@remit/mailbox-service 0.0.90 → 0.0.91

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@remit/mailbox-service",
3
- "version": "0.0.90",
3
+ "version": "0.0.91",
4
4
  "type": "module",
5
5
  "main": "src/index.ts",
6
6
  "types": "src/index.ts",
@@ -9,6 +9,10 @@
9
9
  "types": "./src/index.ts",
10
10
  "default": "./src/index.ts"
11
11
  },
12
+ "./filter-config": {
13
+ "types": "./src/filters/filter-config.ts",
14
+ "default": "./src/filters/filter-config.ts"
15
+ },
12
16
  "./body-sync-queue": {
13
17
  "types": "./src/body-sync-queue.ts",
14
18
  "default": "./src/body-sync-queue.ts"
@@ -42,6 +46,7 @@
42
46
  "@remit/data-ports": "*",
43
47
  "@remit/domain-enums": "*",
44
48
  "@remit/mail-oauth-service": "*",
49
+ "@remit/search-service": "*",
45
50
  "@remit/secrets-service": "*",
46
51
  "@remit/sqs-client": "*",
47
52
  "@remit/storage-service": "*",
@@ -0,0 +1,229 @@
1
+ import assert from "node:assert/strict";
2
+ import { describe, it } from "node:test";
3
+ import type {
4
+ FilterItem,
5
+ IFilterAnchorRepository,
6
+ IFilterRepository,
7
+ IMessageLabelRepository,
8
+ } from "@remit/data-ports";
9
+ import { FilterMatchOperator, FilterState } from "@remit/domain-enums";
10
+ import { _resetEmbeddingCapabilityForTest } from "@remit/search-service/capability";
11
+ import type { PlacementMoveService } from "../placement-move.js";
12
+ import { buildFilterConfig, type FilterConfigDeps } from "./filter-config.js";
13
+ import { type FilterMessage, NO_ACTION } from "./match.js";
14
+ import { _resetMessageEmbedderForTest } from "./message-embedder.js";
15
+ import { FilterPipeline, type MessageEmbedder } from "./pipeline.js";
16
+
17
+ const anchorOnlyFilter = (destinationMailboxId: string): FilterItem =>
18
+ ({
19
+ filterId: "flt-semantic",
20
+ accountConfigId: "cfg-1",
21
+ name: "receipts",
22
+ scope: "Standing",
23
+ state: FilterState.Active,
24
+ hasAnchor: true,
25
+ ruleChangedAt: 1,
26
+ matchOperator: FilterMatchOperator.Or,
27
+ literalClauses: [],
28
+ actionLabelId: NO_ACTION,
29
+ actionMailboxId: destinationMailboxId,
30
+ createdAt: 1,
31
+ updatedAt: 1,
32
+ }) as unknown as FilterItem;
33
+
34
+ const CURRENT_MODEL = "test-model@3";
35
+
36
+ const repos = (filter: FilterItem): FilterConfigDeps => ({
37
+ filterService: {
38
+ listByAccountAndState: async () => [filter],
39
+ refreshExpiry: async (f: FilterItem) => f,
40
+ } as unknown as IFilterRepository,
41
+ filterAnchorService: {
42
+ get: async () => ({
43
+ anchorEmbedding: [1, 0, 0],
44
+ anchorEmbeddingId: CURRENT_MODEL,
45
+ }),
46
+ } as unknown as IFilterAnchorRepository,
47
+ messageLabelService: {} as unknown as IMessageLabelRepository,
48
+ placementMoveService: {} as unknown as PlacementMoveService,
49
+ });
50
+
51
+ const message: FilterMessage = {
52
+ from: "billing@stripe.com",
53
+ fromName: "Stripe",
54
+ subject: "Your receipt",
55
+ text: "Thanks for your payment",
56
+ listId: "",
57
+ };
58
+
59
+ describe("buildFilterConfig", () => {
60
+ it("returns undefined without a placement mover — no move path, filters stay off", () => {
61
+ const deps = repos(anchorOnlyFilter("mbx-archive"));
62
+ deps.placementMoveService = undefined;
63
+
64
+ assert.equal(buildFilterConfig(deps), undefined);
65
+ });
66
+
67
+ it("wires an embedder into the config", () => {
68
+ const config = buildFilterConfig(repos(anchorOnlyFilter("mbx-archive")));
69
+
70
+ assert.ok(config);
71
+ assert.ok(config.embedder);
72
+ });
73
+
74
+ it("lights up a semantic anchor-only filter on the body-sync pass", async () => {
