plugin-ai-api 1.0.24 → 1.0.25

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 (54) hide show
  1. package/dist/client/757.56952e321dc399b7.js +10 -0
  2. package/dist/client/index.js +1 -1
  3. package/dist/client-v2/757.db678ca1aa6c422c.js +10 -0
  4. package/dist/client-v2/index.js +1 -1
  5. package/dist/externalVersion.js +8 -8
  6. package/dist/locale/en-US.json +1 -0
  7. package/dist/locale/vi-VN.json +1 -0
  8. package/dist/locale/zh-CN.json +1 -0
  9. package/dist/server/billing.js +6 -1
  10. package/dist/server/collections/ai-api-config.js +6 -0
  11. package/dist/server/collections/ai-api-usage-records.js +1 -0
  12. package/dist/server/migrations/20260813000000-add-prompt-cache-tokens.js +69 -0
  13. package/dist/server/plugin.js +10 -0
  14. package/dist/server/resource/ai-api-config.js +5 -0
  15. package/dist/server/resource/ai-api-usage-monitor.js +3 -1
  16. package/dist/server/routes/chat-completions.js +89 -10
  17. package/dist/server/routes/completions.js +32 -10
  18. package/dist/server/services/file-processor.js +262 -0
  19. package/dist/server/usage.js +33 -3
  20. package/dist/server/utils/direct-llm-context.js +150 -15
  21. package/dist/server/utils/openai-format.js +21 -2
  22. package/dist/swagger.js +42 -3
  23. package/package.json +1 -1
  24. package/src/client-v2/pages/UsagePage.tsx +9 -0
  25. package/src/locale/en-US.json +1 -0
  26. package/src/locale/vi-VN.json +1 -0
  27. package/src/locale/zh-CN.json +1 -0
  28. package/src/server/__tests__/direct-llm-context.test.ts +87 -6
  29. package/src/server/__tests__/openai-format.test.ts +12 -2
  30. package/src/server/__tests__/request-body.test.ts +45 -2
  31. package/src/server/__tests__/usage-route.test.ts +120 -3
  32. package/src/server/__tests__/usage.test.ts +19 -0
  33. package/src/server/billing.ts +6 -1
  34. package/src/server/collections/ai-api-config.ts +8 -0
  35. package/src/server/collections/ai-api-role-permissions.ts +41 -41
  36. package/src/server/collections/ai-api-usage-records.ts +1 -0
  37. package/src/server/index.ts +10 -10
  38. package/src/server/middleware/rate-limit.ts +70 -70
  39. package/src/server/migrations/20260813000000-add-prompt-cache-tokens.ts +46 -0
  40. package/src/server/plugin.ts +20 -0
  41. package/src/server/resource/ai-api-config.ts +5 -0
  42. package/src/server/resource/ai-api-usage-monitor.ts +3 -0
  43. package/src/server/routes/chat-completions.ts +134 -11
  44. package/src/server/routes/completions.ts +33 -7
  45. package/src/server/services/__tests__/file-processor.test.ts +184 -0
  46. package/src/server/services/file-processor.ts +323 -0
  47. package/src/server/usage.ts +47 -1
  48. package/src/server/utils/direct-llm-context.ts +198 -20
  49. package/src/server/utils/openai-format.ts +25 -2
  50. package/src/server/utils/rate-limiter.ts +83 -83
  51. package/src/server/utils/resolve-service.ts +82 -82
  52. package/src/swagger.ts +45 -3
  53. package/dist/client/757.a01403fb7a1bea01.js +0 -10
  54. package/dist/client-v2/757.a117ce1cf7119cea.js +0 -10
@@ -33,6 +33,7 @@ var import_openai_format = require("../utils/openai-format");
33
33
  var import_resolve_service = require("../utils/resolve-service");
34
34
  var import_user_permissions = require("../utils/user-permissions");
35
35
  var import_streaming = require("../utils/streaming");
36
+ var import_chat_completions = require("./chat-completions");
36
37
  var import_usage = require("../usage");
37
38
  var import_billing = require("../billing");
38
39
  var import_direct_llm_context = require("../utils/direct-llm-context");
@@ -141,6 +142,8 @@ async function handleCompletions(ctx, plugin) {
141
142
  ]);
142
143
  const completionId = (0, import_openai_format.generateCompletionId)().replace("chatcmpl-", "cmpl-");
143
144
  const chatModel = provider.createModel();
145
+ const providerRequestParameters = (0, import_chat_completions.getProviderRequestParameters)(body);
146
