|
1
|
"""Google Vertex AI provider implementation.""" |
|
2
|
|
|
3
|
import logging |
|
4
|
import os |
|
5
|
from pathlib import Path |
|
6
|
from typing import Optional |
|
7
|
|
|
8
|
from dotenv import load_dotenv |
|
9
|
|
|
10
|
from video_processor.providers.base import BaseProvider, ModelInfo, ProviderRegistry |
|
11
|
|
|
12
|
load_dotenv() |
|
13
|
logger = logging.getLogger(__name__) |
|
14
|
|
|
15
|
# Curated list of models available on Vertex AI |
|
16
|
_VERTEX_MODELS = [ |
|
17
|
ModelInfo( |
|
18
|
id="gemini-2.0-flash", |
|
19
|
provider="vertex", |
|
20
|
display_name="Gemini 2.0 Flash", |
|
21
|
capabilities=["chat", "vision", "audio"], |
|
22
|
), |
|
23
|
ModelInfo( |
|
24
|
id="gemini-2.0-pro", |
|
25
|
provider="vertex", |
|
26
|
display_name="Gemini 2.0 Pro", |
|
27
|
capabilities=["chat", "vision", "audio"], |
|
28
|
), |
|
29
|
ModelInfo( |
|
30
|
id="gemini-1.5-pro", |
|
31
|
provider="vertex", |
|
32
|
display_name="Gemini 1.5 Pro", |
|
33
|
capabilities=["chat", "vision", "audio"], |
|
34
|
), |
|
35
|
ModelInfo( |
|
36
|
id="gemini-1.5-flash", |
|
37
|
provider="vertex", |
|
38
|
display_name="Gemini 1.5 Flash", |
|
39
|
capabilities=["chat", "vision", "audio"], |
|
40
|
), |
|
41
|
] |
|
42
|
|
|
43
|
|
|
44
|
class VertexProvider(BaseProvider): |
|
45
|
"""Google Vertex AI provider using google-genai SDK with Vertex config.""" |
|
46
|
|
|
47
|
provider_name = "vertex" |
|
48
|
|
|
49
|
def __init__( |
|
50
|
self, |
|
51
|
project: Optional[str] = None, |
|
52
|
location: Optional[str] = None, |
|
53
|
): |
|
54
|
try: |
|
55
|
from google import genai |
|
56
|
from google.genai import types # noqa: F401 |
|
57
|
except ImportError: |
|
58
|
raise ImportError( |
|
59
|
"google-cloud-aiplatform or google-genai package not installed. " |
|
60
|
"Install with: pip install google-cloud-aiplatform" |
|
61
|
) |
|
62
|
|
|
63
|
self._genai = genai |
|
64
|
self._project = project or os.getenv("GOOGLE_CLOUD_PROJECT") |
|
65
|
self._location = location or os.getenv("GOOGLE_CLOUD_REGION", "us-central1") |
|
66
|
|
|
67
|
if not self._project: |
|
68
|
raise ValueError("GOOGLE_CLOUD_PROJECT not set") |
|
69
|
|
|
70
|
self.client = genai.Client( |
|
71
|
vertexai=True, |
|
72
|
project=self._project, |
|
73
|
location=self._location, |
|
74
|
) |
|
75
|
self._last_usage = {} |
|
76
|
|
|
77
|
def chat( |
|
78
|
self, |
|
79
|
messages: list[dict], |
|
80
|
max_tokens: int = 4096, |
|
81
|
temperature: float = 0.7, |
|
82
|
model: Optional[str] = None, |
|
83
|
) -> str: |
|
84
|
from google.genai import types |
|
85
|
|
|
86
|
model = model or "gemini-2.0-flash" |
|
87
|
if model.startswith("vertex/"): |
|
88
|
model = model[len("vertex/") :] |
|
89
|
|
|
90
|
contents = [] |
|
91
|
for msg in messages: |
|
92
|
role = "user" if msg["role"] == "user" else "model" |
|
93
|
contents.append( |
|
94
|
types.Content( |
|
95
|
role=role, |
|
96
|
parts=[types.Part.from_text(text=msg["content"])], |
|
97
|
) |
|
98
|
) |
|
99
|
|
|
100
|
response = self.client.models.generate_content( |
|
101
|
model=model, |
|
102
|
contents=contents, |
|
103
|
config=types.GenerateContentConfig( |
|
104
|
max_output_tokens=max_tokens, |
|
105
|
temperature=temperature, |
|
106
|
), |
|
107
|
) |
|
108
|
um = getattr(response, "usage_metadata", None) |
|
109
|
self._last_usage = { |
|
110
|
"input_tokens": getattr(um, "prompt_token_count", 0) if um else 0, |
|
111
|
"output_tokens": getattr(um, "candidates_token_count", 0) if um else 0, |
|
112
|
} |
|
113
|
return response.text or "" |
|
114
|
|
|
115
|
def analyze_image( |
|
116
|
self, |
|
117
|
image_bytes: bytes, |
|
118
