anonyma 1.0.0 → 1.1.0

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.
Files changed (183) hide show
  1. package/CHANGELOG.md +273 -144
  2. package/README.md +427 -143
  3. package/dist/ai/index.cjs +522 -0
  4. package/dist/ai/index.cjs.map +1 -0
  5. package/dist/ai/index.d.cts +322 -0
  6. package/dist/ai/index.d.ts +322 -0
  7. package/dist/ai/index.js +505 -0
  8. package/dist/ai/index.js.map +1 -0
  9. package/dist/audit/index.cjs +359 -0
  10. package/dist/audit/index.cjs.map +1 -0
  11. package/dist/audit/index.d.cts +287 -0
  12. package/dist/audit/index.d.ts +287 -0
  13. package/dist/audit/index.js +340 -0
  14. package/dist/audit/index.js.map +1 -0
  15. package/dist/chunk-26IHMQHL.cjs +662 -0
  16. package/dist/chunk-26IHMQHL.cjs.map +1 -0
  17. package/dist/chunk-2F2GWGTB.js +174 -0
  18. package/dist/chunk-2F2GWGTB.js.map +1 -0
  19. package/dist/chunk-2NNMJ55L.cjs +45 -0
  20. package/dist/chunk-2NNMJ55L.cjs.map +1 -0
  21. package/dist/chunk-2YG7R4KJ.cjs +114 -0
  22. package/dist/chunk-2YG7R4KJ.cjs.map +1 -0
  23. package/dist/chunk-2ZZHLCEJ.js +237 -0
  24. package/dist/chunk-2ZZHLCEJ.js.map +1 -0
  25. package/dist/chunk-3CY6ATKS.cjs +308 -0
  26. package/dist/chunk-3CY6ATKS.cjs.map +1 -0
  27. package/dist/chunk-3D5ZCJ7R.js +268 -0
  28. package/dist/chunk-3D5ZCJ7R.js.map +1 -0
  29. package/dist/chunk-6LDAIJBL.js +103 -0
  30. package/dist/chunk-6LDAIJBL.js.map +1 -0
  31. package/dist/chunk-7HK2V5Q4.js +111 -0
  32. package/dist/chunk-7HK2V5Q4.js.map +1 -0
  33. package/dist/chunk-AAP472NG.js +177 -0
  34. package/dist/chunk-AAP472NG.js.map +1 -0
  35. package/dist/chunk-AP4JIUNW.js +42 -0
  36. package/dist/chunk-AP4JIUNW.js.map +1 -0
  37. package/dist/chunk-BBJPQPGP.js +457 -0
  38. package/dist/chunk-BBJPQPGP.js.map +1 -0
  39. package/dist/chunk-CRXRJZEV.cjs +82 -0
  40. package/dist/chunk-CRXRJZEV.cjs.map +1 -0
  41. package/dist/chunk-E3PICZYI.js +20 -0
  42. package/dist/chunk-E3PICZYI.js.map +1 -0
  43. package/dist/chunk-F7O2FA3F.cjs +1255 -0
  44. package/dist/chunk-F7O2FA3F.cjs.map +1 -0
  45. package/dist/chunk-FB56NAAH.cjs +243 -0
  46. package/dist/chunk-FB56NAAH.cjs.map +1 -0
  47. package/dist/chunk-G4AN7FND.js +1063 -0
  48. package/dist/chunk-G4AN7FND.js.map +1 -0
  49. package/dist/chunk-GS3FRH4K.js +294 -0
  50. package/dist/chunk-GS3FRH4K.js.map +1 -0
  51. package/dist/chunk-GYMG5OI3.cjs +186 -0
  52. package/dist/chunk-GYMG5OI3.cjs.map +1 -0
  53. package/dist/chunk-IHEIYZPL.cjs +301 -0
  54. package/dist/chunk-IHEIYZPL.cjs.map +1 -0
  55. package/dist/chunk-IJAMLYOT.cjs +1087 -0
  56. package/dist/chunk-IJAMLYOT.cjs.map +1 -0
  57. package/dist/chunk-ISL5B4LU.cjs +181 -0
  58. package/dist/chunk-ISL5B4LU.cjs.map +1 -0
  59. package/dist/chunk-JKW36IPT.js +179 -0
  60. package/dist/chunk-JKW36IPT.js.map +1 -0
  61. package/dist/chunk-KOCQBDI4.cjs +179 -0
  62. package/dist/chunk-KOCQBDI4.cjs.map +1 -0
  63. package/dist/chunk-OGNLB2KA.js +79 -0
  64. package/dist/chunk-OGNLB2KA.js.map +1 -0
  65. package/dist/chunk-PKHZ7CVI.js +28 -0
  66. package/dist/chunk-PKHZ7CVI.js.map +1 -0
  67. package/dist/chunk-QYD4F6T7.cjs +114 -0
  68. package/dist/chunk-QYD4F6T7.cjs.map +1 -0
  69. package/dist/chunk-RCW5IPPR.js +1288 -0
  70. package/dist/chunk-RCW5IPPR.js.map +1 -0
  71. package/dist/chunk-RMMKYHUN.cjs +35 -0
  72. package/dist/chunk-RMMKYHUN.cjs.map +1 -0
  73. package/dist/chunk-S4BJ2ADB.cjs +22 -0
  74. package/dist/chunk-S4BJ2ADB.cjs.map +1 -0
  75. package/dist/chunk-S7NH6CHY.cjs +462 -0
  76. package/dist/chunk-S7NH6CHY.cjs.map +1 -0
  77. package/dist/chunk-SIHPDZXA.js +176 -0
  78. package/dist/chunk-SIHPDZXA.js.map +1 -0
  79. package/dist/chunk-TW7PM3UJ.cjs +30 -0
  80. package/dist/chunk-TW7PM3UJ.cjs.map +1 -0
  81. package/dist/chunk-UA3EOFBS.cjs +1322 -0
  82. package/dist/chunk-UA3EOFBS.cjs.map +1 -0
  83. package/dist/chunk-VE4CCW7W.js +32 -0
  84. package/dist/chunk-VE4CCW7W.js.map +1 -0
  85. package/dist/chunk-W6RBCHYF.js +651 -0
  86. package/dist/chunk-W6RBCHYF.js.map +1 -0
  87. package/dist/chunk-WLJN7GIS.cjs +189 -0
  88. package/dist/chunk-WLJN7GIS.cjs.map +1 -0
  89. package/dist/chunk-YMLR375Z.js +1244 -0
  90. package/dist/chunk-YMLR375Z.js.map +1 -0
  91. package/dist/compliance/index.cjs +88 -0