+ (0, import_chat_completions.applyProviderRequestParameters)(chatModel, providerRequestParameters);
144
147
  (0, import_billing.markLlmProviderAttempted)(ctx);
145
148
  if (stream) {
146
149
  await handleStreamingTextCompletion(
@@ -149,10 +152,18 @@ async function handleCompletions(ctx, plugin) {
149
152
  langchainMessages,
150
153
  completionId,
151
154
  body.model,
152
- body.stream_options
155
+ body.stream_options,
156
+ providerRequestParameters
153
157
  );
154
158
  } else {
155
- await handleNonStreamingTextCompletion(ctx, chatModel, langchainMessages, completionId, body.model);
159
+ await handleNonStreamingTextCompletion(
160
+ ctx,
161
+ chatModel,
162
+ langchainMessages,
163
+ completionId,
164
+ body.model,
165
+ providerRequestParameters
166
+ );
156
167
  }
157
168
  } catch (err) {
158
169
  ctx.log.error("AI API completions error:", err);
@@ -170,8 +181,8 @@ async function handleCompletions(ctx, plugin) {
170
181
  }
171
182
  }
172
183
  }
173
- async function handleNonStreamingTextCompletion(ctx, chatModel, messages, completionId, modelName) {
174
- const result = await chatModel.invoke(messages);
184
+ async function handleNonStreamingTextCompletion(ctx, chatModel, messages, completionId, modelName, providerRequestParameters) {
185
+ const result = await chatModel.invoke(messages, providerRequestParameters);
175
186
  let text = "";
176
187
  if (typeof result.content === "string") {
177
188
  text = result.content;
@@ -179,10 +190,15 @@ async function handleNonStreamingTextCompletion(ctx, chatModel, messages, comple
179
190
  const textPart = result.content.find((c) => c.type === "text");
180
191
  text = (textPart == null ? void 0 : textPart.text) || JSON.stringify(result.content);
181
192
  }
182
- const usage = (0, import_usage.setAiApiUsageResult)(ctx, result.usage_metadata, {
183
- gatewayResponseId: completionId,
184
- providerRequestId: (0, import_usage.extractProviderRequestId)(result)
185
- });
193
+ const usage = (0, import_usage.setAiApiUsageResult)(
194
+ ctx,
195
+ result.usage_metadata,
196
+ {
197
+ gatewayResponseId: completionId,
198
+ providerRequestId: (0, import_usage.extractProviderRequestId)(result)
199
+ },
200
+ result.response_metadata
201
+ );
186
202
  ctx.status = 200;
187
203
  ctx.body = {
188
204
  id: completionId,
@@ -198,10 +214,15 @@ async function handleNonStreamingTextCompletion(ctx, chatModel, messages, comple
198
214
  finish_reason: "stop"
199
215
  }
200
216
  ],
201
- usage: usage ?? { prompt_tokens: 0, completion_tokens: 0, total_tokens: 0 }
217
+ usage: usage ? {
218
+ prompt_tokens: usage.prompt_tokens,
219
+ completion_tokens: usage.completion_tokens,
220
+ total_tokens: usage.total_tokens,
221
+ prompt_tokens_details: { cached_tokens: usage.prompt_cache_tokens ?? null }
222
+ } : { prompt_tokens: 0, completion_tokens: 0, total_tokens: 0, prompt_tokens_details: { cached_tokens: null } }
202
223
  };
203
224
  }
204
- async function handleStreamingTextCompletion(ctx, chatModel, messages, completionId, modelName, streamOptions) {
225
+ async function handleStreamingTextCompletion(ctx, chatModel, messages, completionId, modelName, streamOptions, providerRequestParameters) {
205
226
  ctx.set({
206
227
  "Content-Type": "text/event-stream",
207
228
  "Cache-Control": "no-cache",
@@ -214,6 +235,7 @@ async function handleStreamingTextCompletion(ctx, chatModel, messages, completio
214
235
  let providerRequestId;
215
236
  try {
216
237
  const stream = await chatModel.stream(messages, {
238
+ ...providerRequestParameters,
217
239
  stream_options: { ...streamOptions, include_usage: true },
218
240
  signal: requestAbort.signal
219
241
  });
@@ -0,0 +1,262 @@
1
+ /**
2
+ * This file is part of the NocoBase (R) project.
3
+ * Copyright (c) 2020-2024 NocoBase Co., Ltd.
4
+ * Authors: NocoBase Team.
5
+ *
6
+ * This project is dual-licensed under AGPL-3.0 and NocoBase Commercial License.
7
+ * For more information, please refer to: https://www.nocobase.com/agreement.
