File size: 16,960 Bytes
077865a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
import type {
  ChatMessage,
  ChatCompletionResponse,
  ChatCompletionChunk,
  ChatToolCall,
  ChatToolChoice,
  ChatToolDefinition,
  TokenUsage,
} from '@freellmapi/shared/types.js';
import { BaseProvider, providerHttpError, type CompletionOptions } from './base.js';
import { contentToString } from '../lib/content.js';

const API_BASE = 'https://generativelanguage.googleapis.com/v1beta';

interface GeminiPart {
  text?: string;
  inlineData?: {
    mimeType: string;
    data: string;
  };
  thoughtSignature?: string;
  functionCall?: {
    id?: string;
    name?: string;
    args?: unknown;
  };
  functionResponse?: {
    id?: string;
    name?: string;
    response?: unknown;
  };
}

interface GeminiCandidate {
  content?: { parts?: GeminiPart[] };
  finishReason?: string;
}

interface GeminiResponse {
  candidates?: GeminiCandidate[];
  usageMetadata?: {
    promptTokenCount?: number;
    candidatesTokenCount?: number;
    totalTokenCount?: number;
  };
}

function safeParseObject(raw: string): Record<string, unknown> {
  try {
    const parsed = JSON.parse(raw) as unknown;
    if (parsed && typeof parsed === 'object' && !Array.isArray(parsed)) {
      return parsed as Record<string, unknown>;
    }
    return { value: parsed };
  } catch {
    return { value: raw };
  }
}

function normalizeGeminiArgs(args: unknown): string {
  if (typeof args === 'string') return args;
  return JSON.stringify(args ?? {});
}

function toGeminiFinishReason(finishReason?: string): string {
  const r = (finishReason ?? '').toUpperCase();
  if (!r) return 'stop';
  if (r === 'MAX_TOKENS') return 'length';
  if (r === 'SAFETY' || r === 'RECITATION' || r === 'BLOCKLIST' || r === 'PROHIBITED_CONTENT' || r === 'SPII') {
    return 'content_filter';
  }
  return 'stop';
}

// Google Gemini accepts only a subset of JSON Schema (~OpenAPI 3.0).
// Strip fields that opencode / other strict-JSON-Schema clients send but
// Google rejects with 400 "Unknown name '<field>'".
const GEMINI_UNSUPPORTED_SCHEMA_KEYS = new Set([
  '$schema', '$id', '$ref', '$defs', '$comment',
  'definitions',
  'exclusiveMinimum', 'exclusiveMaximum',
  'patternProperties', 'unevaluatedProperties', 'unevaluatedItems',
  'if', 'then', 'else',
  'contentEncoding', 'contentMediaType', 'contentSchema',
  'dependentRequired', 'dependentSchemas', 'dependencies',
  'additionalProperties',
  'examples', 'const', 'readOnly', 'writeOnly',
  'uniqueItems',
  'not', 'allOf', 'oneOf',
  'prefixItems',
  'contains', 'minContains', 'maxContains',
  'propertyNames',
  'multipleOf',
  'deprecated',
]);

export function sanitizeForGemini(schema: unknown): unknown {
  if (Array.isArray(schema)) {
    return schema.map(sanitizeForGemini);
  }
  if (schema && typeof schema === 'object') {
    const out: Record<string, unknown> = {};
    for (const [k, v] of Object.entries(schema as Record<string, unknown>)) {
      if (GEMINI_UNSUPPORTED_SCHEMA_KEYS.has(k)) continue;
      out[k] = sanitizeForGemini(v);
    }
    return out;
  }
  return schema;
}

function toGeminiTools(tools?: ChatToolDefinition[]): Array<{ functionDeclarations: Array<Record<string, unknown>> }> | undefined {
  if (!tools || tools.length === 0) return undefined;

  return [{
    functionDeclarations: tools.map(t => ({
      name: t.function.name,
      description: t.function.description,
      parameters: sanitizeForGemini(t.function.parameters),
    })),
  }];
}

function toGeminiToolConfig(toolChoice?: ChatToolChoice): { functionCallingConfig: Record<string, unknown> } | undefined {
  if (!toolChoice) return undefined;

  if (typeof toolChoice === 'string') {
    const mode =
      toolChoice === 'none'
        ? 'NONE'
        : toolChoice === 'required'
          ? 'ANY'
          : 'AUTO';
    return { functionCallingConfig: { mode } };
  }

