|
1
|
"""OpenAI provider implementation.""" |
|
2
|
|
|
3
|
import base64 |
|
4
|
import logging |
|
5
|
import os |
|
6
|
from pathlib import Path |
|
7
|
from typing import Optional |
|
8
|
|
|
9
|
from dotenv import load_dotenv |
|
10
|
from openai import OpenAI |
|
11
|
|
|
12
|
from video_processor.providers.base import BaseProvider, ModelInfo, ProviderRegistry |
|
13
|
|
|
14
|
load_dotenv() |
|
15
|
logger = logging.getLogger(__name__) |
|
16
|
|
|
17
|
# Models known to have vision capability |
|
18
|
_VISION_MODELS = { |
|
19
|
"gpt-4o", |
|
20
|
"gpt-4o-mini", |
|
21
|
"gpt-4-turbo", |
|
22
|
"gpt-4.1", |
|
23
|
"gpt-4.1-mini", |
|
24
|
"gpt-4.1-nano", |
|
25
|
"o1", |
|
26
|
"o3", |
|
27
|
"o3-mini", |
|
28
|
"o4-mini", |
|
29
|
} |
|
30
|
_AUDIO_MODELS = {"whisper-1"} |
|
31
|
|
|
32
|
|
|
33
|
class OpenAIProvider(BaseProvider): |
|
34
|
"""OpenAI API provider.""" |
|
35
|
|
|
36
|
provider_name = "openai" |
|
37
|
|
|
38
|
def __init__(self, api_key: Optional[str] = None): |
|
39
|
self.api_key = api_key or os.getenv("OPENAI_API_KEY") |
|
40
|
if not self.api_key: |
|
41
|
raise ValueError("OPENAI_API_KEY not set") |
|
42
|
self.client = OpenAI(api_key=self.api_key) |
|
43
|
|
|
44
|
def chat( |
|
45
|
self, |
|
46
|
messages: list[dict], |
|
47
|
max_tokens: int = 4096, |
|
48
|
temperature: float = 0.7, |
|
49
|
model: Optional[str] = None, |
|
50
|
) -> str: |
|
51
|
model = model or "gpt-4o-mini" |
|
52
|
response = self.client.chat.completions.create( |
|
53
|
model=model, |
|
54
|
messages=messages, |
|
55
|
max_tokens=max_tokens, |
|
56
|
temperature=temperature, |
|
57
|
) |
|
58
|
self._last_usage = { |
|
59
|
"input_tokens": getattr(response.usage, "prompt_tokens", 0) if response.usage else 0, |
|
60
|
"output_tokens": getattr(response.usage, "completion_tokens", 0) |
|
61
|
if response.usage |
|
62
|
else 0, |
|
63
|
} |
|
64
|
return response.choices[0].message.content or "" |
|
65
|
|
|
66
|
def analyze_image( |
|
67
|
self, |
|
68
|
image_bytes: bytes, |
|
69
|
prompt: str, |
|
70
|
max_tokens: int = 4096, |
|
71
|
model: Optional[str] = None, |
|
72
|
) -> str: |
|
73
|
model = model or "gpt-4o-mini" |
|
74
|
b64 = base64.b64encode(image_bytes).decode() |
|
75
|
response = self.client.chat.completions.create( |
|
76
|
model=model, |
|
77
|
messages=[ |
|
78
|
{ |
|
79
|
"role": "user", |
|
80
|
"content": [ |
|
81
|
{"type": "text", "text": prompt}, |
|
82
|
{ |
|
83
|
"type": "image_url", |
|
84
|
"image_url": {"url": f"data:image/jpeg;base64,{b64}"}, |
|
85
|
}, |
|
86
|
], |
|
87
|
} |
|
88
|
], |
|
89
|
max_tokens=max_tokens, |
|
90
|
) |
|
91
|
self._last_usage = { |
|
92
|
"input_tokens": getattr(response.usage, "prompt_tokens", 0) if response.usage else 0, |
|
93
|
"output_tokens": getattr(response.usage, "completion_tokens", 0) |
|
94
|
if response.usage |
|
95
|
else 0, |
|
96
|
} |
|
97
|
return response.choices[0].message.content or "" |
|
98
|
|
|
99
|
# Whisper API limit is 25MB |
|
100
|
_MAX_FILE_SIZE = 25 * 1024 * 1024 |
|
101
|
|
|
102
|
def transcribe_audio( |
|
103
|
self, |
|
104
|
audio_path: str | Path, |
|
105
|
language: Optional[str] = None, |
|
106
|
model: Optional[str] = None, |
|
107
|
) -> dict: |
|
108
|
model = model or "whisper-1" |
|
109
|
audio_path = Path(audio_path) |
|
110
|
file_size = audio_path.stat().st_size |
|
111
|
|
|
112
|
if file_size <= self._MAX_FILE_SIZE: |
|
113
|
return self._transcribe_single(audio_path, language, model) |
|
114
|
|
|
115
|
# File too large — split into chunks and transcribe each |
|
116
|
logger.info( |
|
117
|
f"Audio file {file_size / 1024 / 1024:.1f}MB exceeds Whisper 25MB limit, chunking..." |
|
118
|
) |
|
119
|
return self._transcribe_chunked(audio_path, language, model) |
|
120
|
|
|
121
|
def _transcribe_single(self, audio_path: Path, language: Optional[str], model: str) -> dict: |
|
122
|
"""Transcribe a single audio file.""" |
|
123
|
with open(audio_path, "rb") as f: |
|
124
|
kwargs = {"model": model, "file": f} |
|
125
|
if language: |
|
126
|
kwargs["language"] = language |
|
127
|
