|
1
|
"""Ollama provider implementation using OpenAI-compatible API.""" |
|
2
|
|
|
3
|
import base64 |
|
4
|
import logging |
|
5
|
import os |
|
6
|
from pathlib import Path |
|
7
|
from typing import Optional |
|
8
|
|
|
9
|
import requests |
|
10
|
from openai import OpenAI |
|
11
|
|
|
12
|
from video_processor.providers.base import BaseProvider, ModelInfo, ProviderRegistry |
|
13
|
|
|
14
|
logger = logging.getLogger(__name__) |
|
15
|
|
|
16
|
# Known vision-capable model families (base name before the colon/tag) |
|
17
|
_VISION_FAMILIES = { |
|
18
|
"llava", |
|
19
|
"llava-llama3", |
|
20
|
"llava-phi3", |
|
21
|
"llama3.2-vision", |
|
22
|
"moondream", |
|
23
|
"bakllava", |
|
24
|
"minicpm-v", |
|
25
|
"deepseek-vl", |
|
26
|
"internvl2", |
|
27
|
} |
|
28
|
|
|
29
|
DEFAULT_HOST = "http://localhost:11434" |
|
30
|
|
|
31
|
|
|
32
|
class OllamaProvider(BaseProvider): |
|
33
|
"""Ollama local LLM provider via OpenAI-compatible API.""" |
|
34
|
|
|
35
|
provider_name = "ollama" |
|
36
|
|
|
37
|
def __init__(self, host: Optional[str] = None): |
|
38
|
self.host = host or os.getenv("OLLAMA_HOST", DEFAULT_HOST) |
|
39
|
self.client = OpenAI( |
|
40
|
base_url=f"{self.host}/v1", |
|
41
|
api_key="ollama", |
|
42
|
) |
|
43
|
self._models_cache: Optional[list[ModelInfo]] = None |
|
44
|
|
|
45
|
@staticmethod |
|
46
|
def is_available(host: Optional[str] = None) -> bool: |
|
47
|
"""Check if an Ollama server is running and reachable.""" |
|
48
|
host = host or os.getenv("OLLAMA_HOST", DEFAULT_HOST) |
|
49
|
try: |
|
50
|
resp = requests.get(f"{host}/api/tags", timeout=3) |
|
51
|
return resp.status_code == 200 |
|
52
|
except Exception: |
|
53
|
return False |
|
54
|
|
|
55
|
@property |
|
56
|
def _default_model(self) -> str: |
|
57
|
models = self._get_models() |
|
58
|
for m in models: |
|
59
|
if "chat" in m.capabilities: |
|
60
|
return m.id |
|
61
|
return "llama3.2:latest" |
|
62
|
|
|
63
|
@property |
|
64
|
def _default_vision_model(self) -> Optional[str]: |
|
65
|
models = self._get_models() |
|
66
|
for m in models: |
|
67
|
if "vision" in m.capabilities: |
|
68
|
return m.id |
|
69
|
return None |
|
70
|
|
|
71
|
def _get_models(self) -> list[ModelInfo]: |
|
72
|
if self._models_cache is None: |
|
73
|
self._models_cache = self.list_models() |
|
74
|
return self._models_cache |
|
75
|
|
|
76
|
def chat( |
|
77
|
self, |
|
78
|
messages: list[dict], |
|
79
|
max_tokens: int = 4096, |
|
80
|
temperature: float = 0.7, |
|
81
|
model: Optional[str] = None, |
|
82
|
) -> str: |
|
83
|
model = model or self._default_model |
|
84
|
response = self.client.chat.completions.create( |
|
85
|
model=model, |
|
86
|
messages=messages, |
|
87
|
max_tokens=max_tokens, |
|
88
|
temperature=temperature, |
|
89
|
) |
|
90
|
self._last_usage = { |
|
91
|
"input_tokens": (getattr(response.usage, "prompt_tokens", 0) or 0) |
|
92
|
if response.usage |
|
