fix(edge): meter and size-cap knowledge embeddings in embed-chunks and search-knowledge

server/supabase/functions/embed-chunks/handler.test.ts

@@ -0,0 +1,166 @@
+// Regression tests for the embed-chunks cost guard (redteam r1 #18): before
+// the guard any signed-in account could have an unbounded number of chunks of
+// any size embedded by the paid provider, with no tier budget.
+
+import { createEmbedChunksHandler, type KnowledgeChunkRow, type KnowledgeIndexStore } from './handler.ts'
+import { EMBEDDING_DIMENSIONS, type EmbeddingProvider } from '../_shared/openai-embeddings.ts'
+import { EMBEDDING_LIMITS, EMBEDDING_QUOTA } from '../_shared/embedding-quota.ts'
+import { MemoryUsageStore } from '../_shared/embedding-quota.fake.ts'
+import type { PlanQuotaTier } from '../_shared/core-contract.generated.ts'
+
+function assert(condition: boolean, message: string): asserts condition {
+  if (!condition) throw new Error(message)
+}
+
+const USER = 'user-1'
+const DOC = '0b6c1f3e-4a5d-4e7f-8a9b-1c2d3e4f5a6b'
+const NOW = new Date('2026-09-28T12:00:00Z')
+const TODAY = '2026-09-28'
+
+class FakeKnowledgeStore implements KnowledgeIndexStore {
+  embedded = new Set<string>()
+  indexed = false
+  listLimit = -1
+  constructor(public chunks: KnowledgeChunkRow[], public owner: string | null = USER) {}
+  documentOwner(): Promise<string | null> {
+    return Promise.resolve(this.owner)
+  }
+  countChunks(_doc: string, pendingOnly: boolean): Promise<number> {
+    return Promise.resolve(pendingOnly ? this.pending().length : this.chunks.length)
+  }
+  listPendingChunks(_doc: string, limit: number): Promise<KnowledgeChunkRow[]> {
+    this.listLimit = limit
+    return Promise.resolve(this.pending().slice(0, limit))
+  }
+  saveEmbedding(_doc: string, chunkId: string): Promise<boolean> {
+    this.embedded.add(chunkId)
+    return Promise.resolve(true)
+  }
+  clearIndexed(): Promise<boolean> {
+    this.indexed = false
+    return Promise.resolve(true)
+  }
+  markIndexed(): Promise<boolean> {
+    this.indexed = true
+    return Promise.resolve(true)
+  }
+  private pending(): KnowledgeChunkRow[] {
+    return this.chunks.filter((chunk) => !this.embedded.has(chunk.id))
+  }
+}
+
+class FakeProvider implements EmbeddingProvider {
+  calls = 0
+  inputs = 0
+  constructor(private fail = false) {}
+  embed(input: string | readonly string[]) {
+    this.calls += 1
+    const list = typeof input === 'string' ? [input] : input
+    this.inputs += list.length
+    if (this.fail) return Promise.resolve({ ok: false as const, reason: 'upstream' as const })
+    return Promise.resolve({
+      ok: true as const,
+      data: list.map((_, index) => ({ index, embedding: new Array(EMBEDDING_DIMENSIONS).fill(0.1) })),
+    })
+  }
+}
+
+function chunks(count: number, size = 800): KnowledgeChunkRow[] {
+  return Array.from({ length: count }, (_, i) => ({ id: `c${i}`, content: 'x'.repeat(size) }))
+}
+
+function setup(options: {
+  tier?: PlanQuotaTier
+  chunks: KnowledgeChunkRow[]
+  providerFails?: boolean
+  usageFails?: boolean
+}) {
+  const store = new FakeKnowledgeStore(options.chunks)
+  const provider = new FakeProvider(options.providerFails)
+  const usage = new MemoryUsageStore(options.tier ?? 'free', TODAY)
+  usage.failIncrement = options.usageFails ?? false
+  const handler = createEmbedChunksHandler({
+    authenticate: () => Promise.resolve({ id: USER }),
+    embeddingProvider: () => provider,
+    knowledgeStore: () => store,
+    usageStore: () => usage,
+    now: () => NOW,
+  })
+  const call = () => handler(new Request('https://edge/embed-chunks', {
+    method: 'POST',
+    body: JSON.stringify({ document_id: DOC }),
+  }))
+  return { store, provider, usage, call }
+}
+
+Deno.test('a free account over its weekly embedding budget is refused before any provider call', async () => {
+  const { provider, usage, call } = setup({ chunks: chunks(10) })
+  usage.rows.set('2026-09-26', EMBEDDING_QUOTA.free.limit - 5)
+  const response = await call()
