Skip to content

LLM Providers

Silkweb supports multiple LLM backends through a unified LLMProvider interface.

Base class

LLMProvider 

LLMProvider(*, model: str, api_key: str | None = None, timeout_s: float = 30.0, retry: RetryConfig | None = None)

Bases: ABC

Source code in silkweb/llm/providers/base.py 
155
156
157
158
159
160
161
162
163
164
165
166
def __init__(
    self,
    *,
    model: str,
    api_key: str | None = None,
    timeout_s: float = 30.0,
    retry: RetryConfig | None = None,
) -> None:
    self.model = model
    self.api_key = api_key
    self.timeout_s = timeout_s
    self.retry = retry or RetryConfig()

unwrap 

unwrap() -> LLMProvider

Return the underlying concrete provider (identity for non-wrapped).

Source code in silkweb/llm/providers/base.py 
168
169
170
def unwrap(self) -> LLMProvider:
    """Return the underlying concrete provider (identity for non-wrapped)."""
    return self

Ollama

OllamaProvider 

OllamaProvider(*, model: str, api_key: str | None = None, timeout_s: float = 30.0, retry: RetryConfig | None = None)

Bases: LLMProvider

Source code in silkweb/llm/providers/base.py 
155
156
157
158
159
160
161
162
163
164
165
166
def __init__(
    self,
    *,
    model: str,
    api_key: str | None = None,
    timeout_s: float = 30.0,
    retry: RetryConfig | None = None,
) -> None:
    self.model = model
    self.api_key = api_key
    self.timeout_s = timeout_s
    self.retry = retry or RetryConfig()

OpenAI

OpenAIProvider 

OpenAIProvider(*, model: str, api_key: str | None = None, timeout_s: float = 30.0, retry: RetryConfig | None = None)

Bases: LLMProvider

Source code in silkweb/llm/providers/base.py 
155
156
157
158
159
160
161
162
163
164
165
166
def __init__(
    self,
    *,
    model: str,
    api_key: str | None = None,
    timeout_s: float = 30.0,
    retry: RetryConfig | None = None,
) -> None:
    self.model = model
    self.api_key = api_key
    self.timeout_s = timeout_s
    self.retry = retry or RetryConfig()

Anthropic

AnthropicProvider 

AnthropicProvider(*, model: str, api_key: str | None = None, timeout_s: float = 30.0, retry: RetryConfig | None = None)

Bases: LLMProvider

Source code in silkweb/llm/providers/base.py 
155
156
157
158
159
160
161
162
163
164
165
166
def __init__(
    self,
    *,
    model: str,
    api_key: str | None = None,
    timeout_s: float = 30.0,
    retry: RetryConfig | None = None,
) -> None:
    self.model = model
    self.api_key = api_key
    self.timeout_s = timeout_s
    self.retry = retry or RetryConfig()

llama.cpp

LlamaCppProvider 

LlamaCppProvider(*, model: str, api_key: str | None = None, timeout_s: float = 30.0, retry=None, model_path: str | None = None, n_ctx: int = 8192, **kwargs: Any)

Bases: LLMProvider

Source code in silkweb/llm/providers/llamacpp.py 
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
def __init__(
    self,
    *,
    model: str,
    api_key: str | None = None,
    timeout_s: float = 30.0,
    retry=None,
    model_path: str | None = None,
    n_ctx: int = 8192,
    **kwargs: Any,
) -> None:
    super().__init__(model=model, api_key=api_key, timeout_s=timeout_s, retry=retry)
    self.model_path = model_path or model
    self.n_ctx = n_ctx
    self._llama = None
    self._llama_kwargs = kwargs

Registry

parse_model_uri 

parse_model_uri(uri: str) -> ParsedModelURI

Parse provider/model URIs like ollama/qwen2.5:14b.

Source code in silkweb/llm/providers/registry.py 
36
37
38
39
40
41
42
43
44
45
46
47
48
49
def parse_model_uri(uri: str) -> ParsedModelURI:
    """
    Parse `provider/model` URIs like `ollama/qwen2.5:14b`.
    """
    if not uri or "/" not in uri:
        raise SilkwebConfigError(message="Invalid model URI.", key="model", value=uri)
    provider, model = uri.split("/", 1)
    provider = provider.strip().lower()
    model = model.strip()
    if provider not in {"ollama", "openai", "anthropic", "llamacpp"}:
        raise SilkwebConfigError(message="Unknown model provider.", key="provider", value=provider)
    if not model:
        raise SilkwebConfigError(message="Missing model name in URI.", key="model", value=uri)
    return ParsedModelURI(provider=provider, model=model)  # type: ignore[arg-type]

create_provider 

create_provider(uri: str, *, api_key: str | None = None, timeout_s: float | None = None, **kwargs: Any) -> LLMProvider
Source code in silkweb/llm/providers/registry.py 
 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
def create_provider(
    uri: str,
    *,
    api_key: str | None = None,
    timeout_s: float | None = None,
    **kwargs: Any,
) -> LLMProvider:
    if timeout_s is None:
        from ...config import get_config

        timeout_s = get_config().llm_timeout_ms / 1000.0

    parsed = parse_model_uri(uri)

