Initial Gemini implementation (#149)

* Initial Gemini implementation

* Fix

* Update requirements.txt

* Add doc

* Update requirements.txt

council/__init__.py

@@ -6,5 +6,5 @@
 from .controllers import BasicController, ControllerBase, ExecutionUnit, LLMController
 from .evaluators import BasicEvaluator, EvaluatorBase, LLMEvaluator
 from .filters import BasicFilter, FilterBase
-from .llm import AnthropicLLM, AzureLLM, OpenAILLM
+from .llm import AnthropicLLM, AzureLLM, OpenAILLM, GeminiLLM
 from .runners import DoWhile, If, Parallel, ParallelFor, RunnerGenerator, RunnerPredicate, Sequential, While

council/llm/__init__.py

@@ -25,6 +25,9 @@
 from .anthropic_llm_configuration import AnthropicLLMConfiguration
 from .anthropic_llm import AnthropicLLM
 
+from .gemini_llm_configuration import GeminiLLMConfiguration
+from .gemini_llm import GeminiLLM
+
 
 def get_default_llm(max_retries: Optional[int] = None) -> LLMBase:
     provider = read_env_str("COUNCIL_DEFAULT_LLM_PROVIDER", default=LLMProviders.OpenAI).unwrap()
@@ -37,6 +40,8 @@ def get_default_llm(max_retries: Optional[int] = None) -> LLMBase:
         llm = AzureLLM.from_env()
     elif provider == LLMProviders.Anthropic.lower():
         llm = AnthropicLLM.from_env()
+    elif provider == LLMProviders.Gemini.lower():
+        llm = GeminiLLM.from_env()
 
     if llm is None:
         raise ValueError(f"Provider {provider} not supported by council.")
@@ -67,5 +72,7 @@ def _build_llm(llm_config: LLMConfigObject) -> LLMBase:
         return OpenAILLM.from_config(llm_config)
     elif provider.is_of_kind(LLMProviders.Anthropic):
         return AnthropicLLM.from_config(llm_config)
+    elif provider.is_of_kind(LLMProviders.Gemini):
+        return GeminiLLM.from_config(llm_config)
 
     raise ValueError(f"Provider `{provider.kind}` not supported by Council")

council/llm/gemini_llm.py

@@ -0,0 +1,81 @@
+from __future__ import annotations
+
+from typing import Any, List, Sequence, Tuple
+
+import google.generativeai as genai  # type: ignore
+from council.contexts import Consumption, LLMContext
+from council.llm import (
+    GeminiLLMConfiguration,
+    LLMBase,
+    LLMConfigObject,
+    LLMMessage,
+    LLMMessageRole,
+    LLMProviders,
+    LLMResult,
+)
+from google.ai.generativelanguage_v1 import HarmCategory  # type: ignore
+from google.generativeai.types import HarmBlockThreshold  # type: ignore
+
+
+class GeminiLLM(LLMBase[GeminiLLMConfiguration]):
+    def __init__(self, config: GeminiLLMConfiguration) -> None:
+        """
+        Initialize a new instance.
+
+        Args:
+            config(GeminiLLMConfiguration): configuration for the instance
+        """
+        super().__init__(name=f"{self.__class__.__name__}", configuration=config)
+        genai.configure(api_key=config.api_key.value)
+        self._model = genai.GenerativeModel(
+            config.model_name(),
+            safety_settings={
+                HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: HarmBlockThreshold.HARM_BLOCK_THRESHOLD_UNSPECIFIED,
+            },
+        )
+
+    def _post_chat_request(self, context: LLMContext, messages: Sequence[LLMMessage], **kwargs: Any) -> LLMResult:
+        history, last = self._to_chat_history(messages=messages)
+        chat = self._model.start_chat(history=history)
+        response = chat.send_message(last)
+        return LLMResult(choices=[response.text], consumptions=self.to_consumptions())
+
+    def to_consumptions(self) -> Sequence[Consumption]:
+        model = self._configuration.model_name()
+        return [
+            Consumption(1, "call", f"{model}"),
+        ]
+
+    @staticmethod
+    def from_env() -> GeminiLLM:
+        """
+        Helper function that create a new instance by getting the configuration from environment variables.
+
+        Returns:
+            GeminiLLM
+        """
+
+        return GeminiLLM(GeminiLLMConfiguration.from_env())
+
+    @staticmethod
+    def from_config(config_object: LLMConfigObject) -> GeminiLLM:
+        provider = config_object.spec.provider
+        if not provider.is_of_kind(LLMProviders.Gemini):
+            raise ValueError(f"Invalid LLM provider, actual {provider}, expected {LLMProviders.Gemini}")
+
+        config = GeminiLLMConfiguration.from_spec(config_object.spec)
+        return GeminiLLM(config=config)
+
+    @staticmethod
+    def _to_chat_history(messages: Sequence[LLMMessage]) -> Tuple[List[Any], Any]:
+        history = []
+        for message in messages[:-1]:
+            if message.is_of_role(LLMMessageRole.System):
+                history.append({"role": "user", "parts": [{"text": f"System Prompt: {message.content}"}]})
+                history.append({"role": "model", "parts": [{"text": "Understood"}]})
+            elif message.is_of_role(LLMMessageRole.User):
+                history.append({"role": "user", "parts": [{"text": message.content}]})
+            elif message.is_of_role(LLMMessageRole.Assistant):
+                history.append({"role": "model", "parts": [{"text": message.content}]})
+        last = messages[-1].content
+        return history, last

