Skip to content

Language Models API

LanguageModel

Bases: SynalinksSaveable

A language model API wrapper.

A language model is a type of AI model designed to generate, and interpret human language. It is trained on large amounts of text data to learn patterns and structures in language. Language models can perform various tasks such as text generation, translation, summarization, and answering questions.

We support providers that implement constrained structured output like OpenAI, Azure, Ollama or Mistral. In addition we support providers that otherwise allow to constrain the use of a specific tool like Groq or Anthropic.

For the complete list of models, please refer to the providers documentation.

Using OpenAI models

import synalinks
import os

os.environ["OPENAI_API_KEY"] = "your-api-key"

language_model = synalinks.LanguageModel(
    model="openai/gpt-4o-mini",
)

Using Groq models

import synalinks
import os

os.environ["GROQ_API_KEY"] = "your-api-key"

language_model = synalinks.LanguageModel(
    model="groq/llama3-8b-8192",
)

Using Anthropic models

import synalinks
import os

os.environ["ANTHROPIC_API_KEY"] = "your-api-key"

language_model = synalinks.LanguageModel(
    model="anthropic/claude-3-sonnet-20240229",
)

Using Mistral models

import synalinks
import os

os.environ["MISTRAL_API_KEY"] = "your-api-key"

language_model = synalinks.LanguageModel(
    model="mistral/codestral-latest",
)

Using Ollama models

import synalinks
import os

language_model = synalinks.LanguageModel(
    model="ollama/mistral",
)

Using Azure OpenAI models

import synalinks
import os

os.environ["AZURE_API_KEY"] = "your-api-key"
os.environ["AZURE_API_BASE"] = "your-api-key"
os.environ["AZURE_API_VERSION"] = "your-api-key"

language_model = synalinks.LanguageModel(
    model="azure/<your_deployment_name>",
)

Using Google Gemini models

import synalinks
import os

os.environ["GEMINI_API_KEY"] = "your-api-key"

language_model = synalinks.LanguageModel(
    model="gemini/gemini-1.5-pro",
)

To cascade models in case there is anything wrong with the model provider (hence making your pipelines more robust). Use the fallback argument like in this example:

import synalinks
import os

os.environ["OPENAI_API_KEY"] = "your-api-key"
os.environ["ANTHROPIC_API_KEY"] = "your-api-key"

language_model = synalinks.LanguageModel(
    model="anthropic/claude-3-sonnet-20240229",
    fallback=synalinks.LanguageModel(
        model="openai/gpt-4o-mini",
    )
)

Note: Obviously, use an .env file and .gitignore to avoid putting your API keys in the code or a config file that can lead to leackage when pushing it into repositories.

Parameters:

Name Type Description Default
model str

The model to use.

 None
api_base str

Optional. The endpoint to use.

 None
timeout int

Optional. The timeout value in seconds (Default to 600).

 600
retry int

Optional. The number of retry (default to 5).

 5
fallback LanguageModel

Optional. The language model to fallback if anything is wrong.

 None
caching bool

Optional. Enable caching of LM calls (Default to True).

 True
Source code in synalinks/src/language_models/language_model.py 
 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
@synalinks_export(
    [
        "synalinks.LanguageModel",
        "synalinks.language_models.LanguageModel",
    ]
)
class LanguageModel(SynalinksSaveable):
    """A language model API wrapper.

    A language model is a type of AI model designed to generate, and interpret human
    language. It is trained on large amounts of text data to learn patterns and
    structures in language. Language models can perform various tasks such as text
    generation, translation, summarization, and answering questions.

    We support providers that implement *constrained structured output*
    like OpenAI, Azure, Ollama or Mistral. In addition we support providers that otherwise
    allow to constrain the use of a specific tool like Groq or Anthropic.

    For the complete list of models, please refer to the providers documentation.

    **Using OpenAI models**

    ```python
    import synalinks
    import os

    os.environ["OPENAI_API_KEY"] = "your-api-key"

    language_model = synalinks.LanguageModel(
        model="openai/gpt-4o-mini",
    )
    ```

    **Using Groq models**

    ```python
    import synalinks
    import os

    os.environ["GROQ_API_KEY"] = "your-api-key"

    language_model = synalinks.LanguageModel(
        model="groq/llama3-8b-8192",
    )
    ```

    **Using Anthropic models**

    ```python
    import synalinks
    import os

    os.environ["ANTHROPIC_API_KEY"] = "your-api-key"

    language_model = synalinks.LanguageModel(
        model="anthropic/claude-3-sonnet-20240229",
    )
    ```