8
+ */
9
+
10
+ var __defProp = Object.defineProperty;
11
+ var __getOwnPropDesc = Object.getOwnPropertyDescriptor;
12
+ var __getOwnPropNames = Object.getOwnPropertyNames;
13
+ var __hasOwnProp = Object.prototype.hasOwnProperty;
14
+ var __export = (target, all) => {
15
+ for (var name in all)
16
+ __defProp(target, name, { get: all[name], enumerable: true });
17
+ };
18
+ var __copyProps = (to, from, except, desc) => {
19
+ if (from && typeof from === "object" || typeof from === "function") {
20
+ for (let key of __getOwnPropNames(from))
21
+ if (!__hasOwnProp.call(to, key) && key !== except)
22
+ __defProp(to, key, { get: () => from[key], enumerable: !(desc = __getOwnPropDesc(from, key)) || desc.enumerable });
23
+ }
24
+ return to;
25
+ };
26
+ var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: true }), mod);
27
+ var file_processor_exports = {};
28
+ __export(file_processor_exports, {
29
+ FileProcessorError: () => FileProcessorError,
30
+ FileProcessorService: () => FileProcessorService,
31
+ base64FileForwarder: () => base64FileForwarder,
32
+ fetchFileAsBase64: () => fetchFileAsBase64,
33
+ httpFileUrlFetcher: () => httpFileUrlFetcher,
34
+ pdfFileProcessor: () => pdfFileProcessor
35
+ });
36
+ module.exports = __toCommonJS(file_processor_exports);
37
+ var import_path = require("path");
38
+ class FileProcessorError extends Error {
39
+ constructor(code, message) {
40
+ super(message);
41
+ this.code = code;
42
+ this.name = "FileProcessorError";
43
+ }
44
+ }
45
+ const DEFAULT_MAX_FILE_SIZE = 50 * 1024 * 1024;
46
+ const DEFAULT_TIMEOUT_MS = 3e4;
47
+ const ALLOWED_PROTOCOLS = /* @__PURE__ */ new Set(["http:", "https:"]);
48
+ class FileProcessorService {
49
+ processors = [];
50
+ pdfRenderer = null;
51
+ register(processor) {
52
+ this.unregister(processor.name);
53
+ this.processors.push(processor);
54
+ }
55
+ unregister(name) {
56
+ this.processors = this.processors.filter((p) => p.name !== name);
57
+ }
58
+ async process(block, context) {
59
+ const processor = [...this.processors].reverse().find((p) => p.canHandle(block));
60
+ if (!processor) {
61
+ return block;
62
+ }
63
+ return processor.process(block, context);
64
+ }
65
+ list() {
66
+ return [...this.processors];
67
+ }
68
+ /**
69
+ * Register a renderer used to convert PDF pages into PNG images when
70
+ * `pdfRenderPagesAsImages` is enabled in the AI API config.
71
+ */
72
+ registerPdfRenderer(renderer) {
73
+ this.pdfRenderer = renderer;
74
+ }
75
+ unregisterPdfRenderer() {
76
+ this.pdfRenderer = null;
77
+ }
78
+ getPdfRenderer() {
79
+ return this.pdfRenderer;
80
+ }
81
+ }
82
+ function isRecord(value) {
83
+ return typeof value === "object" && value !== null && !Array.isArray(value);
84
+ }
85
+ function getUrlString(value) {
86
+ if (typeof value === "string") return value;
87
+ if (isRecord(value) && typeof value.url === "string") return value.url;
88
+ return void 0;
89
+ }
90
+ function extractFilename(url, contentDisposition) {
91
+ if (contentDisposition) {
92
+ const match = contentDisposition.match(/filename="?([^"]+)"?/);
93
+ if (match) return match[1];
94
+ }
95
+ try {
96
+ const pathname = new URL(url).pathname;
97
+ if (pathname) return (0, import_path.basename)(pathname);
98
+ } catch {
99
+ }
100
+ return void 0;
101
+ }
102
+ async function fetchFileAsBase64(url, options = {}) {
103
+ const maxSize = options.maxSizeBytes ?? DEFAULT_MAX_FILE_SIZE;
104
+ const timeoutMs = options.timeoutMs ?? DEFAULT_TIMEOUT_MS;
105
+ const allowedProtocols = options.allowedProtocols ? new Set(options.allowedProtocols) : ALLOWED_PROTOCOLS;
106
+ let protocol;
107
+ try {
108
+ protocol = new URL(url).protocol;
109
+ } catch {
110
+ throw new FileProcessorError("invalid_url", `File URL '${url}' is not a valid URL.`);
111
+ }
112
+ if (!allowedProtocols.has(protocol)) {
