synth-ai 0.1.0.dev10__py3-none-any.whl → 0.1.0.dev12__py3-none-any.whl

public_tests/test_all_structured_outputs.py

@@ -68,6 +68,12 @@ def models():
             temperature=0.1,
             structured_output_mode="stringified_json",
         ),
+        "mistral-small-latest": LM(
+            model_name="mistral-small-latest",
+            formatting_model_name="gpt-4o-mini",
+            temperature=0.1,
+            structured_output_mode="stringified_json",
+        ),
     }
 
 

synth_ai/zyk/lms/core/all.py

@@ -7,6 +7,7 @@ from synth_ai.zyk.lms.vendors.core.openai_api import (
 from synth_ai.zyk.lms.vendors.supported.deepseek import DeepSeekAPI
 from synth_ai.zyk.lms.vendors.supported.together import TogetherAPI
 from synth_ai.zyk.lms.vendors.supported.groq import GroqAPI
+from synth_ai.zyk.lms.vendors.core.mistral_api import MistralAPI
 
 
 class OpenAIClient(OpenAIPrivate):
@@ -39,3 +40,8 @@ class TogetherClient(TogetherAPI):
 class GroqClient(GroqAPI):
     def __init__(self):
         super().__init__()
+
+
+class MistralClient(MistralAPI):
+    def __init__(self):
+        super().__init__()

synth_ai/zyk/lms/core/vendor_clients.py

@@ -9,6 +9,7 @@ from synth_ai.zyk.lms.core.all import (
     OpenAIStructuredOutputClient,
     TogetherClient,
     GroqAPI,
+    MistralAPI,
 )
 
 openai_naming_regexes: List[Pattern] = [
@@ -33,6 +34,19 @@ together_naming_regexes: List[Pattern] = [
 groq_naming_regexes: List[Pattern] = [
     re.compile(r"^llama-3.3-70b-versatile$"),
     re.compile(r"^llama-3.1-8b-instant$"),
+    re.compile(r"^qwen-2.5-32b$"),
+    re.compile(r"^deepseek-r1-distill-qwen-32b$"),
+    re.compile(r"^deepseek-r1-distill-llama-70b-specdec$"),
+    re.compile(r"^deepseek-r1-distill-llama-70b$"),
+    re.compile(r"^llama-3.3-70b-specdec$"),
+    re.compile(r"^llama-3.2-1b-preview$"),
+    re.compile(r"^llama-3.2-3b-preview$"),
+    re.compile(r"^llama-3.2-11b-vision-preview$"),
+    re.compile(r"^llama-3.2-90b-vision-preview$"),
+]
+
+mistral_naming_regexes: List[Pattern] = [
+    re.compile(r"^mistral-.*$"),
 ]
 
 
@@ -64,5 +78,7 @@ def get_client(
         return TogetherClient()
     elif any(regex.match(model_name) for regex in groq_naming_regexes):
         return GroqAPI()
+    elif any(regex.match(model_name) for regex in mistral_naming_regexes):
+        return MistralAPI()
     else:
         raise ValueError(f"Invalid model name: {model_name}")