    **Using Mistral models**

    ```python
    import synalinks
    import os

    os.environ["MISTRAL_API_KEY"] = "your-api-key"

    language_model = synalinks.LanguageModel(
        model="mistral/codestral-latest",
    )
    ```

    **Using Ollama models**

    ```python
    import synalinks
    import os

    language_model = synalinks.LanguageModel(
        model="ollama/mistral",
    )
    ```

    **Using Azure OpenAI models**

    ```python
    import synalinks
    import os

    os.environ["AZURE_API_KEY"] = "your-api-key"
    os.environ["AZURE_API_BASE"] = "your-api-key"
    os.environ["AZURE_API_VERSION"] = "your-api-key"

    language_model = synalinks.LanguageModel(
        model="azure/<your_deployment_name>",
    )
    ```

    **Using Google Gemini models**

    ```python
    import synalinks
    import os

    os.environ["GEMINI_API_KEY"] = "your-api-key"

    language_model = synalinks.LanguageModel(
        model="gemini/gemini-1.5-pro",
    )
    ```

    To cascade models in case there is anything wrong with
    the model provider (hence making your pipelines more robust).
    Use the `fallback` argument like in this example:

    ```python
    import synalinks
    import os

    os.environ["OPENAI_API_KEY"] = "your-api-key"
    os.environ["ANTHROPIC_API_KEY"] = "your-api-key"

    language_model = synalinks.LanguageModel(
        model="anthropic/claude-3-sonnet-20240229",
        fallback=synalinks.LanguageModel(
            model="openai/gpt-4o-mini",
        )
    )
    ```

    **Note**: Obviously, use an `.env` file and `.gitignore` to avoid
    putting your API keys in the code or a config file that can lead to
    leackage when pushing it into repositories.

    Args:
        model (str): The model to use.
        api_base (str): Optional. The endpoint to use.
        timeout (int): Optional. The timeout value in seconds (Default to 600).
        retry (int): Optional. The number of retry (default to 5).
        fallback (LanguageModel): Optional. The language model to fallback
            if anything is wrong.
        caching (bool): Optional. Enable caching of LM calls (Default to True).
    """

    def __init__(
        self,
        model=None,
        api_base=None,
        timeout=600,
        retry=5,
        fallback=None,
        caching=True,
    ):
        if model is None:
            raise ValueError("You need to set the `model` argument for any LanguageModel")
        model_provider = model.split("/")[0]
        if model_provider == "ollama":
            # Switch from `ollama` to `ollama_chat`
            # because it have better performance due to the chat prompts
            model = model.replace("ollama", "ollama_chat")
        self.model = model
        self.fallback = fallback
        if self.model.startswith("ollama") and not api_base:
            self.api_base = "http://localhost:11434"
        else:
            self.api_base = api_base
        self.timeout = timeout
        self.retry = retry
        self.caching = caching

    async def __call__(self, messages, schema=None, streaming=False, **kwargs):
        """
        Call method to generate a response using the language model.

