||
|
||||
it('copies the document unchanged when it fits the budget', async () => {
|
||||
const sliced = await slicePdfForExtraction(await makePdf(3), 8)
|
||||
expect(await pageWidths(sliced!)).toEqual([500, 501, 502])
|
||||
})
|
||||
|
||||
it('returns null on an unparseable buffer', async () => {
|
||||
const garbage = new TextEncoder().encode('not a pdf').buffer as ArrayBuffer
|
||||
expect(await slicePdfForExtraction(garbage, 8)).toBeNull()
|
||||
})
|
||||
})
|
||||
|
||||
describe('maxPagesForAutoExtract', () => {
|
||||
const ENV = [
|
||||
'AWS_ACCESS_KEY_ID',
|
||||
'AWS_SECRET_ACCESS_KEY',
|
||||
'ANTHROPIC_API_KEY',
|
||||
'AI_PROVIDER',
|
||||
'AI_BASE_URL',
|
||||
'AI_API_KEY',
|
||||
'AI_MODEL',
|
||||
'AI_PDF_MODE',
|
||||
] as const
|
||||
let saved: Record<string, string | undefined> = {}
|
||||
beforeEach(() => {
|
||||
saved = {}
|
||||
for (const k of ENV) {
|
||||
saved[k] = process.env[k]
|
||||
delete process.env[k]
|
||||
}
|
||||
})
|
||||
afterEach(() => {
|
||||
for (const k of ENV) {
|
||||
if (saved[k] === undefined) delete process.env[k]
|
||||
else process.env[k] = saved[k]
|
||||
}
|
||||
})
|
||||
|
||||
it('uses the higher budget when the backend reads PDFs natively (Claude)', () => {
|
||||
process.env.AWS_ACCESS_KEY_ID = 'AKIAEXAMPLE'
|
||||
process.env.AWS_SECRET_ACCESS_KEY = 'secret'
|
||||
expect(maxPagesForAutoExtract()).toBe(MAX_PAGES_FOR_AUTO_EXTRACT_NATIVE)
|
||||
})
|
||||
|
||||
it('keeps the conservative budget on a rasterizing OpenAI-compatible backend', () => {
|
||||
process.env.AI_PROVIDER = 'openai-compatible'
|
||||
process.env.AI_BASE_URL = 'http://localhost:8000/v1'
|
||||
process.env.AI_MODEL = 'some-model'
|
||||
expect(maxPagesForAutoExtract()).toBe(MAX_PAGES_FOR_AUTO_EXTRACT)
|
||||
})
|
||||
|
||||
it('follows AI_PDF_MODE=native on an OpenAI-compatible backend', () => {
|
||||
process.env.AI_PROVIDER = 'openai-compatible'
|
||||
process.env.AI_BASE_URL = 'http://localhost:8000/v1'
|
||||
process.env.AI_MODEL = 'some-model'
|
||||
process.env.AI_PDF_MODE = 'native'
|
||||
expect(maxPagesForAutoExtract()).toBe(MAX_PAGES_FOR_AUTO_EXTRACT_NATIVE)
|
||||
})
|
||||
})
|
||||
@@ -137,9 +137,12 @@ function buildCtx(supabase: unknown): ExtensionContext {
|
||||
} as ExtensionContext
|
||||
}
|
||||
|
||||
// The last page gets a distinct size so tests can assert the slice kept it:
|
||||
// slicePdfForExtraction takes the first pages PLUS the last (totals page).