113
+ throw new FileProcessorError("unsupported_protocol", `File URL protocol '${protocol}' is not allowed.`);
114
+ }
115
+ const controller = new AbortController();
116
+ const timeout = setTimeout(() => controller.abort(), timeoutMs);
117
+ try {
118
+ const response = await fetch(url, {
119
+ signal: controller.signal,
120
+ redirect: "follow"
121
+ });
122
+ if (!response.ok) {
123
+ throw new FileProcessorError(
124
+ "fetch_failed",
125
+ `Failed to fetch file from '${url}': ${response.status} ${response.statusText}`
126
+ );
127
+ }
128
+ const contentLength = response.headers.get("content-length");
129
+ if (contentLength && Number(contentLength) > maxSize) {
130
+ throw new FileProcessorError("file_too_large", `File at '${url}' exceeds maximum allowed size.`);
131
+ }
132
+ const contentType = response.headers.get("content-type") || void 0;
133
+ if (options.allowedContentTypes && contentType && !options.allowedContentTypes.some((type) => contentType.includes(type))) {
134
+ throw new FileProcessorError("content_type_not_allowed", `File content type '${contentType}' is not allowed.`);
135
+ }
136
+ const buffer = Buffer.from(await response.arrayBuffer());
137
+ if (buffer.length > maxSize) {
138
+ throw new FileProcessorError("file_too_large", `File at '${url}' exceeds maximum allowed size.`);
139
+ }
140
+ const mimeType = (contentType == null ? void 0 : contentType.split(";")[0].trim()) ?? "application/octet-stream";
141
+ const contentDisposition = response.headers.get("content-disposition");
142
+ const filename = extractFilename(url, contentDisposition) ?? "file";
143
+ return {
144
+ fileData: `data:${mimeType};base64,${buffer.toString("base64")}`,
145
+ mimeType,
146
+ filename
147
+ };
148
+ } finally {
149
+ clearTimeout(timeout);
150
+ }
151
+ }
152
+ const base64FileForwarder = {
153
+ name: "base64FileForwarder",
154
+ canHandle(block) {
155
+ if (block.type !== "file") return false;
156
+ const file = isRecord(block.file) ? block.file : void 0;
157
+ if (!file) return false;
158
+ const fileData = String(file.file_data ?? "");
159
+ return fileData.startsWith("data:") && fileData.includes(";base64,");
160
+ },
161
+ async process(block) {
162
+ return block;
163
+ }
164
+ };
165
+ const httpFileUrlFetcher = {
166
+ name: "httpFileUrlFetcher",
167
+ canHandle(block) {
168
+ if (block.type !== "file_url") return false;
169
+ const fileUrl = isRecord(block.file_url) ? block.file_url : void 0;
170
+ if (!fileUrl) return false;
171
+ const url = String(fileUrl.url ?? "");
172
+ return url.startsWith("http://") || url.startsWith("https://");
173
+ },
174
+ async process(block) {
175
+ const fileUrl = isRecord(block.file_url) ? block.file_url : void 0;
176
+ const url = String((fileUrl == null ? void 0 : fileUrl.url) ?? "");
177
+ if (!url) {
178
+ throw new FileProcessorError("missing_url", "file_url block requires a 'url' property.");
179
+ }
180
+ const { fileData, mimeType, filename } = await fetchFileAsBase64(url);
181
+ return {
182
+ type: "file",
183
+ file: {
184
+ file_data: fileData,
185
+ mime_type: mimeType,
186
+ filename: filename || "file"
187
+ }
188
+ };
189
+ }
190
+ };
191
+ function decodeBase64DataUrl(url) {
192
+ const match = /^data:([^;]+);base64,([A-Za-z0-9+/]+=*)$/.exec(url);
193
+ if (!match) return void 0;
194
+ try {
195
+ const buffer = Buffer.from(match[2], "base64");
196
+ return { mimeType: match[1].toLowerCase(), buffer };
197
+ } catch {
198
+ return void 0;
199
+ }
200
+ }
201
+ function isPdfBuffer(buffer) {
202
+ return buffer.length >= 4 && buffer.toString("binary", 0, 4) === "%PDF";
203
+ }
204
+ function isPdfFileBlock(block) {
205
+ if (block.type !== "file") return false;
206
+ const file = isRecord(block.file) ? block.file : void 0;
207
+ if (!file) return false;