        Args:
            messages (dict): A formatted dict of chat messages.
            schema (dict): The target JSON schema for structed output (optional).
                If None, output a ChatMessage-like answer.
            streaming (bool): Enable streaming (optional). Default to False.
                Can be enabled only if schema is None.
            **kwargs (keyword arguments): The additional keywords arguments
                forwarded to the LM call.
        Returns:
            (dict): The generated structured response.
        """
        formatted_messages = messages.get_json().get("messages", [])
        json_instance = {}
        input_kwargs = copy.deepcopy(kwargs)
        schema = copy.deepcopy(schema)
        if schema:
            if self.model.startswith("groq"):
                # Use a tool created on the fly for groq
                kwargs.update(
                    {
                        "tools": [
                            {
                                "function": {
                                    "name": "structured_output",
                                    "description": "Generate a valid JSON output",
                                    "parameters": schema.get("properties"),
                                },
                                "type": "function",
                            }
                        ],
                        "tool_choice": {
                            "type": "function",
                            "function": {"name": "structured_output"},
                        },
                    }
                )
            elif self.model.startswith("anthropic"):
                # Use a tool created on the fly for anthropic
                kwargs.update(
                    {
                        "tools": [
                            {
                                "name": "structured_output",
                                "description": "Generate a valid JSON output",
                                "input_schema": {
                                    "type": "object",
                                    "properties": schema.get("properties"),
                                    "required": schema.get("required"),
                                },
                            }
                        ],
                        "tool_choice": {
                            "type": "tool",
                            "name": "structured_output",
                        },
                    }
                )
            elif self.model.startswith("ollama") or self.model.startswith("mistral"):
                # Use constrained structured output for ollama/mistral
                kwargs.update(
                    {
                        "response_format": {
                            "type": "json_schema",
                            "json_schema": {"schema": schema},
                            "strict": True,
                        },
                    }
                )
            elif self.model.startswith("openai") or self.model.startswith("azure"):
                # Use constrained structured output for openai
                # OpenAI require the field  "additionalProperties"
                # Also OpenAI disallow the field "description" in $ref
                if "properties" in schema:
                    for prop_key, prop_value in schema["properties"].items():
                        if "$ref" in prop_value and "description" in prop_value:
                            del prop_value["description"]
                kwargs.update(
                    {
                        "response_format": {
                            "type": "json_schema",
                            "json_schema": {
                                "name": "structured_output",
                                "strict": True,
                                "schema": schema,
                            },
                        }
                    }
                )
            elif self.model.startswith("gemini"):
                kwargs.update(
                    {
                        "response_format": {
                            "type": "json_schema",
                            "json_schema": schema,
                            "strict": True,
                        }
                    }
                )
            elif self.model.startswith("xai"):
                kwargs.update(
                    {
                        "response_format": {
                            "type": "json_schema",
                            "json_schema": schema,
                            "strict": True,
                        }
                    }
                )
            else:
                provider = self.model.split("/")[0]
                raise ValueError(
                    f"LM provider '{provider}' not supported yet, please ensure that"
                    " they support constrained structured output and fill an issue."
                )

        if self.api_base:
            kwargs.update(
                {
                    "api_base": self.api_base,
                }
            )
        if streaming and schema:
            streaming = False
        if streaming:
            kwargs.update({"stream": True})
        for i in range(self.retry):
            try:
                response_str = ""
                response = await litellm.acompletion(
                    model=self.model,
                    messages=formatted_messages,
                    timeout=self.timeout,
                    caching=self.caching,
                    **kwargs,
                )
                if streaming:
                    return StreamingIterator(response)
                if (
                    self.model.startswith("groq") or self.model.startswith("anthropic")
                ) and schema:
                    response_str = response["choices"][0]["message"]["tool_calls"][0][
                        "function"
                    ]["arguments"]
                else:
                    response_str = response["choices"][0]["message"]["content"].strip()
                if schema:
                    json_instance = json.loads(response_str)
                else:
                    json_instance = {
                        "role": ChatRole.ASSISTANT,
                        "content": response_str,
                        "tool_call_id": None,
                        "tool_calls": [],
                    }
                return json_instance
            except Exception as e:
                warnings.warn(
                    f"Error occured while trying to call {self}: "
                    + str(e)
                    + f"\nReceived response={shorten_text(response_str)}"
                )
            await asyncio.sleep(1)
        if self.fallback:
            return self.fallback(
                messages,
                schema=schema,
                streaming=streaming,
                **input_kwargs,
            )
        else:
            return None

    def _obj_type(self):
        return "LanguageModel"

    def get_config(self):
        config = {
            "model": self.model,
            "api_base": self.api_base,
            "timeout": self.timeout,
            "retry": self.retry,
            "caching": self.caching,
        }
        if self.fallback:
            fallback_config = {
                "fallback": serialization_lib.serialize_synalinks_object(
                    self.fallback,
                )
            }
            return {**fallback_config, **config}
        else:
            return config

    @classmethod
    def from_config(cls, config):
        if "fallback" in config:
            fallback = serialization_lib.deserialize_synalinks_object(
                config.pop("fallback")
            )
            return cls(fallback=fallback, **config)
        else:
            return cls(**config)

    def __repr__(self):
        api_base = f" api_base={self.api_base}" if self.api_base else ""
        return f"<LanguageModel model={self.model}{api_base}>"

__call__(messages, schema=None, streaming=False, **kwargs) async

Call method to generate a response using the language model.

Parameters:

Name Type Description Default
messages dict

A formatted dict of chat messages.

 required
schema dict

The target JSON schema for structed output (optional). If None, output a ChatMessage-like answer.

 None
streaming bool

Enable streaming (optional). Default to False. Can be enabled only if schema is None.

 False
**kwargs keyword arguments

The additional keywords arguments forwarded to the LM call.

 {}

Returns: (dict): The generated structured response.

Source code in synalinks/src/language_models/language_model.py 
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
async def __call__(self, messages, schema=None, streaming=False, **kwargs):
    """
    Call method to generate a response using the language model.