synth_ai/zyk/lms/vendors/core/mistral_api.py

@@ -0,0 +1,221 @@
+import json
+import os
+from typing import Any, Dict, List, Tuple, Type
+
+import pydantic
+from mistralai import Mistral  # use Mistral as both sync and async client
+from pydantic import BaseModel
+
+from synth_ai.zyk.lms.caching.initialize import get_cache_handler
+from synth_ai.zyk.lms.vendors.base import VendorBase
+from synth_ai.zyk.lms.vendors.constants import SPECIAL_BASE_TEMPS
+from synth_ai.zyk.lms.vendors.core.openai_api import OpenAIStructuredOutputClient
+from synth_ai.zyk.lms.vendors.retries import BACKOFF_TOLERANCE, backoff
+
+# Since the mistralai package doesn't expose an exceptions module,
+# we fallback to catching all Exceptions for retry.
+MISTRAL_EXCEPTIONS_TO_RETRY: Tuple[Type[Exception], ...] = (Exception,)
+
+
+class MistralAPI(VendorBase):
+    used_for_structured_outputs: bool = True
+    exceptions_to_retry: Tuple = MISTRAL_EXCEPTIONS_TO_RETRY
+    _openai_fallback: Any
+
+    def __init__(
+        self,
+        exceptions_to_retry: Tuple[Type[Exception], ...] = MISTRAL_EXCEPTIONS_TO_RETRY,
+        used_for_structured_outputs: bool = False,
+    ):
+        self.used_for_structured_outputs = used_for_structured_outputs
+        self.exceptions_to_retry = exceptions_to_retry
+        self._openai_fallback = None
+
+    @backoff.on_exception(
+        backoff.expo,
+        MISTRAL_EXCEPTIONS_TO_RETRY,
+        max_tries=BACKOFF_TOLERANCE,
+        on_giveup=lambda e: print(e),
+    )
+    async def _hit_api_async(
+        self,
+        model: str,
+        messages: List[Dict[str, Any]],
+        lm_config: Dict[str, Any],
+        use_ephemeral_cache_only: bool = False,
+    ) -> str:
+        assert (
+            lm_config.get("response_model", None) is None
+        ), "response_model is not supported for standard calls"
+        used_cache_handler = get_cache_handler(use_ephemeral_cache_only)
+        cache_result = used_cache_handler.hit_managed_cache(
+            model, messages, lm_config=lm_config
+        )
+        if cache_result:
+            return (
+                cache_result["response"]
+                if isinstance(cache_result, dict)
+                else cache_result
+            )
+
+        mistral_messages = [
+            {"role": msg["role"], "content": msg["content"]} for msg in messages
+        ]
+        async with Mistral(api_key=os.getenv("MISTRAL_API_KEY", "")) as client:
+            response = await client.chat.complete_async(
+                model=model,
+                messages=mistral_messages,
+                max_tokens=lm_config.get("max_tokens", 4096),
+                temperature=lm_config.get(
+                    "temperature", SPECIAL_BASE_TEMPS.get(model, 0)
+                ),
+                stream=False,
+            )
+        api_result = response.choices[0].message.content
+        used_cache_handler.add_to_managed_cache(
+            model, messages, lm_config=lm_config, output=api_result
+        )
+        return api_result
+
+    @backoff.on_exception(
+        backoff.expo,
+        MISTRAL_EXCEPTIONS_TO_RETRY,
+        max_tries=BACKOFF_TOLERANCE,
+        on_giveup=lambda e: print(e),
+    )
+    def _hit_api_sync(
+        self,
+        model: str,
+        messages: List[Dict[str, Any]],
+        lm_config: Dict[str, Any],
+        use_ephemeral_cache_only: bool = False,
+    ) -> str:
+        assert (
+            lm_config.get("response_model", None) is None
+        ), "response_model is not supported for standard calls"
+        used_cache_handler = get_cache_handler(use_ephemeral_cache_only)
+        cache_result = used_cache_handler.hit_managed_cache(
+            model, messages, lm_config=lm_config
+        )
+        if cache_result:
+            return (
+                cache_result["response"]
+                if isinstance(cache_result, dict)
+                else cache_result
+            )
+
+        mistral_messages = [
+            {"role": msg["role"], "content": msg["content"]} for msg in messages
+        ]
+        with Mistral(api_key=os.getenv("MISTRAL_API_KEY", "")) as client:
+            response = client.chat.complete(