75
+ let embedCalls = 0;
76
+ const embedder: MessageEmbedder = {
77
+ embed: async () => {
78
+ embedCalls += 1;
79
+ return [1, 0, 0];
80
+ },
81
+ embeddingId: CURRENT_MODEL,
82
+ };
83
+ const config = buildFilterConfig(
84
+ repos(anchorOnlyFilter("mbx-archive")),
85
+ embedder,
86
+ );
87
+ assert.ok(config);
88
+
89
+ const decision = await new FilterPipeline(config, {
90
+ info: () => {},
91
+ }).evaluate("cfg-1", "m-1", message);
92
+
93
+ assert.deepEqual(decision.move, {
94
+ destinationMailboxId: "mbx-archive",
95
+ filterId: "flt-semantic",
96
+ });
97
+ assert.equal(embedCalls, 1);
98
+ });
99
+
100
+ it("does not match when the message embedding diverges from the anchor", async () => {
101
+ const embedder: MessageEmbedder = {
102
+ embed: async () => [0, 1, 0],
103
+ embeddingId: CURRENT_MODEL,
104
+ };
105
+ const config = buildFilterConfig(
106
+ repos(anchorOnlyFilter("mbx-archive")),
107
+ embedder,
108
+ );
109
+ assert.ok(config);
110
+
111
+ const decision = await new FilterPipeline(config, {
112
+ info: () => {},
113
+ }).evaluate("cfg-1", "m-2", message);
114
+
115
+ assert.equal(decision.move, undefined);
116
+ });
117
+ });
118
+
119
+ // An instance with semantic search off (issue #1068). `off` selects an embedder
120
+ // that throws by design, and the throw lands in the pipeline's per-filter catch:
121
+ // one error-level `filter_anchor_match_failed` for every semantic filter on
122
+ // every message it syncs. The pipeline already has a designed skip for a
123
+ // missing embedder, so the config passes none.
124
+ describe("buildFilterConfig with semantic search off", () => {
125
+ const withProviderOff = async (body: () => Promise<void>): Promise<void> => {
126
+ const previous = process.env.SEARCH_EMBEDDING_PROVIDER;
127
+ process.env.SEARCH_EMBEDDING_PROVIDER = "off";
128
+ _resetMessageEmbedderForTest();
129
+ try {
130
+ await body();
131
+ } finally {
132
+ if (previous === undefined) {
133
+ delete process.env.SEARCH_EMBEDDING_PROVIDER;
134
+ } else {
135
+ process.env.SEARCH_EMBEDDING_PROVIDER = previous;
136
+ }
137
+ _resetMessageEmbedderForTest();
138
+ }
139
+ };
140
+
141
+ it("wires no embedder", async () => {
142
+ await withProviderOff(async () => {
143
+ const config = buildFilterConfig(repos(anchorOnlyFilter("mbx-archive")));
144
+
145
+ assert.ok(config);
146
+ assert.equal(config.embedder, undefined);
147
+ });
148
+ });
149
+
150
+ it("skips the semantic filter without logging an error", async () => {
151
+ await withProviderOff(async () => {
152
+ const config = buildFilterConfig(repos(anchorOnlyFilter("mbx-archive")));
153
+ assert.ok(config);
154
+
155
+ const errors: unknown[] = [];
156
+ const decision = await new FilterPipeline(config, {
157
+ info: () => {},
158
+ debug: () => {},
159
+ warn: () => {},
160
+ error: (fields) => errors.push(fields.alert),
161
+ }).evaluate("cfg-1", "m-off", message);
162
+
163
+ assert.deepEqual(decision.labels, []);
164
+ assert.equal(decision.move, undefined);
165
+ assert.deepEqual(errors, []);
166
+ });
167
+ });
168
+ });
169
+
170
+ // The self-host `local` profile asks for the Transformers.js embedder, but the
171
+ // backend and imap-worker images ship without `@huggingface/transformers`
172
+ // (npm-scripts/docker-bundle.mjs). The first embed throws ERR_MODULE_NOT_FOUND,
173
+ // which is the same capability absence `off` states and must take the same
174
+ // designed skip — not an error-level `filter_anchor_match_failed` per semantic
175
+ // filter per message on a user-facing GET.