|
||||
async function makePdfBuffer(pageCount: number): Promise<Uint8Array> {
|
||||
const pdf = await PDFDocument.create()
|
||||
for (let i = 0; i < pageCount; i++) pdf.addPage([612, 792])
|
||||
for (let i = 0; i < pageCount - 1; i++) pdf.addPage([612, 792])
|
||||
pdf.addPage([400, 600])
|
||||
return pdf.save()
|
||||
}
|
||||
|
||||
@@ -165,7 +168,7 @@ beforeEach(() => {
|
||||
})
|
||||
|
||||
describe('POST /upload: staged extraction + page-count gate (issue #553)', () => {
|
||||
it('defers long PDFs: responds processing, then slices to 3 pages and flips to received', async () => {
|
||||
it('defers long PDFs: responds processing, then slices to 8 pages (incl. the last) and flips to received', async () => {
|
||||
const captured: { row?: Record<string, unknown> } = {}
|
||||
const flip: FlipCapture = { filters: [] }
|
||||
const supabase = makeSupabase(captured, flip)
|
||||
@@ -175,7 +178,7 @@ describe('POST /upload: staged extraction + page-count gate (issue #553)', () =>
|
||||
rawText: 'ok',
|
||||
})
|
||||
|
||||
const req = await makeUploadRequest(6)
|
||||
const req = await makeUploadRequest(10)
|
||||
const res = await uploadRoute.handler(req, buildCtx(supabase))
|
||||
const { status, body } = await parseJsonResponse<{ data: Record<string, unknown> }>(res)
|
||||
|
||||
@@ -186,16 +189,20 @@ describe('POST /upload: staged extraction + page-count gate (issue #553)', () =>
|
||||
expect(body.data.extracted_data).toBeNull()
|
||||
expect(body.data.extraction_skipped).toBe(false)
|
||||
expect(body.data.skip_reason).toBeNull()
|
||||
expect(body.data.page_count).toBe(6)
|
||||
expect(body.data.page_count).toBe(10)
|
||||
expect(captured.row?.status).toBe('processing')
|
||||
expect(captured.row?.extracted_data).toBeNull()
|
||||
expect(captured.row?.extraction_skipped).toBe(false)
|
||||
|
||||
// The deferred worker extracts from the sliced copy, not the original.
|
||||
// The deferred worker extracts from the sliced copy, not the original:
|
||||
// the first 7 pages plus the LAST page (the distinct 400x600 one), where
|
||||
// invoice totals usually sit.
|
||||
await vi.waitFor(() => expect(extractInvoiceFields).toHaveBeenCalledOnce())
|
||||
const sentBuffer = vi.mocked(extractInvoiceFields).mock.calls[0][0].buffer
|
||||
const sentPdf = await PDFDocument.load(sentBuffer)
|
||||
expect(sentPdf.getPageCount()).toBe(3)
|
||||
expect(sentPdf.getPageCount()).toBe(8)
|
||||
const lastPage = sentPdf.getPage(7).getSize()
|
||||
expect(lastPage).toEqual({ width: 400, height: 600 })
|
||||
|
||||
// ...and CAS-flips the processing row to received, with the truncation
|
||||
// recorded in extracted_data.pages rather than as a skip.
|
||||
@@ -203,8 +210,8 @@ describe('POST /upload: staged extraction + page-count gate (issue #553)', () =>
|
||||
expect(flip.payload?.status).toBe('received')
|
||||
expect(flip.payload?.extraction_skipped).toBe(false)
|
||||
expect((flip.payload?.extracted_data as { pages?: unknown })?.pages).toEqual({
|
||||
total: 6,
|
||||
analyzed: 3,
|
||||
total: 10,
|
||||
analyzed: 8,
|
||||
})
|
||||
expect(flip.filters).toEqual([
|
||||
['id', 'inbox-1'],
|
||||
@@ -212,7 +219,7 @@ describe('POST /upload: staged extraction + page-count gate (issue #553)', () =>
|
||||
])
|
||||
})
|
||||
|
||||
it('defers PDFs at or below the page-count limit and extracts the full buffer', async () => {
|
||||
it('defers PDFs at or below the native page budget (6 pages) and extracts the full buffer', async () => {
|
||||
const captured: { row?: Record<string, unknown> } = {}
|
||||
const flip: FlipCapture = { filters: [] }
|
||||
const supabase = makeSupabase(captured, flip)
|
||||
@@ -222,7 +229,9 @@ describe('POST /upload: staged extraction + page-count gate (issue #553)', () =>
|
||||
rawText: 'ok',
|
||||
})
|
||||
|
||||
const req = await makeUploadRequest(2)
|
||||
// 6 pages gated on the old budget of 3; on the native budget of 8 the
|
||||
// whole document is read (this is the regression the raise fixes).