208
+ const fileData = String(file.file_data ?? "");
209
+ if (!fileData.startsWith("data:")) return false;
210
+ const mimeType = String(file.mime_type ?? "").toLowerCase();
211
+ if (mimeType === "application/pdf") return true;
212
+ if (fileData.startsWith("data:application/pdf")) return true;
213
+ const decoded = decodeBase64DataUrl(fileData);
214
+ if (decoded && isPdfBuffer(decoded.buffer)) return true;
215
+ return false;
216
+ }
217
+ function getPluginFromContext(context) {
218
+ var _a, _b, _c;
219
+ return (_c = (_b = (_a = context.ctx.app) == null ? void 0 : _a.pm) == null ? void 0 : _b.get) == null ? void 0 : _c.call(_b, "plugin-ai-api");
220
+ }
221
+ const pdfFileProcessor = {
222
+ name: "pdfFileProcessor",
223
+ canHandle(block) {
224
+ return isPdfFileBlock(block);
225
+ },
226
+ async process(block, context) {
227
+ var _a, _b, _c, _d;
228
+ const config = await context.ctx.db.getRepository("aiApiConfig").findOne();
229
+ if (!(config == null ? void 0 : config.pdfRenderPagesAsImages)) {
230
+ return block;
231
+ }
232
+ const plugin = getPluginFromContext(context);
233
+ const renderer = (_b = (_a = plugin == null ? void 0 : plugin.fileProcessorService) == null ? void 0 : _a.getPdfRenderer) == null ? void 0 : _b.call(_a);
234
+ if (!renderer) {
235
+ (_d = (_c = context.ctx.log) == null ? void 0 : _c.warn) == null ? void 0 : _d.call(
236
+ _c,
237
+ "[pdfFileProcessor] pdfRenderPagesAsImages is enabled but no PdfToImageRenderer is registered. Forwarding PDF as a file block."
238
+ );
239
+ return block;
240
+ }
241
+ const file = isRecord(block.file) ? block.file : void 0;
242
+ const fileData = String((file == null ? void 0 : file.file_data) ?? "");
243
+ const decoded = decodeBase64DataUrl(fileData);
244
+ if (!decoded) {
245
+ return block;
246
+ }
247
+ const pages = await renderer.render(decoded.buffer);
248
+ return pages.map((buffer) => ({
249
+ type: "image_url",
250
+ image_url: { url: `data:image/png;base64,${buffer.toString("base64")}` }
251
+ }));
252
+ }
253
+ };
254
+ // Annotate the CommonJS export names for ESM import in node:
255
+ 0 && (module.exports = {
256
+ FileProcessorError,
257
+ FileProcessorService,
258
+ base64FileForwarder,
259
+ fetchFileAsBase64,
260
+ httpFileUrlFetcher,
261
+ pdfFileProcessor
262
+ });
@@ -42,6 +42,24 @@ function getAiApiState(ctx) {
42
42
  function normalizeTokenCount(value) {
43
43
  return typeof value === "number" && Number.isSafeInteger(value) && value >= 0 ? value : null;
44
44
  }
45
+ function extractPromptCacheTokens(source) {
46
+ const inputDetails = source.input_token_details;
47
+ if (inputDetails && typeof inputDetails === "object") {
48
+ const cacheRead = inputDetails.cache_read;
49
+ if (typeof cacheRead === "number") return cacheRead;
50
+ const cachedTokens = inputDetails.cached_tokens;
51
+ if (typeof cachedTokens === "number") return cachedTokens;
52
+ }
53
+ const promptDetails = source.prompt_tokens_details;
54
+ if (promptDetails && typeof promptDetails === "object") {
55
+ const cachedTokens = promptDetails.cached_tokens;
56
+ if (typeof cachedTokens === "number") return cachedTokens;
57
+ }
58
+ if (typeof source.cached_tokens === "number") return source.cached_tokens;
59
+ if (typeof source.cache_read_tokens === "number") return source.cache_read_tokens;
60
+ if (typeof source.cache_read === "number") return source.cache_read;
61
+ return null;
62
+ }
45
63
  function normalizeUsage(value) {
46
64
  if (!value || typeof value !== "object") return void 0;
47
65
  const source = value;
@@ -55,11 +73,22 @@ function normalizeUsage(value) {
55
73
  return {
56
74
  prompt_tokens: prompt,
57
75
  completion_tokens: completion,
58
- total_tokens: total
76
+ total_tokens: total,
77
+ prompt_cache_tokens: extractPromptCacheTokens(source)
59
78
  };
60
79