  return {
    functionCallingConfig: {
      mode: 'ANY',
      allowedFunctionNames: [toolChoice.function.name],
    },
  };
}

const MAX_IMAGE_BYTES = 8 * 1024 * 1024; // 8 MB cap on fetched/inlined images

// Pull the URL out of an OpenAI image content block. Accepts both the object
// form `{ image_url: { url } }` and the shorthand `{ image_url: '...' }`.
function extractImageUrl(block: unknown): string | undefined {
  const iu = (block as { image_url?: unknown })?.image_url;
  if (typeof iu === 'string') return iu;
  if (iu && typeof (iu as { url?: unknown }).url === 'string') return (iu as { url: string }).url;
  return undefined;
}

// Convert an image URL to a Gemini inlineData part. Handles base64 `data:` URLs
// directly; for `http(s)` URLs we fetch and inline because the Gemini API does
// not fetch external URLs itself. Fetching a user-supplied URL is a minor SSRF
// surface, acceptable for a single-user self-hosted proxy; we still restrict to
// http/https and cap the size. Returns null (part skipped) on any failure.
async function imageUrlToInlineData(url: string): Promise<{ mimeType: string; data: string } | null> {
  const dataMatch = /^data:([^;,]+)?(;base64)?,(.*)$/s.exec(url);
  if (dataMatch) {
    const mimeType = dataMatch[1] || 'application/octet-stream';
    const isBase64 = Boolean(dataMatch[2]);
    const payload = dataMatch[3] ?? '';
    const data = isBase64
      ? payload
      : Buffer.from(decodeURIComponent(payload)).toString('base64');
    return { mimeType, data };
  }
  if (/^https?:\/\//i.test(url)) {
    try {
      const res = await fetch(url);
      if (!res.ok) return null;
      const buf = Buffer.from(await res.arrayBuffer());
      if (buf.length === 0 || buf.length > MAX_IMAGE_BYTES) return null;
      const mimeType = res.headers.get('content-type')?.split(';')[0]?.trim() || 'image/jpeg';
      return { mimeType, data: buf.toString('base64') };
    } catch {
      return null;
    }
  }
  return null;
}

// Build Gemini parts for a user message: joined text first, then any images as
// inlineData. Non-array content (string/null) collapses to a single text part.
async function userContentToParts(content: ChatMessage['content']): Promise<GeminiPart[]> {
  const parts: GeminiPart[] = [];
  const text = contentToString(content);
  if (text.length > 0) parts.push({ text });

  if (Array.isArray(content)) {
    for (const block of content) {
      const type = (block as { type?: string })?.type;
      if (type !== 'image_url' && type !== 'image') continue;
      const url = extractImageUrl(block);
      if (!url) continue;
      const inlineData = await imageUrlToInlineData(url);
      if (inlineData) parts.push({ inlineData });
    }
  }

  // Gemini rejects empty `parts`; keep at least one (possibly empty) text part.
  if (parts.length === 0) parts.push({ text: '' });
  return parts;
}

// Translate OpenAI messages to Gemini format. Content may arrive as a string,
// null, or the OpenAI multimodal array envelope. System/assistant/tool messages
// flatten to text; user messages additionally carry images as inlineData parts.
async function toGeminiContents(messages: ChatMessage[]) {
  const systemMessages = messages
    .filter(m => m.role === 'system')
    .map(m => contentToString(m.content))
    .filter(s => s.length > 0);

  const toolNameByCallId = new Map<string, string>();
  for (const m of messages) {
    for (const tc of m.tool_calls ?? []) {
      toolNameByCallId.set(tc.id, tc.function.name);
    }
  }

  const contents = (await Promise.all(messages
    .filter(m => m.role !== 'system')
    .map(async (m): Promise<{ role: 'user' | 'model'; parts: GeminiPart[] } | null> => {
      if (m.role === 'assistant') {
        const parts: GeminiPart[] = [];

        const assistantText = contentToString(m.content);
        if (assistantText.length > 0) {
          parts.push({ text: assistantText });
        }

        for (const call of m.tool_calls ?? []) {
          parts.push({
            thoughtSignature: call.thought_signature,
            functionCall: {
              id: call.id,
              name: call.function.name,
              args: safeParseObject(call.function.arguments),
            },
          });
        }