|
||||
const req = await makeUploadRequest(6)
|
||||
const res = await uploadRoute.handler(req, buildCtx(supabase))
|
||||
const { status, body } = await parseJsonResponse<{ data: Record<string, unknown> }>(res)
|
||||
|
||||
@@ -230,12 +239,12 @@ describe('POST /upload: staged extraction + page-count gate (issue #553)', () =>
|
||||
expect(body.data.status).toBe('processing')
|
||||
expect(body.data.extraction_skipped).toBe(false)
|
||||
expect(body.data.skip_reason).toBeNull()
|
||||
expect(body.data.page_count).toBe(2)
|
||||
expect(body.data.page_count).toBe(6)
|
||||
|
||||
await vi.waitFor(() => expect(extractInvoiceFields).toHaveBeenCalledOnce())
|
||||
const sentBuffer = vi.mocked(extractInvoiceFields).mock.calls[0][0].buffer
|
||||
const sentPdf = await PDFDocument.load(sentBuffer)
|
||||
expect(sentPdf.getPageCount()).toBe(2)
|
||||
expect(sentPdf.getPageCount()).toBe(6)
|
||||
|
||||
await vi.waitFor(() => expect(flip.payload).toBeDefined())
|
||||
expect(flip.payload?.status).toBe('received')
|
||||
|
||||
@@ -15,7 +15,7 @@ import {
|
||||
countPdfPages,
|
||||
slicePdfForExtraction,
|
||||
MAX_FILE_SIZE,
|
||||
MAX_PAGES_FOR_AUTO_EXTRACT,
|
||||
maxPagesForAutoExtract,
|
||||
UPLOAD_ALLOWED_MIME_TYPES,
|
||||
EMAIL_ALLOWED_MIME_TYPES,
|
||||
ensureHtmlDocument,
|
||||
@@ -758,13 +758,14 @@ export const invoiceInboxExtension: Extension = {
|
||||
})
|
||||
|
||||
// Same page handling as /upload (issue #553): long PDFs extract
|
||||
// from a slice of their first pages; the skip only remains for
|
||||
// unsliceable (encrypted/malformed) PDFs. Sandbox companies skip
|
||||
// Bedrock unconditionally.
|
||||
// from a slice (first pages + the last page); the skip only remains
|
||||
// for unsliceable (encrypted/malformed) PDFs. Sandbox companies
|
||||
// skip Bedrock unconditionally.
|
||||
const maxAutoExtractPages = maxPagesForAutoExtract()
|
||||
const pageCount =
|
||||
file.type === 'application/pdf' ? await countPdfPages(buffer) : null
|
||||
const gatedByPageCount =
|
||||
pageCount != null && pageCount > MAX_PAGES_FOR_AUTO_EXTRACT
|
||||
pageCount != null && pageCount > maxAutoExtractPages
|
||||
const sandbox = await isSandboxCompany(ctx.supabase, ctx.companyId)
|
||||
// Paid-tier gate: no `ai` capability → no Bedrock OCR (seed empty
|
||||
// skeleton; the attached document is still stored). Same paywall as
|
||||
@@ -772,7 +773,7 @@ export const invoiceInboxExtension: Extension = {
|
||||
const hasAiEntitlement = await hasCapability(ctx.supabase, ctx.companyId, CAPABILITY.ai)
|
||||
const slicedBuffer =
|
||||
gatedByPageCount && hasAiEntitlement && !sandbox
|
||||
? await slicePdfForExtraction(buffer, MAX_PAGES_FOR_AUTO_EXTRACT)
|
||||
? await slicePdfForExtraction(buffer, maxAutoExtractPages)
|
||||
: null
|
||||
const skipReason: 'no_ai_entitlement' | 'too_many_pages' | 'sandbox' | null =
|
||||
!hasAiEntitlement
|
||||
@@ -793,7 +794,7 @@ export const invoiceInboxExtension: Extension = {
|
||||
})
|
||||
const { data: extracted } = extraction
|
||||
if (!skipExtraction && slicedBuffer != null && pageCount != null) {
|
||||
extracted.pages = { total: pageCount, analyzed: MAX_PAGES_FOR_AUTO_EXTRACT }
|
||||
extracted.pages = { total: pageCount, analyzed: maxAutoExtractPages }
|
||||
}
|
||||
await mirrorExtractionToDocument(doc.id, {
|
||||
data: extracted,
|
||||
|
||||
@@ -524,13 +524,14 @@ export async function extractInvoiceFields(
|
||||
let rawText: string | null = null
|
||||
let model: string | null = null
|
||||
try {
|
||||
const result = await service.extractFromDocument({
|
||||
const baseMaxTokens = readAiConfig().extractionMaxTokens
|
||||
const request = {
|
||||
document: toDocumentInput(input),
|
||||
system: SYSTEM_PROMPT,
|
||||
instruction: EXTRACTION_INSTRUCTION,
|
||||
maxTokens: readAiConfig().extractionMaxTokens,
|
||||
jsonSchema: EXTRACTION_JSON_SCHEMA,
|
||||
})
|
||||
}
|
||||
let result = await service.extractFromDocument({ ...request, maxTokens: baseMaxTokens })
|
||||
if (!result.ok) {
|
||||
// Not a failure: the deployment cannot read this document at all.