  92. package/dist/compliance/index.cjs.map +1 -0
  93. package/dist/compliance/index.d.cts +277 -0
  94. package/dist/compliance/index.d.ts +277 -0
  95. package/dist/compliance/index.js +47 -0
  96. package/dist/compliance/index.js.map +1 -0
  97. package/dist/crypto.cjs +12 -136
  98. package/dist/crypto.cjs.map +1 -1
  99. package/dist/crypto.d.cts +4 -3
  100. package/dist/crypto.d.ts +4 -3
  101. package/dist/crypto.js +3 -137
  102. package/dist/crypto.js.map +1 -1
  103. package/dist/detectors/index.cjs +146 -1483
  104. package/dist/detectors/index.cjs.map +1 -1
  105. package/dist/detectors/index.d.cts +1 -1
  106. package/dist/detectors/index.d.ts +1 -1
  107. package/dist/detectors/index.js +3 -1449
  108. package/dist/detectors/index.js.map +1 -1
  109. package/dist/engine/index.cjs +441 -0
  110. package/dist/engine/index.cjs.map +1 -0
  111. package/dist/engine/index.d.cts +728 -0
  112. package/dist/engine/index.d.ts +728 -0
  113. package/dist/engine/index.js +189 -0
  114. package/dist/engine/index.js.map +1 -0
  115. package/dist/hash-2PSYNDMM.cjs +14 -0
  116. package/dist/hash-2PSYNDMM.cjs.map +1 -0
  117. package/dist/hash-CELLDNKN.js +5 -0
  118. package/dist/hash-CELLDNKN.js.map +1 -0
  119. package/dist/index.cjs +153 -2816
  120. package/dist/index.cjs.map +1 -1
  121. package/dist/index.d.cts +93 -11
  122. package/dist/index.d.ts +93 -11
  123. package/dist/index.js +13 -2791
  124. package/dist/index.js.map +1 -1
  125. package/dist/json-CKDvlRtl.d.ts +147 -0
  126. package/dist/json-s8ny-RCo.d.cts +147 -0
  127. package/dist/mcp/index.cjs +751 -0
  128. package/dist/mcp/index.cjs.map +1 -0
  129. package/dist/mcp/index.d.cts +226 -0
  130. package/dist/mcp/index.d.ts +226 -0
  131. package/dist/mcp/index.js +744 -0
  132. package/dist/mcp/index.js.map +1 -0
  133. package/dist/middleware/express.cjs +93 -0
  134. package/dist/middleware/express.cjs.map +1 -0
  135. package/dist/middleware/express.d.cts +90 -0
  136. package/dist/middleware/express.d.ts +90 -0
  137. package/dist/middleware/express.js +91 -0
  138. package/dist/middleware/express.js.map +1 -0
  139. package/dist/middleware/hono.cjs +58 -0
  140. package/dist/middleware/hono.cjs.map +1 -0
  141. package/dist/middleware/hono.d.cts +75 -0
  142. package/dist/middleware/hono.d.ts +75 -0
  143. package/dist/middleware/hono.js +56 -0
  144. package/dist/middleware/hono.js.map +1 -0
  145. package/dist/middleware/index.cjs +24 -0
  146. package/dist/middleware/index.cjs.map +1 -0
  147. package/dist/middleware/index.d.cts +106 -0
  148. package/dist/middleware/index.d.ts +106 -0
  149. package/dist/middleware/index.js +15 -0
  150. package/dist/middleware/index.js.map +1 -0
  151. package/dist/schemas.cjs +5 -5
  152. package/dist/schemas.cjs.map +1 -1
  153. package/dist/schemas.d.cts +20 -10
  154. package/dist/schemas.d.ts +20 -10
  155. package/dist/schemas.js +5 -5
  156. package/dist/schemas.js.map +1 -1
  157. package/dist/session-BdRftyZ5.d.cts +102 -0
  158. package/dist/session-Cs2zlNQp.d.ts +102 -0
  159. package/dist/stream.cjs +13 -2262
  160. package/dist/stream.cjs.map +1 -1
  161. package/dist/stream.d.cts +1 -1
  162. package/dist/stream.d.ts +1 -1
  163. package/dist/stream.js +6 -2255
  164. package/dist/stream.js.map +1 -1
  165. package/dist/types--tPm_AjY.d.ts +199 -0
  166. package/dist/types-0BJH_wqc.d.cts +199 -0
  167. package/dist/types-CPUhjvl8.d.cts +629 -0
  168. package/dist/types-Dtht33aq.d.ts +629 -0
  169. package/dist/{types-D-9HzVuy.d.cts → types-EJTj_BKY.d.cts} +38 -7
  170. package/dist/{types-D-9HzVuy.d.ts → types-EJTj_BKY.d.ts} +38 -7
  171. package/dist/types-Zt_bZvwS.d.cts +182 -0
  172. package/dist/types-Zt_bZvwS.d.ts +182 -0
  173. package/dist/validators.cjs +39 -181
  174. package/dist/validators.cjs.map +1 -1
  175. package/dist/validators.js +1 -174
  176. package/dist/validators.js.map +1 -1
  177. package/dist/vault/index.cjs +688 -0
  178. package/dist/vault/index.cjs.map +1 -0
  179. package/dist/vault/index.d.cts +385 -0
  180. package/dist/vault/index.d.ts +385 -0
  181. package/dist/vault/index.js +669 -0
  182. package/dist/vault/index.js.map +1 -0
  183. package/package.json +128 -11
@@ -0,0 +1,322 @@
1
+ import { K as KeyRule } from '../json-s8ny-RCo.cjs';
2