176
+ describe("buildFilterConfig when this deployment carries no embedder", () => {
177
+ const moduleNotFound = (): Error =>
178
+ Object.assign(
179
+ new Error("Cannot find package '@huggingface/transformers'"),
180
+ { code: "ERR_MODULE_NOT_FOUND" },
181
+ );
182
+
183
+ it("skips the semantic filter, logs no error, and stops wiring an embedder", async () => {
184
+ const previous = process.env.SEARCH_EMBEDDING_PROVIDER;
185
+ process.env.SEARCH_EMBEDDING_PROVIDER = "local";
186
+ _resetMessageEmbedderForTest();
187
+ _resetEmbeddingCapabilityForTest();
188
+
189
+ try {
190
+ const unavailable: MessageEmbedder = {
191
+ embed: async () => {
192
+ throw moduleNotFound();
193
+ },
194
+ embeddingId: CURRENT_MODEL,
195
+ };
196
+ const config = buildFilterConfig(
197
+ repos(anchorOnlyFilter("mbx-archive")),
198
+ unavailable,
199
+ );
200
+ assert.ok(config);
201
+
202
+ const errors: unknown[] = [];
203
+ const decision = await new FilterPipeline(config, {
204
+ info: () => {},
205
+ debug: () => {},
206
+ warn: () => {},
207
+ error: (fields) => errors.push(fields.alert),
208
+ }).evaluate("cfg-1", "m-local", message);
209
+
210
+ assert.deepEqual(decision.labels, []);
211
+ assert.equal(decision.move, undefined);
212
+ assert.deepEqual(errors, []);
213
+
214
+ // Probed once, not per filter or per message: every later config on this
215
+ // process wires no embedder at all, so nothing else attempts the embed.
216
+ const next = buildFilterConfig(repos(anchorOnlyFilter("mbx-archive")));
217
+ assert.ok(next);
218
+ assert.equal(next.embedder, undefined);
219
+ } finally {
220
+ if (previous === undefined) {
221
+ delete process.env.SEARCH_EMBEDDING_PROVIDER;
222
+ } else {
223
+ process.env.SEARCH_EMBEDDING_PROVIDER = previous;
224
+ }
225
+ _resetMessageEmbedderForTest();
226
+ _resetEmbeddingCapabilityForTest();
227
+ }
228
+ });
229
+ });
@@ -0,0 +1,69 @@
1
+ import type {
2
+ IFilterAnchorRepository,
3
+ IFilterRepository,
4
+ IMessageLabelRepository,
5
+ } from "@remit/data-ports";
6
+ import { isEmbeddingCapabilityUnavailable } from "@remit/search-service/capability";
7
+ import {
8
+ EMBEDDING_PROVIDER_OFF,
9
+ readEmbeddingProviderFromEnv,
10
+ } from "@remit/search-service/from-env";
11
+ import type { PlacementMoveService } from "../placement-move.js";
12
+ import { getMessageEmbedder } from "./message-embedder.js";
13
+ import type { FilterConfig, MessageEmbedder } from "./pipeline.js";
14
+
15
+ export interface FilterConfigDeps {
16
+ filterService: IFilterRepository;
17
+ filterAnchorService: IFilterAnchorRepository;
18
+ messageLabelService: IMessageLabelRepository;
19
+ placementMoveService?: PlacementMoveService;
20
+ }
21
+
22
+ /**
23
+ * The env-selected embedder, or none on an instance that cannot embed.