|
||||
// `ai_unconfigured` is the self-host "no key yet" case the 30 s
|
||||
@@ -538,6 +539,43 @@ export async function extractInvoiceFields(
|
||||
log.warn('AI extraction skipped', { file_name_hash: fileNameHash, reason: result.skipped })
|
||||
return { data: emptyResult(), rawText: null, skipped: result.skipped }
|
||||
}
|
||||
if (result.truncated) {
|
||||
// The output hit maxTokens mid-JSON (line-item-heavy documents). Left
|
||||
// alone this parsed to nothing and looked like an unreadable document;
|
||||
// one retry at double the cap recovers it. A second truncation falls
|
||||
// through to the normal parse path, which fails visibly in the log
|
||||
// below instead of silently.
|
||||
log.warn('ai_extraction_truncated', {
|
||||
file_name_hash: fileNameHash,
|
||||
max_tokens: baseMaxTokens,
|
||||
retrying: true,
|
||||
})
|
||||
// A retry that THROWS (throttle, network) must not sink the first
|
||||
// response: its text may still parse despite the truncation flag, and
|
||||
// the outer catch would otherwise return the empty skeleton with
|
||||
// rawText null. Swallow locally and continue with the first result.
|
||||
try {
|
||||
const retry = await service.extractFromDocument({
|
||||
...request,
|
||||
maxTokens: baseMaxTokens * 2,
|
||||
})
|
||||
if (retry.ok) {
|
||||
result = retry
|
||||
if (retry.truncated) {
|
||||
log.warn('ai_extraction_truncated', {
|
||||
file_name_hash: fileNameHash,
|
||||
max_tokens: baseMaxTokens * 2,
|
||||
retrying: false,
|
||||
})
|
||||
}
|
||||
}
|
||||
} catch (retryErr) {
|
||||
log.warn('ai_extraction_retry_failed', {
|
||||
file_name_hash: fileNameHash,
|
||||
error: retryErr instanceof Error ? retryErr.message : String(retryErr),
|
||||
})
|
||||
}
|
||||
}
|
||||
rawText = result.text
|
||||
model = result.model
|
||||
|
||||
|
||||
@@ -21,11 +21,16 @@ import { emptyResult } from './extract-invoice-fields'
|
||||
const log = createLogger('invoice-inbox/sweep')
|
||||
|
||||
/**
|
||||
* A deferred extraction is one Bedrock call (the WhatsApp cron budgets
|
||||
* 10-60s for the same call). Two minutes of silence means no live worker
|
||||
* can still deliver a flip that beats the sweep by enough to matter.
|
||||
* A deferred extraction is now up to TWO Bedrock calls on up to 8 native PDF
|
||||
* pages: the base call can spend its full output budget (~90-150s at 8192
|
||||
* tokens) before the truncation retry doubles the cap and runs again. A live
|
||||
* worker can therefore be legitimately silent for several minutes; a cutoff
|
||||
* shorter than its worst case makes the sweep steal the row and discard the
|
||||
* worker's real result via the status CAS. Ten minutes clears the two-call
|
||||
* worst case with margin while still flipping genuinely crashed rows well
|
||||
* before anyone files a support mail about a stuck item.