        if (parts.length === 0) return null;
        return {
          role: 'model',
          parts,
        };
      }

      if (m.role === 'tool') {
        const toolCallId = m.tool_call_id;
        if (!toolCallId) return null;

        const toolName = m.name ?? toolNameByCallId.get(toolCallId) ?? 'tool';
        const response = safeParseObject(contentToString(m.content));

        return {
          role: 'user',
          parts: [{
            functionResponse: {
              id: toolCallId,
              name: toolName,
              response,
            },
          }],
        };
      }

      return {
        role: 'user',
        parts: await userContentToParts(m.content),
      };
    })))
    .filter((entry): entry is { role: 'user' | 'model'; parts: GeminiPart[] } => entry !== null);

  return {
    contents,
    systemInstruction: systemMessages.length > 0
      ? { parts: [{ text: systemMessages.join('\n\n') }] }
      : undefined,
  };
}

function extractToolCalls(parts: GeminiPart[] | undefined): ChatToolCall[] {
  const calls: ChatToolCall[] = [];
  if (!parts) return calls;

  let fallbackIndex = 0;
  for (const part of parts) {
    if (!part.functionCall?.name) continue;

    const id = part.functionCall.id ?? `call_${Date.now()}_${fallbackIndex++}`;
    calls.push({
      id,
      type: 'function',
      function: {
        name: part.functionCall.name,
        arguments: normalizeGeminiArgs(part.functionCall.args),
      },
      thought_signature: part.thoughtSignature,
    });
  }

  return calls;
}

function extractText(parts: GeminiPart[] | undefined): string | null {
  if (!parts) return null;
  const text = parts
    .map(p => p.text ?? '')
    .join('');
  return text.length > 0 ? text : null;
}

export class GoogleProvider extends BaseProvider {
  readonly platform = 'google' as const;
  readonly name = 'Google AI Studio';

  async chatCompletion(
    apiKey: string,
    messages: ChatMessage[],
    modelId: string,
    options?: CompletionOptions,
  ): Promise<ChatCompletionResponse> {
    const { contents, systemInstruction } = await toGeminiContents(messages);

    const body: Record<string, unknown> = {
      contents,
      generationConfig: {
        temperature: options?.temperature,
        maxOutputTokens: options?.max_tokens,
        topP: options?.top_p,
      },
      tools: toGeminiTools(options?.tools),
      toolConfig: toGeminiToolConfig(options?.tool_choice),
    };
    if (systemInstruction) body.systemInstruction = systemInstruction;

    const url = `${API_BASE}/models/${modelId}:generateContent?key=${apiKey}`;
    const res = await this.fetchWithTimeout(url, {
      method: 'POST',
      headers: { 'Content-Type': 'application/json' },
      body: JSON.stringify(body),
    });

    if (!res.ok) {
      const err = await res.json().catch(() => ({}));
      throw providerHttpError(res, `Google API error ${res.status}: ${(err as any).error?.message ?? res.statusText}`);
    }

    const data = await res.json() as GeminiResponse;
    const candidate = data.candidates?.[0];
    const parts = candidate?.content?.parts;
    const toolCalls = extractToolCalls(parts);
    const text = extractText(parts);

    const usage: TokenUsage = {
      prompt_tokens: data.usageMetadata?.promptTokenCount ?? 0,
      completion_tokens: data.usageMetadata?.candidatesTokenCount ?? 0,
      total_tokens: data.usageMetadata?.totalTokenCount ?? 0,
    };

    return {
      id: this.makeId(),
      object: 'chat.completion',
      created: Math.floor(Date.now() / 1000),
      model: modelId,
      choices: [{
        index: 0,
        message: {
          role: 'assistant',
          content: text,
          ...(toolCalls.length > 0 ? { tool_calls: toolCalls } : {}),
        },
        finish_reason: toolCalls.length > 0 ? 'tool_calls' : toGeminiFinishReason(candidate?.finishReason),
      }],
      usage,
      _routed_via: { platform: 'google', model: modelId },
    };
  }

  async *streamChatCompletion(
    apiKey: string,
    messages: ChatMessage[],
    modelId: string,
    options?: CompletionOptions,
  ): AsyncGenerator<ChatCompletionChunk> {
    const { contents, systemInstruction } = await toGeminiContents(messages);