+                model=model,
+                messages=mistral_messages,
+                max_tokens=lm_config.get("max_tokens", 4096),
+                temperature=lm_config.get(
+                    "temperature", SPECIAL_BASE_TEMPS.get(model, 0)
+                ),
+                stream=False,
+            )
+        api_result = response.choices[0].message.content
+        used_cache_handler.add_to_managed_cache(
+            model, messages, lm_config=lm_config, output=api_result
+        )
+        return api_result
+
+    async def _hit_api_async_structured_output(
+        self,
+        model: str,
+        messages: List[Dict[str, Any]],
+        response_model: BaseModel,
+        temperature: float,
+        use_ephemeral_cache_only: bool = False,
+    ) -> Any:
+        try:
+            mistral_messages = [
+                {"role": msg["role"], "content": msg["content"]} for msg in messages
+            ]
+            async with Mistral(api_key=os.getenv("MISTRAL_API_KEY", "")) as client:
+                response = await client.chat.complete_async(
+                    model=model,
+                    messages=mistral_messages,
+                    max_tokens=4096,
+                    temperature=temperature,
+                    stream=False,
+                )
+            result = response.choices[0].message.content
+            parsed = json.loads(result)
+            return response_model(**parsed)
+        except (json.JSONDecodeError, pydantic.ValidationError):
+            if self._openai_fallback is None:
+                self._openai_fallback = OpenAIStructuredOutputClient()
+            return await self._openai_fallback._hit_api_async_structured_output(
+                model="gpt-4o",
+                messages=messages,
+                response_model=response_model,
+                temperature=temperature,
+                use_ephemeral_cache_only=use_ephemeral_cache_only,
+            )
+
+    def _hit_api_sync_structured_output(
+        self,
+        model: str,
+        messages: List[Dict[str, Any]],
+        response_model: BaseModel,
+        temperature: float,
+        use_ephemeral_cache_only: bool = False,
+    ) -> Any:
+        try:
+            mistral_messages = [
+                {"role": msg["role"], "content": msg["content"]} for msg in messages
+            ]
+            with Mistral(api_key=os.getenv("MISTRAL_API_KEY", "")) as client:
+                response = client.chat.complete(
+                    model=model,
+                    messages=mistral_messages,
+                    max_tokens=4096,
+                    temperature=temperature,
+                    stream=False,
+                )
+            result = response.choices[0].message.content
+            parsed = json.loads(result)
+            return response_model(**parsed)
+        except (json.JSONDecodeError, pydantic.ValidationError):
+            print("WARNING - Falling back to OpenAI - THIS IS SLOW")
+            if self._openai_fallback is None:
+                self._openai_fallback = OpenAIStructuredOutputClient()
+            return self._openai_fallback._hit_api_sync_structured_output(
+                model="gpt-4o",
+                messages=messages,
+                response_model=response_model,
+                temperature=temperature,
+                use_ephemeral_cache_only=use_ephemeral_cache_only,
+            )
+
+
+if __name__ == "__main__":
+    import asyncio
+
+    from pydantic import BaseModel
+
+    class TestModel(BaseModel):
+        name: str
+
+    client = MistralAPI(used_for_structured_outputs=True, exceptions_to_retry=[])
+    import time
+
+    t = time.time()
+
+    async def run_async():
+        response = await client._hit_api_async_structured_output(
+            model="mistral-large-latest",
+            messages=[{"role": "user", "content": "What is the capital of the moon?"}],
+            response_model=TestModel,
+            temperature=0.0,
+        )
+        print(response)
+        return response
+
+    response = asyncio.run(run_async())
+    t2 = time.time()
+    print(f"Got {len(response.name)} chars in {t2-t} seconds")