93
|
else 0, |
|
94
|
"output_tokens": (getattr(response.usage, "completion_tokens", 0) or 0) |
|
95
|
if response.usage |
|
96
|
else 0, |
|
97
|
} |
|
98
|
return response.choices[0].message.content or "" |
|
99
|
|
|
100
|
def analyze_image( |
|
101
|
self, |
|
102
|
image_bytes: bytes, |
|
103
|
prompt: str, |
|
104
|
max_tokens: int = 4096, |
|
105
|
model: Optional[str] = None, |
|
106
|
) -> str: |
|
107
|
model = model or self._default_vision_model |
|
108
|
if not model: |
|
109
|
raise RuntimeError( |
|
110
|
"No Ollama vision model available. Install a multimodal model: ollama pull llava" |
|
111
|
) |
|
112
|
b64 = base64.b64encode(image_bytes).decode() |
|
113
|
response = self.client.chat.completions.create( |
|
114
|
model=model, |
|
115
|
messages=[ |
|
116
|
{ |
|
117
|
"role": "user", |
|
118
|
"content": [ |
|
119
|
{"type": "text", "text": prompt}, |
|
120
|
{ |
|
121
|
"type": "image_url", |
|
122
|
"image_url": {"url": f"data:image/jpeg;base64,{b64}"}, |
|
123
|
}, |
|
124
|
], |
|
125
|
} |
|
126
|
], |
|
127
|
max_tokens=max_tokens, |
|
128
|
) |
|
129
|
self._last_usage = { |
|
130
|
"input_tokens": (getattr(response.usage, "prompt_tokens", 0) or 0) |
|
131
|
if response.usage |
|
132
|
else 0, |
|
133
|
"output_tokens": (getattr(response.usage, "completion_tokens", 0) or 0) |
|
134
|
if response.usage |
|
135
|
else 0, |
|
136
|
} |
|
137
|
return response.choices[0].message.content or "" |
|
138
|
|
|
139
|
def transcribe_audio( |
|
140
|
self, |
|
141
|
audio_path: str | Path, |
|
142
|
language: Optional[str] = None, |
|
143
|
model: Optional[str] = None, |
|
144
|
) -> dict: |
|
145
|
raise NotImplementedError( |
|
146
|
"Ollama does not support audio transcription. " |
|
147
|
"Use local Whisper (--transcription-model whisper-local:large) or OpenAI Whisper API." |
|
148
|
) |
|
149
|
|
|
150
|
def list_models(self) -> list[ModelInfo]: |
|
151
|
models = [] |
|
152
|
try: |
|
153
|
resp = requests.get(f"{self.host}/api/tags", timeout=5) |
|
154
|
resp.raise_for_status() |
|
155
|
data = resp.json() |
|
156
|
for m in data.get("models", []): |
|
157
|
name = m.get("name", "") |
|
158
|
caps = ["chat"] |
|
159
|
base_name = name.split(":")[0].lower() |
|
160
|
if base_name in _VISION_FAMILIES or "vision" in base_name: |
|
161
|
caps.append("vision") |
|
162
|
models.append( |
|
163
|
ModelInfo( |
|
164
|
id=name, |
|
165
|
provider="ollama", |
|
166
|
display_name=name, |
|
167
|
capabilities=caps, |
|
168
|
) |
|
169
|
) |
|
170
|
except Exception as e: |
|
171
|
logger.warning(f"Failed to list Ollama models: {e}") |
|
172
|
return sorted(models, key=lambda m: m.id) |
|
173
|
|
|
174
|
|
|
175
|
ProviderRegistry.register( |
|
176
|
name="ollama", |
|
177
|
provider_class=OllamaProvider, |
|
178
|
env_var="", |
|
179
|
model_prefixes=[], |
|
180
|
default_models={"chat": "", "vision": "", "audio": ""}, |
|
181
|
) |
|
182
|
|