|
||||
*/
|
||||
export const PROCESSING_STUCK_MS = 2 * 60 * 1000
|
||||
export const PROCESSING_STUCK_MS = 10 * 60 * 1000
|
||||
const BATCH = 50
|
||||
|
||||
export interface InboxSweepSummary {
|
||||
|
||||
@@ -42,12 +42,24 @@ export function sanitiseMime(raw: string | null | undefined): string {
|
||||
|
||||
export const MAX_FILE_SIZE = 10 * 1024 * 1024
|
||||
|
||||
// AI extraction is tuned for single-page receipts/invoices. Documents above
|
||||
// this page count tend to be sales reports, bank statements, or contracts:
|
||||
// Bedrock churns for minutes and still extracts nothing useful (issue #553).
|
||||
// Above the limit we skip extraction entirely; the document still lands in
|
||||
// the inbox and can be attached to a transaction or converted manually.
|
||||
// Page budget for auto-extraction. Documents far above it tend to be sales
|
||||
// reports, bank statements, or contracts (issue #553); those are sliced (see
|
||||
// slicePdfForExtraction) rather than read whole. The budget depends on how
|
||||
// the backend reads PDFs:
|
||||
// - native (Claude on hosted): the model reads PDF bytes directly and
|
||||
// handles long documents fine; 8 pages covers real multi-page invoices
|
||||
// whose totals sit on a later page without opening the door to 100-page
|
||||
// statements.
|
||||
// - rasterize (OpenAI-compatible self-host): every page becomes a PNG in
|
||||
// the prompt, so the old conservative budget of 3 stays.
|
||||
export const MAX_PAGES_FOR_AUTO_EXTRACT = 3
|
||||
export const MAX_PAGES_FOR_AUTO_EXTRACT_NATIVE = 8
|
||||
|
||||
export function maxPagesForAutoExtract(): number {
|
||||
return getAiStatus().capabilities.pdfNative
|
||||
? MAX_PAGES_FOR_AUTO_EXTRACT_NATIVE
|
||||
: MAX_PAGES_FOR_AUTO_EXTRACT
|
||||
}
|
||||
|
||||
// Returns the page count for a PDF buffer, or null if the buffer isn't a
|
||||
// parseable PDF. Errors fall through so callers can treat "unknown" the same
|
||||
@@ -61,11 +73,13 @@ export async function countPdfPages(buffer: ArrayBuffer): Promise<number | null>
|
||||
}
|
||||
}
|
||||
|
||||
// Long PDFs used to skip extraction entirely (issue #553). Invoice data
|
||||
// almost always sits on the first page(s), so instead we extract from a
|
||||
// slim copy of the first MAX_PAGES_FOR_AUTO_EXTRACT pages and record the
|
||||
// truncation in extracted_data.pages. Returns null when slicing fails
|
||||
// (encrypted/malformed PDF) so the caller can fall back to the old skip.
|
||||
// Long PDFs used to skip extraction entirely (issue #553). Instead we extract
|
||||
// from a slim copy and record the truncation in extracted_data.pages. The
|
||||
// slice is the first maxPages-1 pages PLUS the last page: on multi-page
|
||||
// invoices the totals, OCR number and "Att betala" routinely sit on the final
|
||||
// page, and a first-pages-only slice read everything except the amounts.
|
||||
// Returns null when slicing fails (encrypted/malformed PDF) so the caller can
|
||||
// fall back to the old skip.