  }
61
- function setAiApiUsageResult(ctx, value, metadata = {}) {
62
- const usage = normalizeUsage(value);
80
+ function setAiApiUsageResult(ctx, value, metadata = {}, responseMetadata) {
81
+ let usage = normalizeUsage(value);
82
+ if ((usage == null ? void 0 : usage.prompt_cache_tokens) === null && responseMetadata && typeof responseMetadata === "object") {
83
+ const responseMetaRecord = responseMetadata;
84
+ let cacheTokens = extractPromptCacheTokens(responseMetaRecord);
85
+ if (cacheTokens === null && responseMetaRecord.usage && typeof responseMetaRecord.usage === "object") {
86
+ cacheTokens = extractPromptCacheTokens(responseMetaRecord.usage);
87
+ }
88
+ if (cacheTokens !== null) {
89
+ usage = { ...usage, prompt_cache_tokens: cacheTokens };
90
+ }
91
+ }
63
92
  getAiApiState(ctx).aiApiUsageResult = usage ? { source: "provider", usage, ...metadata } : { source: "unavailable", ...metadata };
64
93
  (0, import_app_observability.addAiApiUsage)(ctx, usage);
65
94
  return usage;
@@ -133,6 +162,7 @@ async function finishUsageRecord(ctx, id, startedAt, status) {
133
162
  inputTokens: (usage == null ? void 0 : usage.prompt_tokens) ?? null,
134
163
  outputTokens: (usage == null ? void 0 : usage.completion_tokens) ?? null,
135
164
  totalTokens: (usage == null ? void 0 : usage.total_tokens) ?? null,
165
+ promptCacheTokens: (usage == null ? void 0 : usage.prompt_cache_tokens) ?? null,
136
166
  resolvedService: ((_c = (_b = state.aiApiLlmBilling) == null ? void 0 : _b.resolution) == null ? void 0 : _c.service) ?? null,
137
167
  resolvedProvider: ((_e = (_d = state.aiApiLlmBilling) == null ? void 0 : _d.resolution) == null ? void 0 : _e.provider) ?? null,
138
168
  resolvedModel: ((_g = (_f = state.aiApiLlmBilling) == null ? void 0 : _f.resolution) == null ? void 0 : _g.model) ?? null,
@@ -27,6 +27,7 @@ var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: tru
27
27
  var direct_llm_context_exports = {};
28
28
  __export(direct_llm_context_exports, {
29
29
  DirectLlmContextError: () => DirectLlmContextError,
30
+ parseImageDimensions: () => parseImageDimensions,
30
31
  prepareDirectLlmContext: () => prepareDirectLlmContext
31
32
  });
32
33
  module.exports = __toCommonJS(direct_llm_context_exports);
@@ -45,20 +46,159 @@ function positiveInteger(value) {
45
46
  const parsed = Number(value);
46
47
  return Number.isSafeInteger(parsed) && parsed > 0 ? parsed : void 0;
47
48
  }
48
- function hasImageContent(value) {
49
- return Array.isArray(value) && value.some(
50
- (block) => typeof block === "object" && block !== null && block.type === "image_url"
51
- );
52
- }
53
49
  function estimateValueTokens(value) {
54
50
  if (value === void 0) return 0;
55
51
  return Math.ceil(Buffer.byteLength(JSON.stringify(value), "utf8") / 3);
56
52
  }
57
- function estimateMessagesTokens(messages) {
58
- return messages.reduce((total, message) => total + estimateValueTokens(message) + 4, 0);
53
+ function isRecord(value) {
54
+ return typeof value === "object" && value !== null && !Array.isArray(value);
55
+ }
56
+ function readBigEndian(buf, offset) {
57
+ return buf.readUInt32BE(offset);
58
+ }
59
+ function readUInt16BE(buf, offset) {
60
+ return buf.readUInt16BE(offset);
61
+ }
62
+ function readLittleEndian(buf, offset) {
63
+ return buf.readUInt16LE(offset);
64
+ }
65
+ function parsePngDimensions(buffer) {
66
+ if (buffer.length < 24) return void 0;
67
+ return { width: readBigEndian(buffer, 16), height: readBigEndian(buffer, 20) };
68
+ }
69
+ function parseJpegDimensions(buffer) {
70
+ let offset = 2;
71
+ while (offset < buffer.length) {
72
+ if (buffer[offset] !== 255) {
73
+ offset++;
74
+ continue;
75
+ }
76
+ const marker = buffer[offset + 1];
77
+ if (marker >= 192 && marker <= 195 || marker >= 197 && marker <= 199 || marker >= 201 && marker <= 203 || marker >= 205 && marker <= 207) {