+ export { J as JsonSanitizeOptions, a as JsonSanitizeResult, s as sanitizeJson, b as sanitizeJsonAsync } from '../json-s8ny-RCo.cjs';
3
+ import { T as TokenizationProvider, A as AsyncChunkTransformer, R as RestoreResult, c as PipelineSpec } from '../types-CPUhjvl8.cjs';
4
+ import { A as AuditFieldEntry } from '../types-Zt_bZvwS.cjs';
5
+ import { S as SessionTokenizer, a as SessionSnapshot, b as SessionTokenizerOptions } from '../session-BdRftyZ5.cjs';
6
+ import '../types-EJTj_BKY.cjs';
7
+
8
+ /**
9
+ * @module ai/restore
10
+ * @description Restoring tokens in model output that arrives as a stream. A
11
+ * token may be split across chunks (`[EMAIL_00` + `01]`), so the transformer
12
+ * holds back a trailing fragment that could still become a token.
13
+ */
14
+
15
+ /**
16
+ * Create a push-style transformer that restores the tokens of `provider` in
17
+ * text arriving in arbitrary chunks.
18
+ *
19
+ * The output is the same as restoring the whole text at once, however the
20
+ * text is split, and accepts the same spellings of a token: a provider with
21
+ * its own `restore()` (a lenient session) is used through it.
22
+ *
23
+ * @param provider - The provider that issued the tokens.
24
+ * @returns An {@link AsyncChunkTransformer}. Await each `push()` before the next.
25
+ *
26
+ * @example
27
+ * ```ts
28
+ * const restorer = createRestoreTransformer(session);
29
+ * for await (const delta of modelStream) ui.append(await restorer.push(delta));
30
+ * ui.append(await restorer.flush());
31
+ * ```
32
+ */
33
+ declare function createRestoreTransformer(provider: TokenizationProvider): AsyncChunkTransformer;
34
+ /**
35
+ * Create a WHATWG `TransformStream` that restores tokens in a stream of text.
36
+ *
37
+ * @param provider - The provider that issued the tokens.
38
+ * @returns A `TransformStream<string, string>`.
39
+ * @throws {@link ValidationError} When `TransformStream` is unavailable.
40
+ *
41
+ * @example
42
+ * ```ts
43
+ * const restored = modelTextStream.pipeThrough(createRestoreStream(session));
44
+ * ```
45
+ */
46
+ declare function createRestoreStream(provider: TokenizationProvider): TransformStream<string, string>;
47
+ /**
48
+ * Restore the tokens inside a JSON document held as text, such as the
49
+ * arguments of a tool call. Values are restored into the parsed document and
50
+ * the document is serialised again, so a restored value that contains a
51
+ * quote, a backslash or a line break can neither break the document nor add
52
+ * members to it. A text that is not JSON is restored as plain text; a
53
+ * document without tokens is returned as it was written.
54
+ *
55
+ * @param text - The JSON document.
56
+ * @param provider - The provider that issued the tokens.
57
+ * @returns The document with restored values.
58
+ * @throws {@link ValidationError} When the document is nested deeper than 512 levels.
59
+ *
60
+ * @example
61
+ * ```ts
62
+ * await restoreJsonText('{"to":"[EMAIL_0001]"}', session); // '{"to":"alice@example.com"}'
63
+ * ```
64
+ */
65
+ declare function restoreJsonText(text: string, provider: TokenizationProvider): Promise<string>;
66
+ /**
67
+ * Reads and writes the text of a structured stream chunk.
68
+ */
69
+ interface ChunkLens<C> {
70
+ /** The text of the chunk, or `undefined` when the chunk carries no text. */
71
+ readonly get: (chunk: C) => string | undefined;
72
+ /** A copy of the chunk with its text replaced. */
73
+ readonly set: (chunk: C, text: string) => C;
74
+ /**
75
+ * Identifies the text block a chunk belongs to, when a stream can carry
76
+ * several at once. Chunks of different blocks are restored independently.
77
+ */
78
+ readonly block?: (chunk: C) => string | undefined;
79
+ }
80
+ /**
81
+ * Lens for chunks shaped `{ type: "text-delta", delta }` or
82
+ * `{ type: "text-delta", textDelta }` (the stream parts of the Vercel AI SDK).
83
+ */
84
+ declare const textDeltaLens: ChunkLens<Record<string, unknown>>;
85
+ /**
86
+ * Lens for chunks whose text is a string member named `content` (message
87
+ * chunks of LangChain) .
88
+ */
89
+ declare const contentLens: ChunkLens<Record<string, unknown>>;
90
+ /**
91
+ * Lens for chunks whose text is a string member named `delta` (response
92
+ * chunks of LlamaIndex).