    const body: Record<string, unknown> = {
      contents,
      generationConfig: {
        temperature: options?.temperature,
        maxOutputTokens: options?.max_tokens,
        topP: options?.top_p,
      },
      tools: toGeminiTools(options?.tools),
      toolConfig: toGeminiToolConfig(options?.tool_choice),
    };
    if (systemInstruction) body.systemInstruction = systemInstruction;

    const url = `${API_BASE}/models/${modelId}:streamGenerateContent?alt=sse&key=${apiKey}`;
    const res = await this.fetchWithTimeout(url, {
      method: 'POST',
      headers: { 'Content-Type': 'application/json' },
      body: JSON.stringify(body),
    });

    if (!res.ok) {
      const err = await res.json().catch(() => ({}));
      throw providerHttpError(res, `Google API error ${res.status}: ${(err as any).error?.message ?? res.statusText}`);
    }

    const reader = res.body?.getReader();
    if (!reader) throw new Error('No response body');

    const decoder = new TextDecoder();
    const id = this.makeId();
    let buffer = '';
    let emittedFinish = false;
    let sawToolCalls = false;

    const seenToolCallKeys = new Set<string>();

    while (true) {
      const { done, value } = await reader.read();
      if (done) break;

      buffer += decoder.decode(value, { stream: true });
      const lines = buffer.split('\n');
      buffer = lines.pop() ?? '';

      for (const line of lines) {
        const trimmed = line.trim();
        if (!trimmed || !trimmed.startsWith('data: ')) continue;
        const raw = trimmed.slice(6);
        if (raw === '[DONE]') {
          if (!emittedFinish) {
            emittedFinish = true;
            yield {
              id,
              object: 'chat.completion.chunk',
              created: Math.floor(Date.now() / 1000),
              model: modelId,
              choices: [{
                index: 0,
                delta: {},
                finish_reason: sawToolCalls ? 'tool_calls' : 'stop',
              }],
            };
          }
          return;
        }

        // Skip malformed SSE frames instead of aborting the whole stream.
        // Matches the defensive parse in openai-compat / cohere / cloudflare:
        // a single corrupt chunk shouldn't take down the rest of the response.
        let chunk: GeminiResponse;
        try {
          chunk = JSON.parse(raw) as GeminiResponse;
        } catch {
          continue;
        }
        const candidate = chunk.candidates?.[0];
        const parts = candidate?.content?.parts ?? [];

        const text = extractText(parts);
        const toolCalls = extractToolCalls(parts).filter(call => {
          const key = `${call.id}:${call.function.name}:${call.function.arguments}`;
          if (seenToolCallKeys.has(key)) return false;
          seenToolCallKeys.add(key);
          return true;
        });

        if ((text && text.length > 0) || toolCalls.length > 0) {
          sawToolCalls = sawToolCalls || toolCalls.length > 0;
          yield {
            id,
            object: 'chat.completion.chunk',
            created: Math.floor(Date.now() / 1000),
            model: modelId,
            choices: [{
              index: 0,
              delta: {
                ...(text ? { content: text } : {}),
                ...(toolCalls.length > 0 ? { tool_calls: toolCalls } : {}),
              },
              finish_reason: null,
            }],
          };
        }

        if (candidate?.finishReason && !emittedFinish) {
          emittedFinish = true;
          yield {
            id,
            object: 'chat.completion.chunk',
            created: Math.floor(Date.now() / 1000),
            model: modelId,
            choices: [{
              index: 0,
              delta: {},
              finish_reason: sawToolCalls ? 'tool_calls' : toGeminiFinishReason(candidate.finishReason),
            }],
          };
          return;
        }
      }
    }

    if (!emittedFinish) {
      yield {
        id,
        object: 'chat.completion.chunk',
        created: Math.floor(Date.now() / 1000),
        model: modelId,
        choices: [{
          index: 0,
          delta: {},
          finish_reason: sawToolCalls ? 'tool_calls' : 'stop',
        }],
      };
    }
  }

  async validateKey(apiKey: string): Promise<boolean> {
    // Transport errors propagate — health.ts marks status='error' without
    // counting toward auto-disable. Only confirmed 401/403 disables a key.
    const res = await this.fetchWithTimeout(
      `${API_BASE}/models?key=${apiKey}`,
      { method: 'GET' },
      10000,
    );
    return res.status !== 401 && res.status !== 403;
  }
}