synth-ai 0.1.0.dev10__tar.gz → 0.1.0.dev11__tar.gz

{synth_ai-0.1.0.dev10 → synth_ai-0.1.0.dev11}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: synth-ai
-Version: 0.1.0.dev10
+Version: 0.1.0.dev11
 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 → synth_ai-0.1.0.dev11}/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-0.1.0.dev10 → synth_ai-0.1.0.dev11}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "synth-ai"
-version = "0.1.0.dev10"
+version = "0.1.0.dev11"
 description = "Software for aiding the best and multiplying the will."
 readme = "README.md"
 authors = [{ name = "Josh Purtell", email = "josh@usesynth.ai" }]
@@ -26,6 +26,8 @@ dependencies = [
     "groq>=0.18.0",
     "pytest-timeout>=2.3.1",
     "lock>=2018.3.25.2110",
+    "ollama>=0.4.7",
+    "mistralai>=1.5.0",
 ]
 requires-python = ">=3.10"
 

{synth_ai-0.1.0.dev10 → synth_ai-0.1.0.dev11}/setup.py

@@ -2,7 +2,7 @@ from setuptools import find_packages, setup
 
 setup(
     name="synth-ai",
-    version="0.1.0.dev10",
+    version="0.1.0.dev11",
     packages=find_packages(),
     install_requires=[
         "openai",

{synth_ai-0.1.0.dev10 → synth_ai-0.1.0.dev11}/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-0.1.0.dev10 → synth_ai-0.1.0.dev11}/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] = [
@@ -35,6 +36,10 @@ groq_naming_regexes: List[Pattern] = [
     re.compile(r"^llama-3.1-8b-instant$"),
 ]
 
+mistral_naming_regexes: List[Pattern] = [
+    re.compile(r"^mistral-.*$"),
+]
+
 
 def get_client(
     model_name: str,
@@ -64,5 +69,7 @@ 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-0.1.0.dev11/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-0.1.0.dev11/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 → synth_ai-0.1.0.dev11}/synth_ai.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: synth-ai
-Version: 0.1.0.dev10
+Version: 0.1.0.dev11
 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 → synth_ai-0.1.0.dev11}/synth_ai.egg-info/SOURCES.txt

@@ -44,12 +44,14 @@ synth_ai/zyk/lms/vendors/retries.py
 synth_ai/zyk/lms/vendors/core/__init__.py
 synth_ai/zyk/lms/vendors/core/anthropic_api.py
 synth_ai/zyk/lms/vendors/core/gemini_api.py
+synth_ai/zyk/lms/vendors/core/mistral_api.py
 synth_ai/zyk/lms/vendors/core/openai_api.py
 synth_ai/zyk/lms/vendors/local/__init__.py
 synth_ai/zyk/lms/vendors/local/ollama.py
 synth_ai/zyk/lms/vendors/supported/__init__.py
 synth_ai/zyk/lms/vendors/supported/deepseek.py
 synth_ai/zyk/lms/vendors/supported/groq.py
+synth_ai/zyk/lms/vendors/supported/ollama.py
 synth_ai/zyk/lms/vendors/supported/together.py
 tests/test_agent.py
 tests/test_recursive_structured_outputs.py

{synth_ai-0.1.0.dev10 → synth_ai-0.1.0.dev11}/synth_ai.egg-info/requires.txt

@@ -12,3 +12,5 @@ datasets>=3.2.0
 groq>=0.18.0
 pytest-timeout>=2.3.1
 lock>=2018.3.25.2110
+ollama>=0.4.7
+mistralai>=1.5.0