93
+ */
94
+ declare const deltaLens: ChunkLens<Record<string, unknown>>;
95
+ /**
96
+ * Restore tokens in a stream of structured chunks. Text that is held back
97
+ * because it may be the start of a token is delivered with the next text
98
+ * chunk of its block, and in any case before the next chunk that carries no
99
+ * text, so the order of the stream is kept: nothing arrives after the chunk
100
+ * that ends its block. A token that is split around a chunk without text is
101
+ * therefore not restored.
102
+ *
103
+ * @param source - The chunks produced by the model SDK.
104
+ * @param lens - How to read and write the text of a chunk.
105
+ * @param provider - The provider that issued the tokens.
106
+ * @returns The chunks with restored text.
107
+ *
108
+ * @example
109
+ * ```ts
110
+ * for await (const chunk of restoreChunks(await chain.stream(input), contentLens, session)) {
111
+ * process.stdout.write(String(chunk.content));
112
+ * }
113
+ * ```
114
+ */
115
+ declare function restoreChunks<C>(source: AsyncIterable<C>, lens: ChunkLens<C>, provider: TokenizationProvider): AsyncGenerator<C, void, undefined>;
116
+
117
+ /**
118
+ * @module ai/guard
119
+ * @description Redaction middleware for LLM calls: personal data is replaced
120
+ * by tokens before a prompt leaves the process, and the tokens are replaced
121
+ * by the original values in what comes back — also when it comes back as a
122
+ * stream. The mapping between tokens and values never leaves the process.
123
+ */
124
+
125
+ /**
126
+ * Options accepted by {@link createLlmGuard}.
127
+ */
128
+ interface LlmGuardOptions {
129
+ /**
130
+ * What to detect. The strategy members of the spec are ignored: every
131
+ * detection is replaced by a session token. Defaults to every built-in
132
+ * category.
133
+ */
134
+ readonly spec?: Omit<PipelineSpec, "defaultStrategy" | "rules">;
135
+ /** Options for the session tokenizer of each exchange. `lenient` defaults to `true`. */
136
+ readonly session?: Omit<SessionTokenizerOptions, "snapshot">;
137
+ /** Message roles that are sent unchanged, for example `["system"]`. */
138
+ readonly skipRoles?: readonly string[];
139
+ /** Keys whose values are tokenized whole, such as `password`. */
140
+ readonly keyRules?: readonly KeyRule[];
141
+ /**
142
+ * Further members that must reach the model byte for byte, by key name at
143
+ * any depth. Signatures, encrypted reasoning, provider options and media
144
+ * payloads are passed on without this. Whatever is passed on is listed in
145
+ * {@link SanitizedMessages.unscanned}.
146
+ */
147
+ readonly passthroughKeys?: readonly string[];
148
+ /**
149
+ * Counts the tokens of a text with the tokenizer of the target model. When
150
+ * given, {@link SanitizedMessages.tokens} reports the prompt size before and
151
+ * after sanitization.
152
+ */
153
+ readonly countTokens?: (text: string) => number;
154
+ }
155
+ /**
156
+ * Sanitized chat messages and an account of what was replaced.
157
+ */
158
+ interface SanitizedMessages<M> {
159
+ /** The messages to send: same length, same structure, tokens in place of personal data. */
160
+ readonly messages: M[];
161
+ /** What was replaced where, without any of the replaced text. */
162
+ readonly fields: readonly AuditFieldEntry[];
163
+ /** Total number of replacements. */
164
+ readonly replaced: number;
165
+ /**
166
+ * JSON Pointers of what was passed on without inspection: media payloads
167
+ * (images, files, audio), members a provider needs byte for byte
168
+ * (signatures, encrypted reasoning), `passthroughKeys`, and values that are
169
+ * not JSON.
170
+ */
171
+ readonly unscanned: readonly string[];
172
+ /** JSON Pointers of members whose key held personal data and was replaced by a token. */
173
+ readonly flaggedKeys: readonly string[];
174
+ /** Prompt size in model tokens before and after, when `countTokens` is configured. */
175
+ readonly tokens?: {
176
+ readonly before: number;
177
+ readonly after: number;
178
+ };
179
+ }
180
+ /**
181
+ * One guarded exchange with a model: a request and its response, or a whole
182
+ * conversation. Everything sanitized through the same exchange shares one
183
+ * token mapping, so a value keeps its token across turns.
184
+ */
185
+ interface LlmExchange {
186
+ /** The session that holds the token mapping. */
187
+ readonly session: SessionTokenizer;
188
+ /** Replace personal data in a text with tokens. */
189
+ readonly sanitizeText: (text: string) => string;
190
+ /**
191
+ * Replace personal data in chat messages with tokens, preserving their
192
+ * structure. Structural members are left alone where the message formats
193
+ * define them — `role`, `name`, identifiers and `type` on a message or a
194
+ * content part, the function name of a tool call — and nowhere else: tool
195
+ * arguments and tool results are sanitized at every depth, object keys
196
+ * included. Of a media part the payload is passed on and listed in
197
+ * `unscanned`; its text (title, context, file name, a text source) is
198
+ * sanitized.
199
+ *
200
+ * @throws ValidationError When a message is not a plain object. A class
201
+ * instance cannot be copied with tokens in place of its content: sanitize
202
+ * the text before the message object is built, with `sanitizeText()`.