|
prompt: str, |
|
119
|
max_tokens: int = 4096, |
|
120
|
model: Optional[str] = None, |
|
121
|
) -> str: |
|
122
|
from google.genai import types |
|
123
|
|
|
124
|
model = model or "gemini-2.0-flash" |
|
125
|
if model.startswith("vertex/"): |
|
126
|
model = model[len("vertex/") :] |
|
127
|
|
|
128
|
response = self.client.models.generate_content( |
|
129
|
model=model, |
|
130
|
contents=[ |
|
131
|
types.Part.from_bytes(data=image_bytes, mime_type="image/jpeg"), |
|
132
|
prompt, |
|
133
|
], |
|
134
|
config=types.GenerateContentConfig( |
|
135
|
max_output_tokens=max_tokens, |
|
136
|
), |
|
137
|
) |
|
138
|
um = getattr(response, "usage_metadata", None) |
|
139
|
self._last_usage = { |
|
140
|
"input_tokens": getattr(um, "prompt_token_count", 0) if um else 0, |
|
141
|
"output_tokens": getattr(um, "candidates_token_count", 0) if um else 0, |
|
142
|
} |
|
143
|
return response.text or "" |
|
144
|
|
|
145
|
def transcribe_audio( |
|
146
|
self, |
|
147
|
audio_path: str | Path, |
|
148
|
language: Optional[str] = None, |
|
149
|
model: Optional[str] = None, |
|
150
|
) -> dict: |
|
151
|
import json |
|
152
|
|
|
153
|
from google.genai import types |
|
154
|
|
|
155
|
model = model or "gemini-2.0-flash" |
|
156
|
if model.startswith("vertex/"): |
|
157
|
model = model[len("vertex/") :] |
|
158
|
|
|
159
|
audio_path = Path(audio_path) |
|
160
|
suffix = audio_path.suffix.lower() |
|
161
|
mime_map = { |
|
162
|
".wav": "audio/wav", |
|
163
|
".mp3": "audio/mpeg", |
|
164
|
".m4a": "audio/mp4", |
|
165
|
".flac": "audio/flac", |
|
166
|
".ogg": "audio/ogg", |
|
167
|
".webm": "audio/webm", |
|
168
|
} |
|
169
|
mime_type = mime_map.get(suffix, "audio/wav") |
|
170
|
audio_bytes = audio_path.read_bytes() |
|
171
|
|
|
172
|
lang_hint = f" The audio is in {language}." if language else "" |
|
173
|
prompt = ( |
|
174
|
f"Transcribe this audio accurately.{lang_hint} " |
|
175
|
"Return a JSON object with keys: " |
|
176
|
'"text" (full transcript), ' |
|
177
|
'"segments" (array of {start, end, text} objects with timestamps in seconds).' |
|
178
|
) |
|
179
|
|
|
180
|
response = self.client.models.generate_content( |
|
181
|
model=model, |
|
182
|
contents=[ |
|
183
|
types.Part.from_bytes(data=audio_bytes, mime_type=mime_type), |
|
184
|
prompt, |
|
185
|
], |
|
186
|
config=types.GenerateContentConfig( |
|
187
|
max_output_tokens=8192, |
|
188
|
response_mime_type="application/json", |
|
189
|
), |
|
190
|
) |
|
191
|
|
|
192
|
try: |
|
193
|
data = json.loads(response.text) |
|
194
|
except (json.JSONDecodeError, TypeError): |
|
195
|
data = {"text": response.text or "", "segments": []} |
|
196
|
|
|
197
|
um = getattr(response, "usage_metadata", None) |
|
198
|
self._last_usage = { |
|
199
|
"input_tokens": getattr(um, "prompt_token_count", 0) if um else 0, |
|
200
|
"output_tokens": getattr(um, "candidates_token_count", 0) if um else 0, |
|
201
|
} |
|
202
|
|
|
203
|
return { |
|
204
|
"text": data.get("text", ""), |
|
205
|
"segments": data.get("segments", []), |
|
206
|
"language": language, |
|
207
|
"duration": None, |
|
208
|
"provider": "vertex", |
|
209
|
"model": model, |
|
210
|
} |
|
211
|
|
|
212
|
def list_models(self) -> list[ModelInfo]: |
|
213
|
return list(_VERTEX_MODELS) |
|
214
|
|
|
215
|
|
|
216
|
ProviderRegistry.register( |
|
217
|
name="vertex", |
|
218
|
provider_class=VertexProvider, |
|
219
|
env_var="GOOGLE_CLOUD_PROJECT", |
|
220
|
model_prefixes=["vertex/"], |
|
221
|
default_models={ |
|
222
|
"chat": "gemini-2.0-flash", |
|
223
|
"vision": "gemini-2.0-flash", |
|
224
|
"audio": "gemini-2.0-flash", |
|
225
|
}, |
|
226
|
) |
|
227
|
|