synth_ai/zyk/lms/vendors/supported/ollama.py

@@ -0,0 +1,14 @@
+from openai import OpenAI, AsyncOpenAI
+from synth_ai.zyk.lms.vendors.openai_standard import OpenAIStandard
+
+
+class OllamaAPI(OpenAIStandard):
+    def __init__(self):
+        self.sync_client = OpenAI(
+            base_url="http://localhost:11434/v1",
+            api_key="ollama",  # required, but unused
+        )
+        self.async_client = AsyncOpenAI(
+            base_url="http://localhost:11434/v1",
+            api_key="ollama",  # required, but unused
+        )

{synth_ai-0.1.0.dev10.dist-info → synth_ai-0.1.0.dev12.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: synth-ai
-Version: 0.1.0.dev10
+Version: 0.1.0.dev12
 Summary: Software for aiding the best and multiplying the will.
 Home-page: https://github.com/synth-laboratories/synth-ai
 Author: Josh Purtell
@@ -49,6 +49,8 @@ Requires-Dist: datasets>=3.2.0
 Requires-Dist: groq>=0.18.0
 Requires-Dist: pytest-timeout>=2.3.1
 Requires-Dist: lock>=2018.3.25.2110
+Requires-Dist: ollama>=0.4.7
+Requires-Dist: mistralai>=1.5.0
 Dynamic: author
 Dynamic: home-page
 