council/llm/gemini_llm_configuration.py

@@ -0,0 +1,104 @@
+from __future__ import annotations
+
+from typing import Any, Final, Optional
+
+from council.utils import Parameter, greater_than_validator, not_empty_validator, prefix_validator, read_env_str
+
+from . import LLMConfigSpec, LLMConfigurationBase
+
+_env_var_prefix: Final[str] = "GEMINI_"
+
+
+def _tv(x: float) -> None:
+    """
+    Temperature and Top_p Validators
+    Sampling temperature to use, between 0. and 1.
+    """
+    if x < 0.0 or x > 1.0:
+        raise ValueError("must be in the range [0.0..1.0]")
+
+
+class GeminiLLMConfiguration(LLMConfigurationBase):
+    def __init__(self, model: str, api_key: str) -> None:
+        """
+        Initialize a new instance
+
+        Args:
+            api_key (str): the api key
+            model (str):
+        """
+        super().__init__()
+        self._model = Parameter.string(name="model", required=True, value=model, validator=prefix_validator("gemini-"))
+        self._api_key = Parameter.string(name="api_key", required=True, value=api_key, validator=not_empty_validator)
+        self._temperature = Parameter.float(name="temperature", required=False, default=0.0, validator=_tv)
+        self._top_p = Parameter.float(name="top_p", required=False, validator=_tv)
+        self._top_k = Parameter.int(name="top_k", required=False, validator=greater_than_validator(0))
+
+    def model_name(self) -> str:
+        return self._model.unwrap()
+
+    @property
+    def model(self) -> Parameter[str]:
+        """
+        Gemini model
+        """
+        return self._model
+
+    @property
+    def api_key(self) -> Parameter[str]:
+        """
+        Gemini API Key
+        """
+        return self._api_key
+
+    @property
+    def temperature(self) -> Parameter[float]:
+        """
+        Amount of randomness injected into the response.
+        Ranges from 0 to 1.
+        Use temp closer to 0 for analytical / multiple choice, and closer to 1 for creative and generative tasks.
+        """
+        return self._temperature
+
+    @property
+    def top_p(self) -> Parameter[float]:
+        """
+        Use nucleus sampling.
+        In nucleus sampling, we compute the cumulative distribution over all the options for each subsequent token in
+        decreasing probability order and cut it off once it reaches a particular probability specified by top_p.
+        """
+        return self._top_p
+
+    @property
+    def top_k(self) -> Parameter[int]:
+        """
+        Only sample from the top K options for each subsequent token.
+        Used to remove "long tail" low probability responses.
+        """
+        return self._top_k
+
+    @staticmethod
+    def from_env() -> GeminiLLMConfiguration:
+        api_key = read_env_str(_env_var_prefix + "API_KEY").unwrap()
+        model = read_env_str(_env_var_prefix + "LLM_MODEL").unwrap()
+        config = GeminiLLMConfiguration(model=model, api_key=api_key)
+        return config
+
+    @staticmethod
+    def from_spec(spec: LLMConfigSpec) -> GeminiLLMConfiguration:
+        api_key = spec.provider.must_get_value("apiKey")
+        model = spec.provider.must_get_value("model")
+        config = GeminiLLMConfiguration(model=str(model), api_key=str(api_key))
+
+        if spec.parameters is not None:
+            value: Optional[Any] = spec.parameters.get("temperature", None)
+            if value is not None:
+                config.temperature.set(float(value))
+            value = spec.parameters.get("topP", None)
+            if value is not None:
+                config.top_p.set(float(value))
+            value = spec.parameters.get("topK", None)
+            if value is not None:
+                config.top_k.set(int(value))
+
+        return config