|
||||
export async function slicePdfForExtraction(
|
||||
buffer: ArrayBuffer,
|
||||
maxPages: number
|
||||
@@ -73,10 +87,12 @@ export async function slicePdfForExtraction(
|
||||
try {
|
||||
const src = await PDFDocument.load(buffer, { updateMetadata: false })
|
||||
const dst = await PDFDocument.create()
|
||||
const pages = await dst.copyPages(
|
||||
src,
|
||||
Array.from({ length: Math.min(maxPages, src.getPageCount()) }, (_, i) => i)
|
||||
)
|
||||
const total = src.getPageCount()
|
||||
const indices =
|
||||
total > maxPages && maxPages >= 2
|
||||
? [...Array.from({ length: maxPages - 1 }, (_, i) => i), total - 1]
|
||||
: Array.from({ length: Math.min(maxPages, total) }, (_, i) => i)
|
||||
const pages = await dst.copyPages(src, indices)
|
||||
for (const page of pages) dst.addPage(page)
|
||||
const bytes = await dst.save()
|
||||
// Copy into a fresh ArrayBuffer: Uint8Array.buffer is ArrayBufferLike
|
||||
@@ -335,16 +351,18 @@ export async function uploadAndExtract(
|
||||
console.error('[invoice-inbox] Failed to append DocumentIngested:', err)
|
||||
}
|
||||
|
||||
// Page-count gate (issue #553): PDFs above MAX_PAGES_FOR_AUTO_EXTRACT
|
||||
// skip extraction. Bedrock would otherwise block the upload response for
|
||||
// minutes on a 6-page sales report and return nothing useful. Images and
|
||||
// non-PDFs are never gated (single-page by definition). countPdfPages
|
||||
// returns null on malformed PDFs: we treat null as "not gated" and fall
|
||||
// through to the existing extraction path so today's behavior is preserved.
|
||||
// Page-count gate (issue #553): PDFs above the auto-extract page budget
|
||||
// are sliced before extraction. Bedrock would otherwise block the upload
|
||||
// response for minutes on a long sales report and return nothing useful.
|
||||
// Images and non-PDFs are never gated (single-page by definition).
|
||||
// countPdfPages returns null on malformed PDFs: we treat null as "not
|
||||
// gated" and fall through to the existing extraction path so today's
|
||||
// behavior is preserved.
|
||||
const maxAutoExtractPages = maxPagesForAutoExtract()
|
||||
const pageCount =
|
||||
file.type === 'application/pdf' ? await countPdfPages(file.buffer) : null
|
||||
const gatedByPageCount =
|
||||
pageCount != null && pageCount > MAX_PAGES_FOR_AUTO_EXTRACT
|
||||
pageCount != null && pageCount > maxAutoExtractPages
|
||||
const sandbox = await isSandboxCompany(supabase, companyId)
|
||||
// Paid-tier gate: AI document OCR (Bedrock, via extractInvoiceFields) is the
|
||||
// `ai` capability. A company without it (free/manual tier) must never trigger
|
||||
@@ -419,6 +437,7 @@ export async function uploadAndExtract(
|
||||
file,
|
||||
pageCount,
|
||||
gatedByPageCount,
|
||||
maxAutoExtractPages,
|
||||
})
|
||||
|
||||
return {
|
||||
@@ -434,12 +453,12 @@ export async function uploadAndExtract(
|
||||
}
|
||||
}
|
||||
|
||||
// Long PDFs are sliced to their first pages instead of skipped, but only
|
||||
// when extraction would actually run: slicing after an entitlement/sandbox/
|
||||
// opt-out verdict would be wasted CPU.
|
||||
// Long PDFs are sliced (first pages + the last page) instead of skipped,
|
||||
// but only when extraction would actually run: slicing after an
|
||||
// entitlement/sandbox/opt-out verdict would be wasted CPU.