{synth_ai-0.1.0.dev10.dist-info → synth_ai-0.1.0.dev12.dist-info}/RECORD

@@ -1,6 +1,6 @@
 private_tests/try_synth_sdk.py,sha256=vk4lUEfpQfLACFl6Qw468t_lsuYxuoIIr05WRgWKGKY,24
 public_tests/test_agent.py,sha256=CjPPWuMWC_TzX1DkDald-bbAxgjXE-HPQvFhq2B--5k,22363
-public_tests/test_all_structured_outputs.py,sha256=SkvQq4paFVh2b3XfeZJ0ihd4LLKgVvVk2Yd4bIXZEEw,6307
+public_tests/test_all_structured_outputs.py,sha256=x7Gj5Ykpw8Ut_XlSOEBHRLJSagYSHDsO1unrMSlv_Vw,6534
 public_tests/test_recursive_structured_outputs.py,sha256=Ne-9XwnOxN7eSpGbNHOpegR-sRj589I84T6y8Z_4QnA,5781
 public_tests/test_structured_outputs.py,sha256=J7sfbGZ7OeB5ONIKpcCTymyayNyAdFfGokC1bcUrSx0,3651
 public_tests/test_synth_sdk.py,sha256=fqkzyzLb_NW4k8EiP2mJ5HZk3lDTi1juyTf9Gv_9wfc,14238
@@ -16,10 +16,10 @@ synth_ai/zyk/lms/caching/handler.py,sha256=sewq5rRfqXHzCEiXvdckbuxYp9ze_EjVSndnU
 synth_ai/zyk/lms/caching/initialize.py,sha256=zZls6RKAax6Z-8oJInGaSg_RPN_fEZ6e_RCX64lMLJw,416
 synth_ai/zyk/lms/caching/persistent.py,sha256=mQmP1z0rWVYjxwso5zIwd51Df2dWZvdHonuqsOY6SFI,2075
 synth_ai/zyk/lms/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-synth_ai/zyk/lms/core/all.py,sha256=kuuu0V1_DVJp8GtZztdwFiTwGVYIdJ946Y1V0D4zz8Y,1034
+synth_ai/zyk/lms/core/all.py,sha256=wakK0HhvYRuaQZmxClURyNf3vUkTbm3OABw3TgpMjOQ,1185
 synth_ai/zyk/lms/core/exceptions.py,sha256=K0BVdAzxVIchsvYZAaHEH1GAWBZvpxhFi-SPcJOjyPQ,205
 synth_ai/zyk/lms/core/main.py,sha256=fdAPBjnyOfLBPtcD0D79tn0f7SrKv49zCP6IhiAeH9Y,9019
-synth_ai/zyk/lms/core/vendor_clients.py,sha256=tuYf9Jio9mnCajPc1dCwurdprknRGnbHv7wwHC5tIdY,2119
+synth_ai/zyk/lms/core/vendor_clients.py,sha256=knvjXV7Pbjm_60cIqiw3mEIjehRZlfQZWXVLmS7m8GE,2746
 synth_ai/zyk/lms/cost/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 synth_ai/zyk/lms/cost/monitor.py,sha256=cSKIvw6WdPZIRubADWxQoh1MdB40T8-jjgfNUeUHIn0,5
 synth_ai/zyk/lms/cost/statefulness.py,sha256=TOsuXL8IjtKOYJ2aJQF8TwJVqn_wQ7AIwJJmdhMye7U,36
@@ -35,18 +35,20 @@ synth_ai/zyk/lms/vendors/retries.py,sha256=m-WvAiPix9ovnO2S-m53Td5VZDWBVBFuHuSK9
 synth_ai/zyk/lms/vendors/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 synth_ai/zyk/lms/vendors/core/anthropic_api.py,sha256=fcb0uVlInADl50MNYMT-IimM9mzO19D8_mSg9Gqp92Q,6986
 synth_ai/zyk/lms/vendors/core/gemini_api.py,sha256=vHwsIv6n1KxzxPtx10yxZcsVDr8j9ZUx8dPE9zNLWjM,5141
+synth_ai/zyk/lms/vendors/core/mistral_api.py,sha256=m11ItQ46VyyCUy6hv6mw5OmiqwHr07wV_NJVNnPHgiA,8080
 synth_ai/zyk/lms/vendors/core/openai_api.py,sha256=bpS7NcOJugF6rVOxsw3FMV30hhPQVkvZ0uV-MppisyE,4967
 synth_ai/zyk/lms/vendors/local/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 synth_ai/zyk/lms/vendors/local/ollama.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 synth_ai/zyk/lms/vendors/supported/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 synth_ai/zyk/lms/vendors/supported/deepseek.py,sha256=diFfdhPMO5bLFZxnYj7VT0v6jKTlOYESBkspUuVa2eY,529
 synth_ai/zyk/lms/vendors/supported/groq.py,sha256=Fbi7QvhdLx0F-VHO5PY-uIQlPR0bo3C9h1MvIOx8nz0,388
+synth_ai/zyk/lms/vendors/supported/ollama.py,sha256=K30VBFRTd7NYyPmyBVRZS2sm0UB651AHp9i3wd55W64,469
 synth_ai/zyk/lms/vendors/supported/together.py,sha256=Ni_jBqqGPN0PkkY-Ew64s3gNKk51k3FCpLSwlNhKbf0,342
 tests/test_agent.py,sha256=CjPPWuMWC_TzX1DkDald-bbAxgjXE-HPQvFhq2B--5k,22363
 tests/test_recursive_structured_outputs.py,sha256=Ne-9XwnOxN7eSpGbNHOpegR-sRj589I84T6y8Z_4QnA,5781
 tests/test_structured_outputs.py,sha256=J7sfbGZ7OeB5ONIKpcCTymyayNyAdFfGokC1bcUrSx0,3651
-synth_ai-0.1.0.dev10.dist-info/LICENSE,sha256=ynhjRQUfqA_RdGRATApfFA_fBAy9cno04sLtLUqxVFM,1069
-synth_ai-0.1.0.dev10.dist-info/METADATA,sha256=0JR0iVQFHPMzRDsm8PxJMsAM76PYPHKONQRNZQ2LsHU,2712
-synth_ai-0.1.0.dev10.dist-info/WHEEL,sha256=In9FTNxeP60KnTkGw7wk6mJPYd_dQSjEZmXdBdMCI-8,91
-synth_ai-0.1.0.dev10.dist-info/top_level.txt,sha256=5GzJO9j-KbJ_4ppxhmCUa_qdhHM4-9cHHNU76yAI8do,42
-synth_ai-0.1.0.dev10.dist-info/RECORD,,
+synth_ai-0.1.0.dev12.dist-info/LICENSE,sha256=ynhjRQUfqA_RdGRATApfFA_fBAy9cno04sLtLUqxVFM,1069
+synth_ai-0.1.0.dev12.dist-info/METADATA,sha256=hg3-nT8oN_Q9AJCqiRy-UhFcmViLIOsxGm0xkgBl2tA,2773
+synth_ai-0.1.0.dev12.dist-info/WHEEL,sha256=In9FTNxeP60KnTkGw7wk6mJPYd_dQSjEZmXdBdMCI-8,91
+synth_ai-0.1.0.dev12.dist-info/top_level.txt,sha256=5GzJO9j-KbJ_4ppxhmCUa_qdhHM4-9cHHNU76yAI8do,42
+synth_ai-0.1.0.dev12.dist-info/RECORD,,