+  assert(response.status === 429, `expected 429, got ${response.status}`)
+  const body = await response.json()
+  assert(body.error === 'quota_exceeded' && body.tier === 'free', 'quota body')
+  assert(provider.calls === 0, 'provider was called despite the exhausted budget')
+  assert((usage.rows.get(TODAY) ?? 0) === 0, 'refused request left units counted')
+})
+
+Deno.test('a document over the chunk cap is refused without spending', async () => {
+  const { provider, usage, call } = setup({ tier: 'enterprise', chunks: chunks(EMBEDDING_LIMITS.maxChunksPerDocument + 1, 10) })
+  const response = await call()
+  assert(response.status === 413, `expected 413, got ${response.status}`)
+  assert((await response.json()).error === 'knowledge_document_too_large', 'error code')
+  assert(provider.calls === 0 && !usage.rows.has(TODAY), 'spent on an oversized document')
+})
+
+Deno.test('a chunk over the size cap is refused without spending', async () => {
+  const list = chunks(3)
+  list[1] = { id: 'huge', content: 'x'.repeat(EMBEDDING_LIMITS.maxChunkChars + 1) }
+  const { provider, call } = setup({ tier: 'enterprise', chunks: list })
+  const response = await call()
+  assert(response.status === 413, `expected 413, got ${response.status}`)
+  const body = await response.json()
+  assert(body.error === 'knowledge_chunk_too_large' && body.chunk_id === 'huge', 'error body')
+  assert(provider.calls === 0, 'provider was called for an oversized chunk')
+})
+
+Deno.test('pending chunks are loaded with a bound', async () => {
+  const { store, call } = setup({ tier: 'pro', chunks: chunks(3) })
+  await call()
+  assert(store.listLimit === EMBEDDING_LIMITS.maxChunksPerDocument, `unbounded pending select (${store.listLimit})`)
+})
+
+Deno.test('a successful run indexes the document and counts the embedded volume', async () => {
+  const { store, provider, usage, call } = setup({ tier: 'pro', chunks: chunks(150) })
+  const response = await call()
+  assert(response.status === 200, `expected 200, got ${response.status}`)
+  const body = await response.json()
+  assert(body.embedded === 150 && body.total === 150 && body.indexed === true, 'response body')
+  assert(store.indexed, 'document not marked indexed')
+  assert(provider.calls === 2 && provider.inputs === 150, 'batching changed')
+  assert(usage.rows.get(TODAY) === 150, `usage ${usage.rows.get(TODAY)}`)
+})
+
+Deno.test('units of batches the provider rejected are refunded', async () => {
+  const { store, usage, call } = setup({ tier: 'pro', chunks: chunks(5), providerFails: true })
+  const response = await call()
+  assert(response.status === 502, `expected 502, got ${response.status}`)
+  const body = await response.json()
+  assert(body.error === 'embedding_failed' && body.remaining === 5, 'error body')
+  assert(!store.indexed, 'failed document marked indexed')
+  assert(usage.rows.get(TODAY) === 0, 'failed batch was charged')
+})
+
+Deno.test('the quota store being down fails closed', async () => {
+  const { provider, call } = setup({ tier: 'pro', chunks: chunks(2), usageFails: true })
+  const response = await call()
+  assert(response.status === 503, `expected 503, got ${response.status}`)
+  assert((await response.json()).error === 'quota_unavailable', 'error code')
+  assert(provider.calls === 0, 'provider called without a reservation')
+})
+
+Deno.test('another user\'s document stays hidden', async () => {
+  const { store, provider, call } = setup({ tier: 'pro', chunks: chunks(2) })
+  store.owner = 'someone-else'
+  const response = await call()
+  assert(response.status === 404 && provider.calls === 0, 'foreign document was embedded')
+})

server/supabase/functions/embed-chunks/handler.ts

@@ -0,0 +1,178 @@
+// server/supabase/functions/embed-chunks/handler.ts
+// Use case: embed a document's pending knowledge chunks.
+// IO is injected through ports (auth, knowledge store, embedding provider,
+// usage store) so the size and quota guards are testable without Supabase or
+// OpenAI. index.ts wires the real adapters.