    Args:
        messages (dict): A formatted dict of chat messages.
        schema (dict): The target JSON schema for structed output (optional).
            If None, output a ChatMessage-like answer.
        streaming (bool): Enable streaming (optional). Default to False.
            Can be enabled only if schema is None.
        **kwargs (keyword arguments): The additional keywords arguments
            forwarded to the LM call.
    Returns:
        (dict): The generated structured response.
    """
    formatted_messages = messages.get_json().get("messages", [])
    json_instance = {}
    input_kwargs = copy.deepcopy(kwargs)
    schema = copy.deepcopy(schema)
    if schema:
        if self.model.startswith("groq"):
            # Use a tool created on the fly for groq
            kwargs.update(
                {
                    "tools": [
                        {
                            "function": {
                                "name": "structured_output",
                                "description": "Generate a valid JSON output",
                                "parameters": schema.get("properties"),
                            },
                            "type": "function",
                        }
                    ],
                    "tool_choice": {
                        "type": "function",
                        "function": {"name": "structured_output"},
                    },
                }
            )
        elif self.model.startswith("anthropic"):
            # Use a tool created on the fly for anthropic
            kwargs.update(
                {
                    "tools": [
                        {
                            "name": "structured_output",
                            "description": "Generate a valid JSON output",
                            "input_schema": {
                                "type": "object",
                                "properties": schema.get("properties"),
                                "required": schema.get("required"),
                            },
                        }
                    ],
                    "tool_choice": {
                        "type": "tool",
                        "name": "structured_output",
                    },
                }
            )
        elif self.model.startswith("ollama") or self.model.startswith("mistral"):
            # Use constrained structured output for ollama/mistral
            kwargs.update(
                {
                    "response_format": {
                        "type": "json_schema",
                        "json_schema": {"schema": schema},
                        "strict": True,
                    },
                }
            )
        elif self.model.startswith("openai") or self.model.startswith("azure"):
            # Use constrained structured output for openai
            # OpenAI require the field  "additionalProperties"
            # Also OpenAI disallow the field "description" in $ref
            if "properties" in schema:
                for prop_key, prop_value in schema["properties"].items():
                    if "$ref" in prop_value and "description" in prop_value:
                        del prop_value["description"]
            kwargs.update(
                {
                    "response_format": {
                        "type": "json_schema",
                        "json_schema": {
                            "name": "structured_output",
                            "strict": True,
                            "schema": schema,
                        },
                    }
                }
            )
        elif self.model.startswith("gemini"):
            kwargs.update(
                {
                    "response_format": {
                        "type": "json_schema",
                        "json_schema": schema,
                        "strict": True,
                    }
                }
            )
        elif self.model.startswith("xai"):
            kwargs.update(
                {
                    "response_format": {
                        "type": "json_schema",
                        "json_schema": schema,
                        "strict": True,
                    }
                }
            )
        else:
            provider = self.model.split("/")[0]
            raise ValueError(
                f"LM provider '{provider}' not supported yet, please ensure that"
                " they support constrained structured output and fill an issue."
            )

    if self.api_base:
        kwargs.update(
            {
                "api_base": self.api_base,
            }
        )
    if streaming and schema:
        streaming = False
    if streaming:
        kwargs.update({"stream": True})
    for i in range(self.retry):
        try:
            response_str = ""
            response = await litellm.acompletion(
                model=self.model,
                messages=formatted_messages,
                timeout=self.timeout,
                caching=self.caching,
                **kwargs,
            )
            if streaming:
                return StreamingIterator(response)
            if (
                self.model.startswith("groq") or self.model.startswith("anthropic")
            ) and schema:
                response_str = response["choices"][0]["message"]["tool_calls"][0][
                    "function"
                ]["arguments"]
            else:
                response_str = response["choices"][0]["message"]["content"].strip()
            if schema:
                json_instance = json.loads(response_str)
            else:
                json_instance = {
                    "role": ChatRole.ASSISTANT,
                    "content": response_str,
                    "tool_call_id": None,
                    "tool_calls": [],
                }
            return json_instance
        except Exception as e:
            warnings.warn(
                f"Error occured while trying to call {self}: "
                + str(e)
                + f"\nReceived response={shorten_text(response_str)}"
            )
        await asyncio.sleep(1)
    if self.fallback:
        return self.fallback(
            messages,
            schema=schema,
            streaming=streaming,
            **input_kwargs,
        )
    else:
        return None