24
+ *
25
+ * `off` embeds nothing by design, and its embedder says so by throwing
26
+ * (EmbeddingDisabledError). The pipeline has a designed skip for a *missing*
27
+ * embedder — semantic filters are passed over and logged at debug — but a
28
+ * present embedder that throws lands in the per-filter catch instead, which is
29
+ * an error-level `filter_anchor_match_failed` for every semantic filter on
30
+ * every synced message. Same outcome, a log nobody can read.
31
+ *
32
+ * `local` on the self-host profiles reaches the same place by a different road:
33
+ * the backend and imap-worker images ship without `@huggingface/transformers`,
34
+ * so the first embed throws ERR_MODULE_NOT_FOUND. The pipeline classifies that
35
+ * one throw through `@remit/search-service/capability` and remembers it, and
36
+ * from then on the config here wires no embedder either — one probe per process,
37
+ * not one per filter per message.
38
+ */
39
+ const embedderFromEnv = (): MessageEmbedder | undefined =>
40
+ readEmbeddingProviderFromEnv() === EMBEDDING_PROVIDER_OFF ||
41
+ isEmbeddingCapabilityUnavailable()
42
+ ? undefined
43
+ : getMessageEmbedder();
44
+
45
+ /**
46
+ * Assemble the filter config a body-materializing pass runs (RFC 034), shared by
47
+ * the imap-worker's sync path and the backend's read-path backfill so both embed
48
+ * under the same model that produced the anchors — a semantic (anchor-only)
49
+ * filter is evaluated on either path instead of silently skipped.
50
+ *
51
+ * Absent the placement mover there is no move path, so filters stay off — a
52
+ * matched filter's actions reuse the same enqueue plumbing the placement mover
53
+ * owns.
54
+ */
55
+ export const buildFilterConfig = (
56
+ deps: FilterConfigDeps,
57
+ embedder?: MessageEmbedder,
58
+ ): FilterConfig | undefined => {
59
+ const { placementMoveService } = deps;
60
+ if (!placementMoveService) return undefined;
61
+ const resolved = embedder ?? embedderFromEnv();
62
+ return {
63
+ filterService: deps.filterService,
64
+ filterAnchorService: deps.filterAnchorService,
65
+ messageLabelService: deps.messageLabelService,
66
+ placementMoveService,
67
+ ...(resolved ? { embedder: resolved } : {}),
68
+ };
69
+ };
@@ -0,0 +1,39 @@
1
+ import assert from "node:assert/strict";
2
+ import { afterEach, describe, it } from "node:test";
3
+ import {
4
+ _resetMessageEmbedderForTest,
5
+ getMessageEmbedder,
6
+ } from "./message-embedder.js";
7
+
8
+ describe("getMessageEmbedder", () => {
9
+ afterEach(() => {
10
+ _resetMessageEmbedderForTest();
11
+ });
12
+
13
+ it("adapts the batch embedding service into a single-text message embedder", async () => {
14
+ const vector = await getMessageEmbedder().embed("invoice from stripe");
15
+
16
+ assert.ok(Array.isArray(vector));
17
+ assert.ok(vector.length > 0);
18
+ assert.ok(vector.every((value) => typeof value === "number"));
19
+ });
20
+
21
+ it("embeds the same text deterministically", async () => {
22
+ const embedder = getMessageEmbedder();
23
+ const a = await embedder.embed("your monthly receipt");
24
+ const b = await embedder.embed("your monthly receipt");
25
+
26
+ assert.deepEqual(a, b);
27
+ });
28
+
29
+ it("memoizes the embedder across calls", () => {
30
+ assert.strictEqual(getMessageEmbedder(), getMessageEmbedder());
31
+ });
32
+
33
+ it("rebuilds after a reset", () => {
34
+ const first = getMessageEmbedder();
35
+ _resetMessageEmbedderForTest();
36
+
37
+ assert.notStrictEqual(getMessageEmbedder(), first);
38
+ });
39
+ });
@@ -0,0 +1,33 @@
1
+ import { buildEmbeddingServiceFromEnv } from "@remit/search-service/from-env";
2
+ import type { MessageEmbedder } from "./pipeline.js";
3
+
4
+ /**
5
+ * Adapt the batch {@link EmbeddingService} the rest of the stack composes from
6
+ * env into the single-text {@link MessageEmbedder} the filter pipeline needs: one
7
+ * incoming message becomes one message-level vector to compare against a filter's
8
+ * persisted anchor (RFC 034 Decision 2.1/2.3).