+
+import { handleCorsPreflightRequest } from '../_shared/cors.ts'
+import { errorResponse, jsonResponse as json } from '../_shared/json-response.ts'
+import { type EmbeddingProvider, isEmbedding } from '../_shared/openai-embeddings.ts'
+import {
+  EMBEDDING_LIMITS,
+  type EmbeddingReservation,
+  type EmbeddingUsageStore,
+  embeddingUnits,
+  firstOversizedChunk,
+  planEmbeddingBatches,
+  quotaExceededBody,
+  refundEmbeddingUnits,
+  reserveEmbeddingUnits,
+} from '../_shared/embedding-quota.ts'
+
+const UUID_PATTERN = /^[0-9a-f]{8}-[0-9a-f]{4}-[1-5][0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12}$/i
+
+export interface KnowledgeChunkRow {
+  id: string
+  content: string
+}
+
+/** Thrown by a KnowledgeIndexStore when the database call fails. */
+export class KnowledgeStorageError extends Error {
+  constructor() {
+    super('knowledge_storage_failed')
+  }
+}
+
+/** Port over knowledge_documents / knowledge_chunks. Read methods throw KnowledgeStorageError. */
+export interface KnowledgeIndexStore {
+  /** Owner of the document, or null when it does not exist. */
+  documentOwner(documentId: string): Promise<string | null>
+  /** Exact row count, or null when the database returned no count. */
+  countChunks(documentId: string, pendingOnly: boolean): Promise<number | null>
+  listPendingChunks(documentId: string, limit: number): Promise<KnowledgeChunkRow[]>
+  /** Returns false when the write failed. */
+  saveEmbedding(documentId: string, chunkId: string, embedding: number[]): Promise<boolean>
+  /** Clear the indexed flag. With a userId the update is scoped to that owner. Returns false on failure. */
+  clearIndexed(documentId: string, userId: string | null): Promise<boolean>
+  /** Set the indexed flag. Returns false when the write failed or matched no row. */
+  markIndexed(documentId: string, userId: string): Promise<boolean>
+}
+
+export interface EmbedChunksDeps {
+  authenticate(req: Request): Promise<{ id: string }>
+  /** null when the provider key is not configured. */
+  embeddingProvider(): EmbeddingProvider | null
+  knowledgeStore(): KnowledgeIndexStore
+  usageStore(): EmbeddingUsageStore
+  now?(): Date
+}
+
+async function embedBatch(
+  provider: EmbeddingProvider,
+  store: KnowledgeIndexStore,
+  documentId: string,
+  batch: KnowledgeChunkRow[],
+): Promise<number> {
+  const outcome = await provider.embed(batch.map((chunk) => chunk.content))
+  if (!outcome.ok || outcome.data.length !== batch.length) return 0
+
+  const received = new Set<number>()
+  let saved = 0
+  for (const item of outcome.data) {
+    if (
+      typeof item.index !== 'number'
+      || !Number.isInteger(item.index)
+      || item.index < 0
+      || item.index >= batch.length
+      || received.has(item.index)
+      || !isEmbedding(item.embedding)
+    ) {
+      continue
+    }
+    received.add(item.index)
+    if (await store.saveEmbedding(documentId, batch[item.index].id, item.embedding)) saved += 1
+  }
+  return saved
+}
+
+export function createEmbedChunksHandler(deps: EmbedChunksDeps): (req: Request) => Promise<Response> {
+  const now = () => deps.now?.() ?? new Date()
+
+  return async (req: Request): Promise<Response> => {
+    const preflight = handleCorsPreflightRequest(req)
+    if (preflight) return preflight
+    if (req.method !== 'POST') return json(405, { error: 'method_not_allowed' })
+
+    try {
+      const user = await deps.authenticate(req)
+      const body = await req.json().catch(() => null) as { document_id?: unknown } | null
+      if (!body || typeof body.document_id !== 'string' || !UUID_PATTERN.test(body.document_id)) {
+        return json(400, { error: 'invalid_document_id' })
+      }
+      const documentId = body.document_id
+
+      const provider = deps.embeddingProvider()
+      if (!provider) return json(503, { error: 'embedding_provider_unavailable' })
+
+      const store = deps.knowledgeStore()
+      const owner = await store.documentOwner(documentId)
+      if (owner !== user.id) return json(404, { error: 'knowledge_document_not_found' })
+
+      const total = await store.countChunks(documentId, false) ?? 0
+      if (total === 0) {
+        await store.clearIndexed(documentId, null)
+        return json(409, { error: 'knowledge_document_empty' })
+      }
+      if (total > EMBEDDING_LIMITS.maxChunksPerDocument) {
+        return json(413, {
+          error: 'knowledge_document_too_large',
+          total,
+          max_chunks: EMBEDDING_LIMITS.maxChunksPerDocument,
+        })
+      }
+
+      const chunks = await store.listPendingChunks(documentId, EMBEDDING_LIMITS.maxChunksPerDocument)
+      const oversized = firstOversizedChunk(chunks)
+      if (oversized !== -1) {
+        return json(413, {
+          error: 'knowledge_chunk_too_large',
+          chunk_id: chunks[oversized].id,
+          max_chars: EMBEDDING_LIMITS.maxChunkChars,
+        })
+      }
+
+      // Reserve the whole run before the first paid call; unspent units are refunded.