response = self.client.audio.transcriptions.create( |
|
128
|
**kwargs, response_format="verbose_json" |
|
129
|
) |
|
130
|
return { |
|
131
|
"text": response.text, |
|
132
|
"segments": [ |
|
133
|
{ |
|
134
|
"start": seg.start, |
|
135
|
"end": seg.end, |
|
136
|
"text": seg.text, |
|
137
|
} |
|
138
|
for seg in (response.segments or []) |
|
139
|
], |
|
140
|
"language": getattr(response, "language", language), |
|
141
|
"duration": getattr(response, "duration", None), |
|
142
|
"provider": "openai", |
|
143
|
"model": model, |
|
144
|
} |
|
145
|
|
|
146
|
def _transcribe_chunked(self, audio_path: Path, language: Optional[str], model: str) -> dict: |
|
147
|
"""Split audio into chunks under 25MB and transcribe each.""" |
|
148
|
import tempfile |
|
149
|
|
|
150
|
from video_processor.extractors.audio_extractor import AudioExtractor |
|
151
|
|
|
152
|
extractor = AudioExtractor() |
|
153
|
audio_data, sr = extractor.load_audio(audio_path) |
|
154
|
total_duration = len(audio_data) / sr |
|
155
|
|
|
156
|
# Calculate chunk duration to stay under 25MB |
|
157
|
# WAV: 16-bit mono = 2 bytes/sample, plus header overhead |
|
158
|
bytes_per_second = sr * 2 |
|
159
|
max_seconds = self._MAX_FILE_SIZE // bytes_per_second |
|
160
|
# Use 80% of max to leave headroom |
|
161
|
chunk_ms = int(max_seconds * 0.8 * 1000) |
|
162
|
|
|
163
|
segments_data = extractor.segment_audio(audio_data, sr, segment_length_ms=chunk_ms) |
|
164
|
logger.info(f"Split into {len(segments_data)} chunks of ~{chunk_ms / 1000:.0f}s each") |
|
165
|
|
|
166
|
all_text = [] |
|
167
|
all_segments = [] |
|
168
|
time_offset = 0.0 |
|
169
|
detected_language = language |
|
170
|
|
|
171
|
with tempfile.TemporaryDirectory() as tmpdir: |
|
172
|
for i, chunk in enumerate(segments_data): |
|
173
|
chunk_path = Path(tmpdir) / f"chunk_{i:03d}.wav" |
|
174
|
extractor.save_segment(chunk, chunk_path, sr) |
|
175
|
|
|
176
|
logger.info(f"Transcribing chunk {i + 1}/{len(segments_data)}...") |
|
177
|
result = self._transcribe_single(chunk_path, language, model) |
|
178
|
|
|
179
|
all_text.append(result["text"]) |
|
180
|
for seg in result.get("segments", []): |
|
181
|
all_segments.append( |
|
182
|
{ |
|
183
|
"start": seg["start"] + time_offset, |
|
184
|
"end": seg["end"] + time_offset, |
|
185
|
"text": seg["text"], |
|
186
|
} |
|
187
|
) |
|
188
|
|
|
189
|
if not detected_language and result.get("language"): |
|
190
|
detected_language = result["language"] |
|
191
|
|
|
192
|
time_offset += len(chunk) / sr |
|
193
|
|
|
194
|
return { |
|
195
|
"text": " ".join(all_text), |
|
196
|
"segments": all_segments, |
|
197
|
"language": detected_language, |
|
198
|
"duration": total_duration, |
|
199
|
"provider": "openai", |
|
200
|
"model": model, |
|
201
|
} |
|
202
|
|
|
203
|
def list_models(self) -> list[ModelInfo]: |
|
204
|
models = [] |
|
205
|
try: |
|
206
|
for m in self.client.models.list(): |
|
207
|
mid = m.id |
|
208
|
caps = [] |
|
209
|
# Infer capabilities from model id |
|
210
|
if any(mid.startswith(p) for p in ("gpt-", "o1", "o3", "o4")): |
|
211
|
caps.append("chat") |
|
212
|
if any(v in mid for v in _VISION_MODELS) or "gpt-4o" in mid or "gpt-4.1" in mid: |
|
213
|
caps.append("vision") |
|
214
|
if mid in _AUDIO_MODELS or mid.startswith("whisper"): |
|
215
|
caps.append("audio") |
|
216
|
if "embedding" in mid: |
|
217
|
caps.append("embedding") |
|
218
|
if caps: |
|
219
|
models.append( |
|
220
|
ModelInfo( |
|
221
|
id=mid, |
|
222
|
provider="openai", |
|
223
|
display_name=mid, |
|
224
|
capabilities=caps, |
|
225
|
) |
|
226
|
) |
|
227
|
except Exception as e: |
|
228
|
logger.warning(f"Failed to list OpenAI models: {e}") |
|
229
|
return sorted(models, key=lambda m: m.id) |
|
230
|
|
|
231
|
|
|
232
|
ProviderRegistry.register( |
|
233
|
name="openai", |
|
234
|
provider_class=OpenAIProvider, |
|
235
|
env_var="OPENAI_API_KEY", |
|
236
|
model_prefixes=["gpt-", "o1", "o3", "o4", "whisper"], |
|
237
|
default_models={"chat": "gpt-4o-mini", "vision": "gpt-4o-mini", "audio": "whisper-1"}, |
|
238
|
) |
|
239
|
|