council/llm/llm_config_object.py

@@ -13,6 +13,7 @@ class LLMProviders(str, Enum):
     OpenAI = "openAISpec"
     Azure = "azureSpec"
     Anthropic = "anthropicSpec"
+    Gemini = "googleGeminiSpec"
 
 
 class LLMProvider:
@@ -43,16 +44,21 @@ def from_dict(cls, values: Dict[str, Any]) -> LLMProvider:
         spec = values.get(LLMProviders.Anthropic)
         if spec is not None:
             return LLMProvider(name, description, spec, LLMProviders.Anthropic)
+        spec = values.get(LLMProviders.Gemini)
+        if spec is not None:
+            return LLMProvider(name, description, spec, LLMProviders.Gemini)
         raise ValueError("Unsupported model provider")
 
     def to_dict(self) -> Dict[str, Any]:
         result: Dict[str, Any] = {"name": self.name, "description": self.description}
         if self.is_of_kind(LLMProviders.OpenAI):
             result[LLMProviders.OpenAI] = self._specs
-        if self.is_of_kind(LLMProviders.Azure):
+        elif self.is_of_kind(LLMProviders.Azure):
             result[LLMProviders.Azure] = self._specs
-        if self.is_of_kind(LLMProviders.Anthropic):
+        elif self.is_of_kind(LLMProviders.Anthropic):
             result[LLMProviders.Anthropic] = self._specs
+        elif self.is_of_kind(LLMProviders.Gemini):
+            result[LLMProviders.Gemini] = self._specs
         return result
 
     def must_get_value(self, key: str) -> Any:

dev-requirements.txt

@@ -11,12 +11,12 @@ types-beautifulsoup4~=4.12.0.7
 # Lint
 black==24.4.2
 mypy==1.10.0
-ruff==0.4.8
+ruff==0.4.10
 pylint==3.2.3
 isort==5.13.2
 
 # Test
 ipykernel==6.26.0
-nbconvert==7.11.0
-nbformat==5.9.2
+nbconvert==7.16.4
+nbformat==5.10.4
 pytest==7.4.4

docs/source/reference/llm/gemini_llm.md

@@ -0,0 +1,11 @@
+# AnthropicLLM
+
+```{eval-rst}
+.. autoclasstree:: council.llm.GeminiLLM
+    :full:
+    :namespace: council
+```
+
+```{eval-rst}
+.. autoclass:: council.llm.GeminiLLM
+```

docs/source/reference/llm/gemini_llm_configuration.md

@@ -0,0 +1,5 @@
+# AnthropicLLMConfiguration
+
+```{eval-rst}
+.. autoclass:: council.llm.GeminiLLMConfiguration
+```

requirements.txt

@@ -5,9 +5,12 @@ python-dotenv~=1.0.0
 more-itertools~=10.1.0
 progressbar~=2.5
 tiktoken~=0.7.0
+toml~=0.10.2
+PyYAML~=6.0.1
 