78
+ if (offset + 9 <= buffer.length) {
79
+ return { height: readUInt16BE(buffer, offset + 5), width: readUInt16BE(buffer, offset + 7) };
80
+ }
81
+ return void 0;
82
+ }
83
+ if (marker === 217 || offset + 4 >= buffer.length) break;
84
+ const segmentLength = buffer.readUInt16BE(offset + 2);
85
+ offset += 2 + segmentLength;
86
+ }
87
+ return void 0;
88
+ }
89
+ function parseGifDimensions(buffer) {
90
+ if (buffer.length < 10) return void 0;
91
+ return { width: readLittleEndian(buffer, 6), height: readLittleEndian(buffer, 8) };
92
+ }
93
+ function parseWebpDimensions(buffer) {
94
+ if (buffer.length < 30) return void 0;
95
+ const riff = buffer.toString("ascii", 0, 4);
96
+ const webp = buffer.toString("ascii", 8, 12);
97
+ if (riff !== "RIFF" || webp !== "WEBP") return void 0;
98
+ const chunkType = buffer.toString("ascii", 12, 16);
99
+ if (chunkType === "VP8 " && buffer.length >= 26) {
100
+ return { width: readLittleEndian(buffer, 26), height: readLittleEndian(buffer, 28) };
101
+ }
102
+ if (chunkType === "VP8L" && buffer.length >= 24) {
103
+ const bits = buffer.readUInt32LE(21);
104
+ return {
105
+ width: (bits & 16383) + 1,
106
+ height: (bits >> 14 & 16383) + 1
107
+ };
108
+ }
109
+ if (chunkType === "VP8X" && buffer.length >= 30) {
110
+ return {
111
+ width: ((buffer[24] | buffer[25] << 8 | buffer[26] << 16) & 16777215) + 1,
112
+ height: ((buffer[27] | buffer[28] << 8 | buffer[29] << 16) & 16777215) + 1
113
+ };
114
+ }
115
+ return void 0;
116
+ }
117
+ function parseImageDimensions(buffer) {
118
+ if (buffer.length < 12) return void 0;
119
+ const header = buffer.toString("binary", 0, 4);
120
+ if (header === "\x89PNG") return parsePngDimensions(buffer);
121
+ if (header === "GIF8") return parseGifDimensions(buffer);
122
+ if (header === "RIFF") return parseWebpDimensions(buffer);
123
+ if (buffer[0] === 255 && buffer[1] === 216) return parseJpegDimensions(buffer);
124
+ return void 0;
59
125
  }
60
- function containsUnsupportedContent(messages) {
61
- return messages.some((message) => hasImageContent(message.content));
126
+ const VISION_TILE_SIZE = 512;
127
+ const VISION_LOW_DETAIL_TOKENS = 85;
128
+ const VISION_TILE_TOKENS = 170;
129
+ const VISION_HTTP_URL_ESTIMATE = 1024;
130
+ const FILE_BASE64_FALLBACK_TOKENS = 1024;
131
+ function estimateVisionTokensForDimensions(width, height) {
132
+ if (!Number.isFinite(width) || !Number.isFinite(height) || width <= 0 || height <= 0) {
133
+ return VISION_LOW_DETAIL_TOKENS;
134
+ }
135
+ const tilesX = Math.ceil(width / VISION_TILE_SIZE);
136
+ const tilesY = Math.ceil(height / VISION_TILE_SIZE);
137
+ return VISION_LOW_DETAIL_TOKENS + tilesX * tilesY * VISION_TILE_TOKENS;
138
+ }
139
+ function decodeBase64DataUrl(url) {
140
+ const match = /^data:([^;]+);base64,([A-Za-z0-9+/]+=?=?)$/.exec(url);
141
+ if (!match) return void 0;
142
+ try {
143
+ const buffer = Buffer.from(match[2], "base64");
144
+ return { mimeType: match[1].toLowerCase(), buffer };
145
+ } catch {
146
+ return void 0;
147
+ }
148
+ }
149
+ function estimateImageUrlTokens(imageUrl) {
150
+ const url = typeof imageUrl === "string" ? imageUrl : isRecord(imageUrl) ? String(imageUrl.url ?? "") : "";
151
+ if (!url) return 0;
152
+ if (url.startsWith("data:")) {
153
+ const decoded = decodeBase64DataUrl(url);
154
+ if (!decoded) return FILE_BASE64_FALLBACK_TOKENS;
155
+ if (!decoded.mimeType.startsWith("image/")) return FILE_BASE64_FALLBACK_TOKENS;
156
+ const dimensions = parseImageDimensions(decoded.buffer);
157
+ return dimensions ? estimateVisionTokensForDimensions(dimensions.width, dimensions.height) : VISION_LOW_DETAIL_TOKENS;
158
+ }
159
+ if (url.startsWith("http://") || url.startsWith("https://")) {
160
+ return VISION_HTTP_URL_ESTIMATE;
161
+ }