203
+ */
204
+ readonly sanitizeMessages: <M>(messages: readonly M[]) => SanitizedMessages<M>;
205
+ /** Replace tokens in model output with the original values. */
206
+ readonly restoreText: (text: string) => RestoreResult;
207
+ /** A transform stream that restores tokens in streamed model output. */
208
+ readonly restoreStream: () => TransformStream<string, string>;
209
+ /** Restore tokens in model output delivered as an async iterable of text. */
210
+ readonly restoreIterable: (source: AsyncIterable<string>) => AsyncGenerator<string, void, undefined>;
211
+ /** A serialisable copy of the token mapping, to continue the conversation later. */
212
+ readonly snapshot: () => SessionSnapshot;
213
+ }
214
+ /**
215
+ * The redaction middleware.
216
+ */
217
+ interface LlmGuard {
218
+ /**
219
+ * Start an exchange.
220
+ *
221
+ * @param snapshot - The snapshot of an earlier exchange to continue from.
222
+ */
223
+ readonly begin: (snapshot?: SessionSnapshot) => LlmExchange;
224
+ }
225
+ /**
226
+ * Create the redaction middleware for LLM calls.
227
+ *
228
+ * @param options - What to detect and how tokens look.
229
+ * @returns The {@link LlmGuard}.
230
+ * @throws {@link ValidationError} When the spec is invalid.
231
+ *
232
+ * @example
233
+ * ```ts
234
+ * import { createLlmGuard } from "anonyma/ai";
235
+ *
236
+ * const guard = createLlmGuard({ skipRoles: ["system"] });
237
+ *
238
+ * const exchange = guard.begin();
239
+ * const { messages } = exchange.sanitizeMessages([
240
+ * { role: "system", content: "You are a support assistant." },
241
+ * { role: "user", content: "Email alice@example.com about invoice 42." },
242
+ * ]);
243
+ * // messages[1].content === "Email [EMAIL_0001] about invoice 42."
244
+ *
245
+ * const stream = await llm.streamText(messages); // "Sure, I wrote to [EMAIL_0001]."
246
+ * for await (const text of exchange.restoreIterable(stream)) {
247
+ * process.stdout.write(text); // "Sure, I wrote to alice@example.com."
248
+ * }
249
+ * ```
250
+ */
251
+ declare function createLlmGuard(options?: LlmGuardOptions): LlmGuard;
252
+ /**
253
+ * The parts of a language-model middleware this adapter implements. The
254
+ * shape matches `LanguageModelMiddleware` of the Vercel AI SDK structurally,
255
+ * so the object can be passed to `wrapLanguageModel()` without this library
256
+ * depending on the SDK.
257
+ */
258
+ interface LanguageModelMiddlewareLike {
259
+ /** Sanitizes the prompt before it reaches the model. */
260
+ readonly transformParams: (input: {
261
+ readonly params: Record<string, unknown>;
262
+ }) => Promise<Record<string, unknown>>;
263
+ /** Restores tokens in a complete response. */
264
+ readonly wrapGenerate: (input: {
265
+ readonly doGenerate: () => PromiseLike<unknown>;
266
+ readonly params: Record<string, unknown>;
267
+ }) => Promise<unknown>;
268
+ /** Restores tokens in a streamed response. */
269
+ readonly wrapStream: (input: {
270
+ readonly doStream: () => PromiseLike<{
271
+ readonly stream: ReadableStream<unknown>;
272
+ }>;
273
+ readonly params: Record<string, unknown>;
274
+ }) => Promise<{
275
+ readonly stream: ReadableStream<unknown>;
276
+ }>;
277
+ }
278
+ /**
279
+ * Options accepted by {@link toLanguageModelMiddleware}.
280
+ */
281
+ interface LanguageModelMiddlewareOptions {
282
+ /**
283
+ * Called with the account of every sanitized prompt: what was replaced,
284
+ * and above all what was passed on without inspection (`unscanned`). Throw
285
+ * from it to stop a request that must not be sent.
286
+ */
287
+ readonly onSanitized?: (report: SanitizedMessages<unknown>) => void;
288
+ }
289
+ /**
290
+ * Adapt a guard to the language-model middleware shape of the Vercel AI SDK.
291
+ *
292
+ * The prompt (`params.prompt`) is sanitized on the way in. On the way out,
293
+ * tokens are restored in the result of `doGenerate()` and in the `text-delta`
294
+ * and `tool-call` parts of `doStream()`; the arguments of a tool call are
295
+ * restored as the JSON document they are. Partial tool input
296
+ * (`tool-input-delta`) is passed on as it is. Each call gets its own exchange.
297
+ *
298
+ * @param guard - The guard to adapt.
299
+ * @param options - A callback that receives the account of each sanitized prompt.
300
+ * @returns An object to pass as `middleware` to `wrapLanguageModel()`.
301
+ *
302
+ * @remarks
303
+ * The adapter is typed structurally and tested against stand-ins for the SDK,
304
+ * not against the SDK itself. Check it against the SDK version you use.
305
+ * `wrapGenerate` and `wrapStream` need the `params` object that
306
+ * `transformParams` returned (a copy made with spread syntax works too) and
307
+ * throw when they get another one: restoring nothing would go unnoticed.