 # LLMs
 anthropic~=0.20.0
+google-generativeai==0.7.0
 
 # Skills
 ## Google
@@ -18,7 +21,3 @@ GoogleNews>=1.6.10
 pymediawiki~=0.7.3
 beautifulsoup4~=4.12.2
 
-toml~=0.10.2
-nbformat~=5.9.2
-nbconvert~=7.11.0
-PyYAML~=6.0.1

tests/data/gemini-llmodel.yaml

@@ -0,0 +1,17 @@
+kind: LLMConfig
+version: 0.1
+metadata:
+  name: an-gemini-deployed-model
+  labels:
+    provider: Google
+spec:
+  description: "Model used to do RST"
+  provider:
+    name: CML-Gemini
+    googleGeminiSpec:
+      model: gemini-pro
+      apiKey:
+        fromEnvVar: GEMINI_API_KEY
+  parameters:
+    temperature: 0.5
+    topK: 8

tests/integration/llm/test_gemini_llm.py

@@ -0,0 +1,35 @@
+import unittest
+
+import dotenv
+from council import LLMContext
+from council.llm import LLMMessage, GeminiLLM
+from council.utils import OsEnviron
+
+
+class TestAnthropicLLM(unittest.TestCase):
+    def test_completion(self):
+        messages = [LLMMessage.user_message("what is the capital of France?")]
+        dotenv.load_dotenv()
+        with OsEnviron("GEMINI_LLM_MODEL", "gemini-1.5-flash"):
+            instance = GeminiLLM.from_env()
+            context = LLMContext.empty()
+            result = instance.post_chat_request(context, messages)
+
+            assert "Paris" in result.choices[0]
+
+    def test_message(self):
+        messages = [LLMMessage.user_message("what is the capital of France?")]
+        dotenv.load_dotenv()
+        with OsEnviron("GEMINI_LLM_MODEL", "gemini-1.0-pro"):
+            instance = GeminiLLM.from_env()
+            context = LLMContext.empty()
+            result = instance.post_chat_request(context, messages)
+
+            assert "Paris" in result.choices[0]
+
+        messages.append(LLMMessage.user_message("give a famous monument of that place"))
+        with OsEnviron("GEMINI_LLM_MODEL", "gemini-1.5-pro"):
+            instance = GeminiLLM.from_env()
+            result = instance.post_chat_request(context, messages)
+
+            assert "Eiffel" in result.choices[0]

tests/unit/__init__.py

@@ -6,6 +6,7 @@ class LLModels:
     OpenAI: str = "openai-llmodel.yaml"
     Anthropic: str = "anthropic-llmodel.yaml"
     AzureWithFallback: str = "azure-with-fallback-llmodel.yaml"
+    Gemini: str = "gemini-llmodel.yaml"
 
 
 class LLMPrompts:

tests/unit/llm/test_gemini_llm_configuration.py

@@ -0,0 +1,22 @@
+import unittest
+from council.llm import GeminiLLMConfiguration
+from council.utils import OsEnviron, ParameterValueException
+
+
+class TestGeminiLLMConfiguration(unittest.TestCase):
+    def test_model_override(self):
+        with OsEnviron("GEMINI_API_KEY", "some-key"), OsEnviron("GEMINI_LLM_MODEL", "gemini-something"):
+            config = GeminiLLMConfiguration.from_env()
+            self.assertEqual("some-key", config.api_key.value)
+            self.assertEqual("gemini-something", config.model.value)
+
+    def test_default(self):
+        config = GeminiLLMConfiguration(model="gemini-something", api_key="some-key")
+        self.assertEqual(0.0, config.temperature.value)
+        self.assertTrue(config.top_p.is_none())
+
+    def test_invalid(self):
+        with self.assertRaises(ParameterValueException):
+            _ = GeminiLLMConfiguration(model="a-gemini-model", api_key="sk-key")
+        with self.assertRaises(ParameterValueException):
+            _ = GeminiLLMConfiguration(model="gemini-model", api_key="")