162
+ return 0;
163
+ }
164
+ function estimateFileBlockTokens(block) {
165
+ const file = isRecord(block.file) ? block.file : void 0;
166
+ if (!file) return 0;
167
+ const fileData = String(file.file_data ?? "");
168
+ if (fileData.startsWith("data:")) {
169
+ const decoded = decodeBase64DataUrl(fileData);
170
+ if (decoded) {
171
+ return Math.max(1, Math.ceil(decoded.buffer.length / 3));
172
+ }
173
+ return FILE_BASE64_FALLBACK_TOKENS;
174
+ }
175
+ return FILE_BASE64_FALLBACK_TOKENS;
176
+ }
177
+ function estimateContentBlockTokens(block) {
178
+ if (!isRecord(block)) return estimateValueTokens(block);
179
+ const type = typeof block.type === "string" ? block.type : void 0;
180
+ if (type === "text") {
181
+ return typeof block.text === "string" ? estimateValueTokens(block.text) + 4 : 4;
182
+ }
183
+ if (type === "image_url") {
184
+ return estimateImageUrlTokens(block.image_url) + 4;
185
+ }
186
+ if (type === "file") {
187
+ return estimateFileBlockTokens(block) + 4;
188
+ }
189
+ if (type === "file_url") {
190
+ return VISION_HTTP_URL_ESTIMATE + 4;
191
+ }
192
+ return estimateValueTokens(block) + 4;
193
+ }
194
+ function estimateMessageTokens(message) {
195
+ if (Array.isArray(message.content)) {
196
+ return message.content.reduce((total, block) => total + estimateContentBlockTokens(block), 4);
197
+ }
198
+ return estimateValueTokens(message) + 4;
199
+ }
200
+ function estimateMessagesTokens(messages) {
201
+ return messages.reduce((total, message) => total + estimateMessageTokens(message), 0);
62
202
  }
63
203
  function isInstruction(message) {
64
204
  return message.role === "system" || message.role === "developer";
@@ -119,12 +259,6 @@ function resolveReservedOutputTokens(options, metadata) {
119
259
  return requested ?? metadata.maxCompletionTokens;
120
260
  }
121
261
  async function prepareDirectLlmContext(ctx, options) {
122
- if (containsUnsupportedContent(options.messages)) {
123
- throw new DirectLlmContextError(
124
- "context_estimation_unsupported",
125
- "Context enforcement does not support image_url content without a model-specific vision token estimator."
126
- );
127
- }
128
262
  const [metadata, behavior] = await Promise.all([
129
263
  loadModelMetadata(ctx, options.serviceName, options.modelId),
130
264
  resolveOverflowBehavior(ctx)
@@ -180,5 +314,6 @@ async function prepareDirectLlmContext(ctx, options) {
180
314
  // Annotate the CommonJS export names for ESM import in node:
181
315
  0 && (module.exports = {
182
316
  DirectLlmContextError,
317
+ parseImageDimensions,
183
318
  prepareDirectLlmContext
184
319
  });
@@ -99,7 +99,19 @@ function toOpenAIResponse(options) {
99
99
  finish_reason: finishReason
100
100
  }
101
101
  ],
102
- usage: usage || { prompt_tokens: null, completion_tokens: null, total_tokens: null }
102
+ usage: usage ? {
103
+ prompt_tokens: usage.prompt_tokens ?? null,
104
+ completion_tokens: usage.completion_tokens ?? null,
105
+ total_tokens: usage.total_tokens ?? null,
106
+ prompt_tokens_details: {
107
+ cached_tokens: usage.prompt_cache_tokens ?? null
108
+ }
109
+ } : {
110
+ prompt_tokens: null,
111
+ completion_tokens: null,
112
+ total_tokens: null,
113
+ prompt_tokens_details: { cached_tokens: null }
114
+ }
103
115
  };
104
116
  }
105
117
  function toOpenAIStreamChunk(options) {
@@ -129,7 +141,14 @@ function toOpenAIUsageChunk(options) {
129
141
  created: Math.floor(Date.now() / 1e3),
130
142
  model,
131
143
  choices: [],
132
- usage
144
+ usage: {
145
+ prompt_tokens: usage.prompt_tokens,
146
+ completion_tokens: usage.completion_tokens,
147
+ total_tokens: usage.total_tokens,
148
+ prompt_tokens_details: {
149
+ cached_tokens: usage.prompt_cache_tokens ?? null
150
+ }
151
+ }
133
152
  };
134
153
  }
135
154
  function toOpenAIEmbeddingsResponse(options) {