|
||||
const slicedBuffer =
|
||||
gatedByPageCount && syncSkipReason === null
|
||||
? await slicePdfForExtraction(file.buffer, MAX_PAGES_FOR_AUTO_EXTRACT)
|
||||
? await slicePdfForExtraction(file.buffer, maxAutoExtractPages)
|
||||
: null
|
||||
// Skip-reason priority: no-AI-entitlement > sandbox > client opt-out >
|
||||
// page-count. Opt-out outranks the page gate (an opted-out caller never
|
||||
@@ -464,7 +483,7 @@ export async function uploadAndExtract(
|
||||
})
|
||||
const { data: extracted, rawText } = extraction
|
||||
if (!skipExtraction && slicedBuffer != null && pageCount != null) {
|
||||
extracted.pages = { total: pageCount, analyzed: MAX_PAGES_FOR_AUTO_EXTRACT }
|
||||
extracted.pages = { total: pageCount, analyzed: maxAutoExtractPages }
|
||||
}
|
||||
|
||||
// Supplier match by org-nr, then VAT number, then case-insensitive name
|
||||
@@ -563,6 +582,9 @@ interface DeferredExtractionJob {
|
||||
file: { name: string; buffer: ArrayBuffer; type: string }
|
||||
pageCount: number | null
|
||||
gatedByPageCount: boolean
|
||||
/** Snapshot of maxPagesForAutoExtract() from the request, so the gate
|
||||
* verdict and the slice/stamp below cannot disagree. */
|
||||
maxAutoExtractPages: number
|
||||
}
|
||||
|
||||
/**
|
||||
@@ -588,7 +610,7 @@ function scheduleDeferredExtraction(job: DeferredExtractionJob): void {
|
||||
let skipReason: string | null = null
|
||||
try {
|
||||
const slicedBuffer = job.gatedByPageCount
|
||||
? await slicePdfForExtraction(job.file.buffer, MAX_PAGES_FOR_AUTO_EXTRACT)
|
||||
? await slicePdfForExtraction(job.file.buffer, job.maxAutoExtractPages)
|
||||
: null
|
||||
if (job.gatedByPageCount && slicedBuffer == null) {
|
||||
extractionSkipped = true
|
||||
@@ -610,7 +632,7 @@ function scheduleDeferredExtraction(job: DeferredExtractionJob): void {
|
||||
skipReason = result.skipped
|
||||
}
|
||||
if (slicedBuffer != null && job.pageCount != null) {
|
||||
extracted.pages = { total: job.pageCount, analyzed: MAX_PAGES_FOR_AUTO_EXTRACT }
|
||||
extracted.pages = { total: job.pageCount, analyzed: job.maxAutoExtractPages }
|
||||
}
|
||||
}
|
||||
} catch (err) {
|
||||
|
||||
@@ -114,6 +114,24 @@ describe('extractFromDocument request shape (hosted regression net)', () => {
|
||||
model: 'eu.anthropic.claude-sonnet-5',
|
||||
usage: { inputTokens: 10, outputTokens: 5, cacheCreationInputTokens: 0, cacheReadInputTokens: 7 },
|
||||
})
|
||||
expect(result.ok && result.truncated).toBeFalsy()
|
||||
})
|
||||
|
||||
it('flags a max_tokens stop as truncated so the caller can retry with a higher cap', async () => {
|
||||
mockCreate.mockResolvedValueOnce({
|
||||
content: [{ type: 'text', text: '{"lineItems":[{"description":"cut of' }],
|
||||
stop_reason: 'max_tokens',
|
||||
usage: { input_tokens: 10, output_tokens: 100, cache_read_input_tokens: 0, cache_creation_input_tokens: 0 },
|
||||
})
|
||||
const svc = createAnthropicFamilyService(readAiConfig())
|
||||
const result = await svc.extractFromDocument({
|
||||
document: { kind: 'text', text: 'x' },
|
||||
system: SYSTEM,
|
||||
instruction: INSTRUCTION,
|
||||
maxTokens: 100,
|
||||
})
|
||||
expect(result.ok).toBe(true)
|
||||
expect(result.ok && result.truncated).toBe(true)
|
||||
})
|
||||
|
||||
it('honours the extraction tier override (legacy BEDROCK_MODEL_ID)', async () => {
|
||||
|
||||
@@ -272,7 +272,13 @@ export function createAnthropicFamilyService(cfg: ResolvedAiConfig): AiService {
|
||||
messages: [{ role: 'user', content: buildAnthropicDocumentContent(req.document, req.instruction) }],
|
||||
}
|
||||
const resp = await getClient().messages.create(params)
|
||||
return { ok: true, text: textOf(resp), model, usage: usageOf(resp) }
|
||||
return {