    ensure_metrics_server()
    metrics = get_metrics()

    class LoggedProvider(LLMProvider):
        def __init__(self, inner: LLMProvider, provider_name: str) -> None:
            super().__init__(
                model=inner.model,
                api_key=getattr(inner, "api_key", None),
                timeout_s=inner.timeout_s,
                retry=inner.retry,
            )
            self._inner = inner
            self._provider_name = provider_name

        def unwrap(self) -> LLMProvider:
            """Return the underlying concrete provider instance."""
            return self._inner

        async def generate(self, messages, system=None, max_tokens=None, temperature=0.2) -> str:  # type: ignore[override]
            t = Timer()
            log_event(
                "llm_call_start", model=self.model, task="generate", provider=self._provider_name
            )
            try:
                out = await self._inner.generate(
                    messages,
                    system=system,
                    max_tokens=max_tokens,
                    temperature=temperature,
                )
                return out
            except Exception as e:
                log_event(
                    "llm_call_complete",
                    model=self.model,
                    task="generate",
                    duration_ms=int(t.seconds() * 1000),
                    error=repr(e),
                )
                raise
            else:
                log_event(
                    "llm_call_complete",
                    model=self.model,
                    task="generate",
                    duration_ms=int(t.seconds() * 1000),
                )
            finally:
                metrics.llm_calls_total.labels(model=self.model, task="generate").inc()
                metrics.llm_duration_seconds.labels(model=self.model, task="generate").observe(
                    t.seconds()
                )

        async def generate_json(
            self, messages, system=None, schema=None, max_tokens=None, temperature=0.2
        ):  # type: ignore[override]
            t = Timer()
            t_wall = time.perf_counter()
            approx_in = _approx_llm_chars(messages, system)
            print(
                f"[silkweb]   llm generate_json → provider: "
                f"chars≈{approx_in:,} max_tokens={max_tokens} "
                f"timeout_s={self.timeout_s} model={self.model!r}",
                flush=True,
            )
            log_event(
                "llm_call_start",
                model=self.model,
                task="generate_json",
                provider=self._provider_name,
            )
            try:
                print(
                    f"[silkweb]   llm generate_json: awaiting {self._provider_name!r} "
                    f"(this is usually network + model time)...",
                    flush=True,
                )
                out = await self._inner.generate_json(
                    messages,
                    system=system,
                    schema=schema,
                    max_tokens=max_tokens,
                    temperature=temperature,
                )
                return out
            except Exception as e:
                log_event(
                    "llm_call_complete",
                    model=self.model,
                    task="generate_json",
                    duration_ms=int(t.seconds() * 1000),
                    error=repr(e),
                )
                print(
                    f"[silkweb]   llm generate_json: failed after {time.perf_counter() - t_wall:.2f}s: "
                    f"{type(e).__name__}",
                    flush=True,
                )
                raise
            else:
                log_event(
                    "llm_call_complete",
                    model=self.model,
                    task="generate_json",
                    duration_ms=int(t.seconds() * 1000),
                )
                print(
                    f"[silkweb]   llm generate_json: done in {time.perf_counter() - t_wall:.2f}s",
                    flush=True,
                )
            finally:
                metrics.llm_calls_total.labels(model=self.model, task="generate_json").inc()
                metrics.llm_duration_seconds.labels(model=self.model, task="generate_json").observe(
                    t.seconds()
                )

        async def embed(self, texts):  # type: ignore[override]
            t = Timer()
            log_event(
                "llm_call_start", model=self.model, task="embed", provider=self._provider_name
            )
            try:
                out = await self._inner.embed(texts)
                return out
            except Exception as e:
                log_event(
                    "llm_call_complete",
                    model=self.model,
                    task="embed",
                    duration_ms=int(t.seconds() * 1000),
                    error=repr(e),
                )
                raise
            else:
                log_event(
                    "llm_call_complete",
                    model=self.model,
                    task="embed",
                    duration_ms=int(t.seconds() * 1000),
                )
            finally:
                metrics.llm_calls_total.labels(model=self.model, task="embed").inc()
                metrics.llm_duration_seconds.labels(model=self.model, task="embed").observe(
                    t.seconds()
                )

    if parsed.provider == "ollama":
        inner = OllamaProvider(model=parsed.model, api_key=api_key, timeout_s=timeout_s, **kwargs)
        return LoggedProvider(inner, "ollama")
    if parsed.provider == "openai":
        inner = OpenAIProvider(model=parsed.model, api_key=api_key, timeout_s=timeout_s, **kwargs)
        return LoggedProvider(inner, "openai")
    if parsed.provider == "anthropic":
        inner = AnthropicProvider(
            model=parsed.model, api_key=api_key, timeout_s=timeout_s, **kwargs
        )
        return LoggedProvider(inner, "anthropic")
    if parsed.provider == "llamacpp":
        inner = LlamaCppProvider(model=parsed.model, api_key=api_key, timeout_s=timeout_s, **kwargs)
        return LoggedProvider(inner, "llamacpp")

    raise SilkwebConfigError(message="Unknown provider.", key="provider", value=parsed.provider)