+      const usage = deps.usageStore()
+      const units = embeddingUnits(chunks.map((chunk) => chunk.content))
+      let reservation: EmbeddingReservation
+      try {
+        reservation = await reserveEmbeddingUnits(usage, user.id, units, now())
+      } catch {
+        return json(503, { error: 'quota_unavailable' })
+      }
+      if (!reservation.allowed) return json(429, quotaExceededBody(reservation, units))
+
+      let embedded = 0
+      let failed = 0
+      let unspentUnits = 0
+      for (const batch of planEmbeddingBatches(chunks)) {
+        let saved = 0
+        try {
+          saved = await embedBatch(provider, store, documentId, batch)
+        } catch {
+          saved = 0
+        }
+        embedded += saved
+        failed += batch.length - saved
+        if (saved === 0) unspentUnits += embeddingUnits(batch.map((chunk) => chunk.content))
+      }
+      await refundEmbeddingUnits(usage, user.id, reservation, unspentUnits, now())
+
+      const remaining = await store.countChunks(documentId, true) ?? total
+      if (failed > 0 || remaining > 0) {
+        if (!await store.clearIndexed(documentId, user.id)) {
+          return json(500, { error: 'knowledge_storage_failed' })
+        }
+        return json(502, { error: 'embedding_failed', embedded, remaining })
+      }
+
+      if (!await store.markIndexed(documentId, user.id)) {
+        return json(500, { error: 'knowledge_storage_failed' })
+      }
+      return json(200, { embedded, total, indexed: true })
+    } catch (error) {
+      if (error instanceof KnowledgeStorageError) return json(500, { error: 'knowledge_storage_failed' })
+      return errorResponse(error)
+    }
+  }
+}

server/supabase/functions/embed-chunks/index.ts

@@ -1,181 +1,26 @@
-import { corsHeaders, handleCorsPreflightRequest } from '../_shared/cors.ts'
-import { requireUser, authErrorResponse, type AuthError } from '../_shared/auth.ts'
+// server/supabase/functions/embed-chunks/index.ts
+// Composition root: wires the real adapters into the embed-chunks use case.
+
+import { requireUser } from '../_shared/auth.ts'
 import { createServiceRoleClient } from '../_shared/quota.ts'
 import { readProviderKey } from '../_shared/provider-key.ts'
+import { createOpenAIEmbeddingProvider } from '../_shared/openai-embeddings.ts'
+import { createSupabaseEmbeddingUsageStore } from '../_shared/embedding-quota.ts'
+import { createEmbedChunksHandler } from './handler.ts'
+import { createSupabaseKnowledgeStore } from './supabase-knowledge-store.ts'
 
-const EMBEDDING_DIMENSIONS = 1536
-const PROVIDER_TIMEOUT_MS = 45_000
-const UUID_PATTERN = /^[0-9a-f]{8}-[0-9a-f]{4}-[1-5][0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12}$/i
-
-interface OpenAIEmbeddingResponse {
-  data?: Array<{ embedding?: unknown; index?: unknown }>
+let serviceClient: ReturnType<typeof createServiceRoleClient> | null = null
+function service(): ReturnType<typeof createServiceRoleClient> {
+  serviceClient ??= createServiceRoleClient()
+  return serviceClient
 }
 
-function json(status: number, body: Record<string, unknown>): Response {
-  return new Response(JSON.stringify(body), {
-    status,
-    headers: { ...corsHeaders, 'Content-Type': 'application/json' },
-  })
-}
-
-function validEmbedding(value: unknown): value is number[] {
-  return Array.isArray(value)
-    && value.length === EMBEDDING_DIMENSIONS
-    && value.every((entry) => typeof entry === 'number' && Number.isFinite(entry))
-}
-
-Deno.serve(async (req: Request) => {
-  const preflight = handleCorsPreflightRequest(req)
-  if (preflight) return preflight
-  if (req.method !== 'POST') return json(405, { error: 'method_not_allowed' })
-
-  try {
-    const user = await requireUser(req)
-    const body = await req.json().catch(() => null) as { document_id?: unknown } | null
-    if (!body || typeof body.document_id !== 'string' || !UUID_PATTERN.test(body.document_id)) {
-      return json(400, { error: 'invalid_document_id' })
-    }
-
-    const openaiKey = readProviderKey('OPENAI_API_KEY')