308
+ *
309
+ * @example
310
+ * ```ts
311
+ * import { wrapLanguageModel } from "ai";
312
+ * import { createLlmGuard, toLanguageModelMiddleware } from "anonyma/ai";
313
+ *
314
+ * const model = wrapLanguageModel({
315
+ * model: baseModel,
316
+ * middleware: toLanguageModelMiddleware(createLlmGuard()),
317
+ * });
318
+ * ```
319
+ */
320
+ declare function toLanguageModelMiddleware(guard: LlmGuard, options?: LanguageModelMiddlewareOptions): LanguageModelMiddlewareLike;
321
+
322
+ export { type ChunkLens, KeyRule, type LanguageModelMiddlewareLike, type LanguageModelMiddlewareOptions, type LlmExchange, type LlmGuard, type LlmGuardOptions, type SanitizedMessages, contentLens, createLlmGuard, createRestoreStream, createRestoreTransformer, deltaLens, restoreChunks, restoreJsonText, textDeltaLens, toLanguageModelMiddleware };
@@ -0,0 +1,322 @@
1
+ import { K as KeyRule } from '../json-CKDvlRtl.js';
2
+ export { J as JsonSanitizeOptions, a as JsonSanitizeResult, s as sanitizeJson, b as sanitizeJsonAsync } from '../json-CKDvlRtl.js';
3
+ import { T as TokenizationProvider, A as AsyncChunkTransformer, R as RestoreResult, c as PipelineSpec } from '../types-Dtht33aq.js';
4
+ import { A as AuditFieldEntry } from '../types-Zt_bZvwS.js';
5
+ import { S as SessionTokenizer, a as SessionSnapshot, b as SessionTokenizerOptions } from '../session-Cs2zlNQp.js';
6
+ import '../types-EJTj_BKY.js';
7
+
8
+ /**
9
+ * @module ai/restore
10
+ * @description Restoring tokens in model output that arrives as a stream. A
11
+ * token may be split across chunks (`[EMAIL_00` + `01]`), so the transformer
12
+ * holds back a trailing fragment that could still become a token.
13
+ */
14
+
15
+ /**
16
+ * Create a push-style transformer that restores the tokens of `provider` in
17
+ * text arriving in arbitrary chunks.
18
+ *
19
+ * The output is the same as restoring the whole text at once, however the
20
+ * text is split, and accepts the same spellings of a token: a provider with
21
+ * its own `restore()` (a lenient session) is used through it.
22
+ *
23
+ * @param provider - The provider that issued the tokens.
24
+ * @returns An {@link AsyncChunkTransformer}. Await each `push()` before the next.
25
+ *
26
+ * @example
27
+ * ```ts
28
+ * const restorer = createRestoreTransformer(session);
29
+ * for await (const delta of modelStream) ui.append(await restorer.push(delta));
30
+ * ui.append(await restorer.flush());
31
+ * ```
32
+ */
33
+ declare function createRestoreTransformer(provider: TokenizationProvider): AsyncChunkTransformer;
34
+ /**
35
+ * Create a WHATWG `TransformStream` that restores tokens in a stream of text.
36
+ *
37
+ * @param provider - The provider that issued the tokens.
38
+ * @returns A `TransformStream<string, string>`.
39
+ * @throws {@link ValidationError} When `TransformStream` is unavailable.
40
+ *
41
+ * @example
42
+ * ```ts
43
+ * const restored = modelTextStream.pipeThrough(createRestoreStream(session));
44
+ * ```
45
+ */
46
+ declare function createRestoreStream(provider: TokenizationProvider): TransformStream<string, string>;
47
+ /**
48
+ * Restore the tokens inside a JSON document held as text, such as the
49
+ * arguments of a tool call. Values are restored into the parsed document and
50
+ * the document is serialised again, so a restored value that contains a
51
+ * quote, a backslash or a line break can neither break the document nor add
52
+ * members to it. A text that is not JSON is restored as plain text; a
53
+ * document without tokens is returned as it was written.
54
+ *
55
+ * @param text - The JSON document.
56
+ * @param provider - The provider that issued the tokens.
57
+ * @returns The document with restored values.
58
+ * @throws {@link ValidationError} When the document is nested deeper than 512 levels.
59
+ *
60
+ * @example
61
+ * ```ts
62
+ * await restoreJsonText('{"to":"[EMAIL_0001]"}', session); // '{"to":"alice@example.com"}'
63
+ * ```
64
+ */
65
+ declare function restoreJsonText(text: string, provider: TokenizationProvider): Promise<string>;
66
+ /**
67
+ * Reads and writes the text of a structured stream chunk.
68
+ */
69
+ interface ChunkLens<C> {
70
+ /** The text of the chunk, or `undefined` when the chunk carries no text. */
71
+ readonly get: (chunk: C) => string | undefined;
72
+ /** A copy of the chunk with its text replaced. */
73
+ readonly set: (chunk: C, text: string) => C;
74
+ /**
75
+ * Identifies the text block a chunk belongs to, when a stream can carry
76
+ * several at once. Chunks of different blocks are restored independently.
77
+ */
78
+ readonly block?: (chunk: C) => string | undefined;
79
+ }
80
+ /**
81
+ * Lens for chunks shaped `{ type: "text-delta", delta }` or
82
+ * `{ type: "text-delta", textDelta }` (the stream parts of the Vercel AI SDK).
83
+ */
84
+ declare const textDeltaLens: ChunkLens<Record<string, unknown>>;
85
+ /**
86
+ * Lens for chunks whose text is a string member named `content` (message
87
+ * chunks of LangChain) .
88
+ */
89
+ declare const contentLens: ChunkLens<Record<string, unknown>>;
90
+ /**
91
+ * Lens for chunks whose text is a string member named `delta` (response
92
+ * chunks of LlamaIndex).
93
+ */
94
+ declare const deltaLens: ChunkLens<Record<string, unknown>>;
95
+ /**
96
+ * Restore tokens in a stream of structured chunks. Text that is held back
97
+ * because it may be the start of a token is delivered with the next text
98
+ * chunk of its block, and in any case before the next chunk that carries no
99
+ * text, so the order of the stream is kept: nothing arrives after the chunk
100
+ * that ends its block. A token that is split around a chunk without text is
101
+ * therefore not restored.