|
||||
ok: true,
|
||||
text: textOf(resp),
|
||||
model,
|
||||
usage: usageOf(resp),
|
||||
...(resp.stop_reason === 'max_tokens' ? { truncated: true } : {}),
|
||||
}
|
||||
},
|
||||
}
|
||||
}
|
||||
|
||||
@@ -233,6 +233,7 @@ export function createOpenAICompatibleService(cfg: ResolvedAiConfig): AiService
|
||||
model,
|
||||
usage: usageOf(result),
|
||||
...(built.pagesRasterized ? { pagesRasterized: built.pagesRasterized } : {}),
|
||||
...(result.finishReason === 'length' ? { truncated: true } : {}),
|
||||
}
|
||||
}
|
||||
const result = await generateText({
|
||||
@@ -247,6 +248,7 @@ export function createOpenAICompatibleService(cfg: ResolvedAiConfig): AiService
|
||||
model,
|
||||
usage: usageOf(result),
|
||||
...(built.pagesRasterized ? { pagesRasterized: built.pagesRasterized } : {}),
|
||||
...(result.finishReason === 'length' ? { truncated: true } : {}),
|
||||
}
|
||||
},
|
||||
}
|
||||
|
||||
+14
-1
@@ -149,7 +149,20 @@ export type ExtractionSkipReason =
|
||||
| 'pdf_rasterize_failed'
|
||||
|
||||
export type ExtractFromDocumentResult =
|
||||
| { ok: true; text: string; model: string; usage: AiUsage; pagesRasterized?: number }
|
||||
| {
|
||||
ok: true
|
||||
text: string
|
||||
model: string
|
||||
usage: AiUsage
|
||||
pagesRasterized?: number
|
||||
/**
|
||||
* The model stopped because it hit maxTokens, so `text` is cut mid-answer
|
||||
* (Anthropic stop_reason 'max_tokens', OpenAI finish_reason 'length').
|
||||
* Callers that parse the text should retry with a higher cap instead of
|
||||
* treating the truncated output as a failed parse.
|
||||
*/
|
||||
truncated?: boolean
|
||||
}
|
||||
| { ok: false; skipped: ExtractionSkipReason }
|
||||
|
||||
export interface AiService {
|
||||
|
||||
+1
-1
@@ -3507,7 +3507,7 @@
|
||||
"payment_cash": "Cash",
|
||||
"payment_invoice": "Invoiced",
|
||||
"payment_other": "Other",
|
||||
"pages_partial_note": "Extracted from the first {analyzed} of {total} pages.",
|
||||
"pages_partial_note": "Extracted from {analyzed} of {total} pages.",
|
||||
"heic_hint": "HEIC images cannot be AI-extracted yet. Upload the receipt as JPEG or PDF, or fill in the fields manually.",
|
||||
"skipped_hint": "AI extraction did not run for this document. You can link the document to a transaction or create a supplier invoice manually.",
|
||||
"retry_overwrite_confirm": "Re-running extraction overwrites the fields, including your own edits. Continue?",
|
||||
|
||||
+1
-1
@@ -3507,7 +3507,7 @@
|
||||
"payment_cash": "Kontant",
|
||||
"payment_invoice": "Faktureras",
|
||||
"payment_other": "Annat",
|
||||
"pages_partial_note": "Tolkad från de första {analyzed} av {total} sidorna.",
|
||||
"pages_partial_note": "Tolkad från {analyzed} av {total} sidor.",
|
||||
"heic_hint": "HEIC-bilder kan inte AI-tolkas ännu. Ladda upp kvittot som JPEG eller PDF, eller fyll i fälten manuellt.",
|
||||
"skipped_hint": "AI-tolkning kördes inte för det här dokumentet. Du kan koppla dokumentet till en transaktion eller skapa leverantörsfaktura manuellt.",
|
||||
"retry_overwrite_confirm": "Ny tolkning skriver över fälten, även ändringar du gjort själv. Fortsätta?",
|
||||
|
||||
+2
-1
@@ -4361,7 +4361,8 @@ export interface InvoiceExtractionResult {
|
||||
confidence: number
|
||||
suggestedTemplateId?: string
|
||||
// Set by the caller (not the model) when a long PDF was sliced before
|
||||
// extraction: fields were read from the first `analyzed` of `total` pages.
|
||||
// extraction: fields were read from `analyzed` of `total` pages (the first
|
||||
// pages plus the last, where totals usually sit).
|
||||
pages?: { total: number; analyzed: number }
|
||||
}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user