-    if (!openaiKey) return json(503, { error: 'embedding_provider_unavailable' })
-
-    const serviceClient = createServiceRoleClient()
-    const documentResult = await serviceClient
-      .from('knowledge_documents')
-      .select('id,user_id')
-      .eq('id', body.document_id)
-      .maybeSingle()
-    if (documentResult.error) return json(500, { error: 'knowledge_storage_failed' })
-    if (!documentResult.data || documentResult.data.user_id !== user.id) {
-      return json(404, { error: 'knowledge_document_not_found' })
-    }
-
-    const totalResult = await serviceClient
-      .from('knowledge_chunks')
-      .select('id', { count: 'exact', head: true })
-      .eq('document_id', body.document_id)
-    if (totalResult.error) return json(500, { error: 'knowledge_storage_failed' })
-    const total = totalResult.count ?? 0
-    if (total === 0) {
-      await serviceClient
-        .from('knowledge_documents')
-        .update({ indexed: false, indexed_at: null })
-        .eq('id', body.document_id)
-      return json(409, { error: 'knowledge_document_empty' })
-    }
-
-    const pendingResult = await serviceClient
-      .from('knowledge_chunks')
-      .select('id,content')
-      .eq('document_id', body.document_id)
-      .is('embedding', null)
-      .order('chunk_index', { ascending: true })
-    if (pendingResult.error) return json(500, { error: 'knowledge_storage_failed' })
-
-    const chunks = (pendingResult.data ?? []) as Array<{ id: string; content: string }>
-    let embedded = 0
-    let failed = 0
-
-    for (let offset = 0; offset < chunks.length; offset += 100) {
-      const batch = chunks.slice(offset, offset + 100)
-      try {
-        const response = await fetch('https://api.openai.com/v1/embeddings', {
-          method: 'POST',
-          headers: {
-            Authorization: `Bearer ${openaiKey}`,
-            'Content-Type': 'application/json',
-          },
-          body: JSON.stringify({
-            model: 'text-embedding-3-small',
-            input: batch.map((chunk) => chunk.content),
-            dimensions: EMBEDDING_DIMENSIONS,
-          }),
-          signal: AbortSignal.timeout(PROVIDER_TIMEOUT_MS),
-        })
-        if (!response.ok) {
-          failed += batch.length
-          continue
-        }
-
-        const payload = await response.json().catch(() => null) as OpenAIEmbeddingResponse | null
-        if (!payload || !Array.isArray(payload.data) || payload.data.length !== batch.length) {
-          failed += batch.length
-          continue
-        }
-
-        const received = new Set<number>()
-        let batchSuccess = 0
-        for (const item of payload.data) {
-          if (
-            typeof item.index !== 'number'
-            || !Number.isInteger(item.index)
-            || item.index < 0
-            || item.index >= batch.length
-            || received.has(item.index)
-            || !validEmbedding(item.embedding)
-          ) {
-            continue
-          }
-          received.add(item.index)
-          const update = await serviceClient
-            .from('knowledge_chunks')
-            .update({ embedding: item.embedding })
-            .eq('id', batch[item.index].id)
-            .eq('document_id', body.document_id)
-          if (!update.error) {
-            embedded += 1
-            batchSuccess += 1
-          }
-        }
-        failed += batch.length - batchSuccess
-      } catch {
-        failed += batch.length
-      }
-    }
-
-    const remainingResult = await serviceClient
-      .from('knowledge_chunks')
-      .select('id', { count: 'exact', head: true })
-      .eq('document_id', body.document_id)
-      .is('embedding', null)
-    if (remainingResult.error) return json(500, { error: 'knowledge_storage_failed' })
-
-    const remaining = remainingResult.count ?? total
-    if (failed > 0 || remaining > 0) {
-      const rollback = await serviceClient
-        .from('knowledge_documents')
-        .update({ indexed: false, indexed_at: null })
-        .eq('id', body.document_id)
-        .eq('user_id', user.id)