9
+ *
10
+ * The embedder is selected by the same `SEARCH_EMBEDDING_*` env the search-index
11
+ * worker reads, so every filter pass embeds under the identical model the anchors
12
+ * were built with. The instance is memoized across warm invocations so a local
13
+ * Transformers.js model loads once per container rather than per event.
14
+ */
15
+ let cached: MessageEmbedder | undefined;
16
+
17
+ export const getMessageEmbedder = (): MessageEmbedder => {
18
+ if (cached) return cached;
19
+ const service = buildEmbeddingServiceFromEnv();
20
+ cached = {
21
+ embed: async (text: string): Promise<number[]> => {
22
+ const [vector] = await service.embed([text]);
23
+ return vector;
24
+ },
25
+ embeddingId: service.embeddingId,
26
+ };
27
+ return cached;
28
+ };
29
+
30
+ /** Reset the singleton — test use only. */
31
+ export const _resetMessageEmbedderForTest = (): void => {
32
+ cached = undefined;
33
+ };
@@ -6,6 +6,11 @@ import type {
6
6
  IMessageLabelRepository,
7
7
  } from "@remit/data-ports";
8
8
  import { FilterState } from "@remit/domain-enums";
9
+ import {
10
+ isEmbeddingCapabilityAbsence,
11
+ isEmbeddingCapabilityUnavailable,
12
+ recordEmbeddingCapabilityAbsence,
13
+ } from "@remit/search-service/capability";
9
14
  import type { PlacementMoveService } from "../placement-move.js";
10
15
  import {
11
16
  type AnchorEmbedder,
@@ -131,7 +136,29 @@ export class FilterPipeline {
131
136
  messageEmbedding = null;
132
137
  return null;
133
138
  }
134
- messageEmbedding = await embedder.embed(buildMatchText(msg));
139
+ messageEmbedding = await embedder
140
+ .embed(buildMatchText(msg))
141
+ .catch((error: unknown) => {
142
+ // This deployment carries no embedding pipeline at all — the
143
+ // container images ship without `@huggingface/transformers`. That is
144
+ // the same absence `SEARCH_EMBEDDING_PROVIDER=off` states, so it takes
145
+ // the same designed skip rather than an error-level
146
+ // `filter_anchor_match_failed` per semantic filter per message. The
147
+ // absence is remembered process-wide, so this is the only embed that
148
+ // pays for it: `buildFilterConfig` wires no embedder afterwards.
149
+ if (!isEmbeddingCapabilityAbsence(error)) throw error;
150
+ if (!isEmbeddingCapabilityUnavailable()) {
151
+ this.log.warn?.(
152
+ {
153
+ accountConfigId,
154
+ error: inspect(error),
155
+ },
156
+ "No embedding pipeline in this deployment; semantic filters are skipped (literal filters still run)",
157
+ );
158
+ }
159
+ recordEmbeddingCapabilityAbsence(error);
160
+ return null;
161
+ });
135
162
  return messageEmbedding;
136
163
  };
137
164