102
+ *
103
+ * @param source - The chunks produced by the model SDK.
104
+ * @param lens - How to read and write the text of a chunk.
105
+ * @param provider - The provider that issued the tokens.
106
+ * @returns The chunks with restored text.
107
+ *
108
+ * @example
109
+ * ```ts
110
+ * for await (const chunk of restoreChunks(await chain.stream(input), contentLens, session)) {
111
+ * process.stdout.write(String(chunk.content));
112
+ * }
113
+ * ```
114
+ */
115
+ declare function restoreChunks<C>(source: AsyncIterable<C>, lens: ChunkLens<C>, provider: TokenizationProvider): AsyncGenerator<C, void, undefined>;
116
+
117
+ /**
118
+ * @module ai/guard
119
+ * @description Redaction middleware for LLM calls: personal data is replaced
120
+ * by tokens before a prompt leaves the process, and the tokens are replaced
121
+ * by the original values in what comes back — also when it comes back as a
122
+ * stream. The mapping between tokens and values never leaves the process.
123
+ */
124
+
125
+ /**
126
+ * Options accepted by {@link createLlmGuard}.
127
+ */
128
+ interface LlmGuardOptions {
129
+ /**
130
+ * What to detect. The strategy members of the spec are ignored: every
131
+ * detection is replaced by a session token. Defaults to every built-in
132
+ * category.
133
+ */
134
+ readonly spec?: Omit<PipelineSpec, "defaultStrategy" | "rules">;
135
+ /** Options for the session tokenizer of each exchange. `lenient` defaults to `true`. */
136
+ readonly session?: Omit<SessionTokenizerOptions, "snapshot">;
137
+ /** Message roles that are sent unchanged, for example `["system"]`. */
138
+ readonly skipRoles?: readonly string[];
139
+ /** Keys whose values are tokenized whole, such as `password`. */
140
+ readonly keyRules?: readonly KeyRule[];
141
+ /**
142
+ * Further members that must reach the model byte for byte, by key name at
143
+ * any depth. Signatures, encrypted reasoning, provider options and media
144
+ * payloads are passed on without this. Whatever is passed on is listed in
145
+ * {@link SanitizedMessages.unscanned}.
146
+ */
147
+ readonly passthroughKeys?: readonly string[];
148
+ /**
149
+ * Counts the tokens of a text with the tokenizer of the target model. When
150
+ * given, {@link SanitizedMessages.tokens} reports the prompt size before and
151
+ * after sanitization.
152
+ */
153
+ readonly countTokens?: (text: string) => number;
154
+ }
155
+ /**
156
+ * Sanitized chat messages and an account of what was replaced.
157
+ */
158
+ interface SanitizedMessages<M> {
159
+ /** The messages to send: same length, same structure, tokens in place of personal data. */
160
+ readonly messages: M[];
161
+ /** What was replaced where, without any of the replaced text. */
162
+ readonly fields: readonly AuditFieldEntry[];
163
+ /** Total number of replacements. */
164
+ readonly replaced: number;
165
+ /**
166
+ * JSON Pointers of what was passed on without inspection: media payloads
167
+ * (images, files, audio), members a provider needs byte for byte
168
+ * (signatures, encrypted reasoning), `passthroughKeys`, and values that are
169
+ * not JSON.
170
+ */
171
+ readonly unscanned: readonly string[];
172
+ /** JSON Pointers of members whose key held personal data and was replaced by a token. */
173
+ readonly flaggedKeys: readonly string[];
174
+ /** Prompt size in model tokens before and after, when `countTokens` is configured. */
175
+ readonly tokens?: {
176
+ readonly before: number;
177
+ readonly after: number;
178
+ };
179
+ }
180
+ /**
181
+ * One guarded exchange with a model: a request and its response, or a whole
182
+ * conversation. Everything sanitized through the same exchange shares one
183
+ * token mapping, so a value keeps its token across turns.
184
+ */
185
+ interface LlmExchange {
186
+ /** The session that holds the token mapping. */
187
+ readonly session: SessionTokenizer;
188
+ /** Replace personal data in a text with tokens. */
189
+ readonly sanitizeText: (text: string) => string;
190
+ /**
191
+ * Replace personal data in chat messages with tokens, preserving their
192
+ * structure. Structural members are left alone where the message formats
193
+ * define them — `role`, `name`, identifiers and `type` on a message or a
194
+ * content part, the function name of a tool call — and nowhere else: tool
195
+ * arguments and tool results are sanitized at every depth, object keys
196
+ * included. Of a media part the payload is passed on and listed in
197
+ * `unscanned`; its text (title, context, file name, a text source) is
198
+ * sanitized.
199
+ *
200
+ * @throws ValidationError When a message is not a plain object. A class
201
+ * instance cannot be copied with tokens in place of its content: sanitize
202
+ * the text before the message object is built, with `sanitizeText()`.
203
+ */
204
+ readonly sanitizeMessages: <M>(messages: readonly M[]) => SanitizedMessages<M>;
205
+ /** Replace tokens in model output with the original values. */
206
+ readonly restoreText: (text: string) => RestoreResult;
207
+ /** A transform stream that restores tokens in streamed model output. */
208
+ readonly restoreStream: () => TransformStream<string, string>;
209
+ /** Restore tokens in model output delivered as an async iterable of text. */
210
+ readonly restoreIterable: (source: AsyncIterable<string>) => AsyncGenerator<string, void, undefined>;
211
+ /** A serialisable copy of the token mapping, to continue the conversation later. */
212
+ readonly snapshot: () => SessionSnapshot;
213
+ }
214
+ /**
215
+ * The redaction middleware.