-      if (rollback.error) return json(500, { error: 'knowledge_storage_failed' })
-      return json(502, {
-        error: 'embedding_failed',
-        embedded,
-        remaining,
-      })
-    }
-
-    const indexed = await serviceClient
-      .from('knowledge_documents')
-      .update({ indexed: true, indexed_at: new Date().toISOString() })
-      .eq('id', body.document_id)
-      .eq('user_id', user.id)
-      .select('id')
-      .maybeSingle()
-    if (indexed.error || !indexed.data) return json(500, { error: 'knowledge_storage_failed' })
-
-    return json(200, { embedded, total, indexed: true })
-  } catch (error) {
-    if (
-      error
-      && typeof error === 'object'
-      && 'status' in error
-      && (error.status === 401 || error.status === 403)
-      && 'message' in error
-      && typeof error.message === 'string'
-    ) {
-      return authErrorResponse(error as AuthError, corsHeaders)
-    }
-    return json(500, { error: 'internal_error' })
-  }
-})
+Deno.serve(createEmbedChunksHandler({
+  authenticate: requireUser,
+  embeddingProvider: () => {
+    const key = readProviderKey('OPENAI_API_KEY')
+    return key ? createOpenAIEmbeddingProvider(key) : null
+  },
+  knowledgeStore: () => createSupabaseKnowledgeStore(service()),
+  usageStore: () => createSupabaseEmbeddingUsageStore(service()),
+}))

server/supabase/functions/embed-chunks/supabase-knowledge-store.ts

@@ -0,0 +1,70 @@
+// server/supabase/functions/embed-chunks/supabase-knowledge-store.ts
+// KnowledgeIndexStore adapter over the service-role Supabase client.
+
+import type { createClient } from '@supabase/supabase-js'
+import { type KnowledgeChunkRow, type KnowledgeIndexStore, KnowledgeStorageError } from './handler.ts'
+
+type SupabaseClient = ReturnType<typeof createClient>
+
+export function createSupabaseKnowledgeStore(client: SupabaseClient): KnowledgeIndexStore {
+  return {
+    async documentOwner(documentId) {
+      const { data, error } = await client
+        .from('knowledge_documents')
+        .select('id,user_id')
+        .eq('id', documentId)
+        .maybeSingle()
+      if (error) throw new KnowledgeStorageError()
+      const row = data as { user_id?: unknown } | null
+      return row && typeof row.user_id === 'string' ? row.user_id : null
+    },
+    async countChunks(documentId, pendingOnly) {
+      let query = client
+        .from('knowledge_chunks')
+        .select('id', { count: 'exact', head: true })
+        .eq('document_id', documentId)
+      if (pendingOnly) query = query.is('embedding', null)
+      const { count, error } = await query
+      if (error) throw new KnowledgeStorageError()
+      return count
+    },
+    async listPendingChunks(documentId, limit) {
+      const { data, error } = await client
+        .from('knowledge_chunks')
+        .select('id,content')
+        .eq('document_id', documentId)
+        .is('embedding', null)
+        .order('chunk_index', { ascending: true })
+        .limit(limit)
+      if (error) throw new KnowledgeStorageError()
+      return (data ?? []) as KnowledgeChunkRow[]
+    },
+    async saveEmbedding(documentId, chunkId, embedding) {
+      const { error } = await client
+        .from('knowledge_chunks')
+        .update({ embedding })
+        .eq('id', chunkId)
+        .eq('document_id', documentId)
+      return !error
+    },
+    async clearIndexed(documentId, userId) {
+      let query = client
+        .from('knowledge_documents')
+        .update({ indexed: false, indexed_at: null })
+        .eq('id', documentId)
+      if (userId !== null) query = query.eq('user_id', userId)
+      const { error } = await query
+      return !error
+    },
+    async markIndexed(documentId, userId) {
+      const { data, error } = await client
+        .from('knowledge_documents')
+        .update({ indexed: true, indexed_at: new Date().toISOString() })
+        .eq('id', documentId)
+        .eq('user_id', userId)
+        .select('id')
+        .maybeSingle()
+      return !error && Boolean(data)
+    },
+  }
+}