216
+ */
217
+ interface LlmGuard {
218
+ /**
219
+ * Start an exchange.
220
+ *
221
+ * @param snapshot - The snapshot of an earlier exchange to continue from.
222
+ */
223
+ readonly begin: (snapshot?: SessionSnapshot) => LlmExchange;
224
+ }
225
+ /**
226
+ * Create the redaction middleware for LLM calls.
227
+ *
228
+ * @param options - What to detect and how tokens look.
229
+ * @returns The {@link LlmGuard}.
230
+ * @throws {@link ValidationError} When the spec is invalid.
231
+ *
232
+ * @example
233
+ * ```ts
234
+ * import { createLlmGuard } from "anonyma/ai";
235
+ *
236
+ * const guard = createLlmGuard({ skipRoles: ["system"] });
237
+ *
238
+ * const exchange = guard.begin();
239
+ * const { messages } = exchange.sanitizeMessages([
240
+ * { role: "system", content: "You are a support assistant." },
241
+ * { role: "user", content: "Email alice@example.com about invoice 42." },
242
+ * ]);
243
+ * // messages[1].content === "Email [EMAIL_0001] about invoice 42."
244
+ *
245
+ * const stream = await llm.streamText(messages); // "Sure, I wrote to [EMAIL_0001]."
246
+ * for await (const text of exchange.restoreIterable(stream)) {
247
+ * process.stdout.write(text); // "Sure, I wrote to alice@example.com."
248
+ * }
249
+ * ```
250
+ */
251
+ declare function createLlmGuard(options?: LlmGuardOptions): LlmGuard;
252
+ /**
253
+ * The parts of a language-model middleware this adapter implements. The
254
+ * shape matches `LanguageModelMiddleware` of the Vercel AI SDK structurally,
255
+ * so the object can be passed to `wrapLanguageModel()` without this library
256
+ * depending on the SDK.
257
+ */
258
+ interface LanguageModelMiddlewareLike {
259
+ /** Sanitizes the prompt before it reaches the model. */
260
+ readonly transformParams: (input: {
261
+ readonly params: Record<string, unknown>;
262
+ }) => Promise<Record<string, unknown>>;
263
+ /** Restores tokens in a complete response. */
264
+ readonly wrapGenerate: (input: {
265
+ readonly doGenerate: () => PromiseLike<unknown>;
266
+ readonly params: Record<string, unknown>;
267
+ }) => Promise<unknown>;
268
+ /** Restores tokens in a streamed response. */
269
+ readonly wrapStream: (input: {
270
+ readonly doStream: () => PromiseLike<{
271
+ readonly stream: ReadableStream<unknown>;
272
+ }>;
273
+ readonly params: Record<string, unknown>;
274
+ }) => Promise<{
275
+ readonly stream: ReadableStream<unknown>;
276
+ }>;
277
+ }
278
+ /**
279
+ * Options accepted by {@link toLanguageModelMiddleware}.
280
+ */
281
+ interface LanguageModelMiddlewareOptions {
282
+ /**
283
+ * Called with the account of every sanitized prompt: what was replaced,
284
+ * and above all what was passed on without inspection (`unscanned`). Throw
285
+ * from it to stop a request that must not be sent.
286
+ */
287
+ readonly onSanitized?: (report: SanitizedMessages<unknown>) => void;
288
+ }
289
+ /**
290
+ * Adapt a guard to the language-model middleware shape of the Vercel AI SDK.
291
+ *
292
+ * The prompt (`params.prompt`) is sanitized on the way in. On the way out,
293
+ * tokens are restored in the result of `doGenerate()` and in the `text-delta`
294
+ * and `tool-call` parts of `doStream()`; the arguments of a tool call are
295
+ * restored as the JSON document they are. Partial tool input
296
+ * (`tool-input-delta`) is passed on as it is. Each call gets its own exchange.
297
+ *
298
+ * @param guard - The guard to adapt.
299
+ * @param options - A callback that receives the account of each sanitized prompt.
300
+ * @returns An object to pass as `middleware` to `wrapLanguageModel()`.
301
+ *
302
+ * @remarks
303
+ * The adapter is typed structurally and tested against stand-ins for the SDK,
304
+ * not against the SDK itself. Check it against the SDK version you use.
305
+ * `wrapGenerate` and `wrapStream` need the `params` object that
306
+ * `transformParams` returned (a copy made with spread syntax works too) and
307
+ * throw when they get another one: restoring nothing would go unnoticed.
308
+ *
309
+ * @example
310
+ * ```ts
311
+ * import { wrapLanguageModel } from "ai";
312
+ * import { createLlmGuard, toLanguageModelMiddleware } from "anonyma/ai";
313
+ *
314
+ * const model = wrapLanguageModel({
315
+ * model: baseModel,
316
+ * middleware: toLanguageModelMiddleware(createLlmGuard()),
317
+ * });
318
+ * ```
319
+ */
320
+ declare function toLanguageModelMiddleware(guard: LlmGuard, options?: LanguageModelMiddlewareOptions): LanguageModelMiddlewareLike;
321
+
322
+ export { type ChunkLens, KeyRule, type LanguageModelMiddlewareLike, type LanguageModelMiddlewareOptions, type LlmExchange, type LlmGuard, type LlmGuardOptions, type SanitizedMessages, contentLens, createLlmGuard, createRestoreStream, createRestoreTransformer, deltaLens, restoreChunks, restoreJsonText, textDeltaLens, toLanguageModelMiddleware };