eval-ai-library 0.3.1__py3-none-any.whl → 0.3.3__py3-none-any.whl

{eval_ai_library-0.3.1.dist-info → eval_ai_library-0.3.3.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: eval-ai-library
-Version: 0.3.1
+Version: 0.3.3
 Summary: Comprehensive AI Model Evaluation Framework with support for multiple LLM providers
 Author-email: Aleksandr Meshkov <alekslynx90@gmail.com>
 License: MIT
@@ -807,6 +807,219 @@ response, cost = await chat_complete(
 )
 ```
 
+## Custom LLM Providers
+
+The library supports custom LLM providers through the `CustomLLMClient` abstract base class. This allows you to integrate any LLM provider, including internal corporate models, locally-hosted models, or custom endpoints.
+
+### Creating a Custom Provider
+
+Implement the `CustomLLMClient` interface:
+```python
+from eval_lib import CustomLLMClient
+from typing import Optional
+from openai import AsyncOpenAI
+
+class InternalLLMClient(CustomLLMClient):
+    """Client for internal corporate LLM or custom endpoint"""
+    
+    def __init__(
+        self,
+        endpoint: str,
+        model: str,
+        api_key: Optional[str] = None,
+        temperature: float = 0.0
+    ):
+        """
+        Args:
+            endpoint: Your internal LLM endpoint URL (e.g., "https://internal-llm.company.com/v1")
+            model: Model name to use
+            api_key: API key if required (optional for local models)
+            temperature: Default temperature
+        """
+        self.endpoint = endpoint
+        self.model = model
+        self.api_key = api_key or "not-needed"  # Some endpoints don't need auth
+        
+        self.client = AsyncOpenAI(
+            api_key=self.api_key,
+            base_url=self.endpoint
+        )
+    
+    async def chat_complete(
+        self,
+        messages: list[dict[str, str]],
+        temperature: float
+    ) -> tuple[str, Optional[float]]:
+        """Generate response from internal LLM"""
+        response = await self.client.chat.completions.create(
+            model=self.model,
+            messages=messages,
+            temperature=temperature,
+        )
+        text = response.choices[0].message.content.strip()
+        cost = None  # Internal models typically don't have API costs
+        return text, cost
+    
+    def get_model_name(self) -> str:
+        """Return model name for logging"""
+        return f"internal:{self.model}"
+```
+
+### Using Custom Providers
+
+Use your custom provider in any metric:
+```python
+import asyncio
+from eval_lib import (
+    evaluate,
+    EvalTestCase,
+    AnswerRelevancyMetric,
+    FaithfulnessMetric
+)
+
+# Create custom internal LLM client
+internal_llm = InternalLLMClient(
+    endpoint="https://internal-llm.company.com/v1",
+    model="company-gpt-v2",
+    api_key="your-internal-key"  # Optional
+)
+
+# Use in metrics
+test_cases = [
+    EvalTestCase(
+        input="What is the capital of France?",
+        actual_output="Paris is the capital.",
+        expected_output="Paris",
+        retrieval_context=["Paris is the capital of France."]
+    )
+]
+
+metrics = [
+    AnswerRelevancyMetric(
+        model=internal_llm,  # ← Your custom LLM
+        threshold=0.7
+    ),
+    FaithfulnessMetric(
+        model=internal_llm,  # ← Same custom client
+        threshold=0.8
+    )
+]
+
+async def run_evaluation():
+    results = await evaluate(
+        test_cases=test_cases,
+        metrics=metrics,
+        verbose=True
+    )
+    return results
+
+asyncio.run(run_evaluation())
+```
+
+### Mixing Standard and Custom Providers
+
+You can mix standard and custom providers in the same evaluation:
+```python
+# Create custom provider
+internal_llm = InternalLLMClient(
+    endpoint="https://internal-llm.company.com/v1",
+    model="company-model"
+)
+
+# Mix standard OpenAI and custom internal LLM
+metrics = [
+    AnswerRelevancyMetric(
+        model="gpt-4o-mini",  # ← Standard OpenAI
+        threshold=0.7
+    ),
+    FaithfulnessMetric(
+        model=internal_llm,  # ← Custom internal LLM
+        threshold=0.8
+    ),
+    ContextualRelevancyMetric(
+        model="anthropic:claude-sonnet-4-0",  # ← Standard Anthropic
+        threshold=0.7
+    )
+]
+
+results = await evaluate(test_cases=test_cases, metrics=metrics)
+```
+
+### Custom Provider Use Cases
+
+**When to use custom providers:**
+
+1. **Internal Corporate LLMs**: Connect to your company's proprietary models
+2. **Local Models**: Integrate locally-hosted models (vLLM, TGI, LM Studio, Ollama with custom setup)
+3. **Fine-tuned Models**: Use your own fine-tuned models hosted anywhere
+4. **Research Models**: Connect to experimental or research models
+5. **Custom Endpoints**: Any LLM accessible via HTTP endpoint
+
+**Example: Local Model with vLLM**
+```python
+# vLLM server running on localhost:8000
+local_model = InternalLLMClient(
+    endpoint="http://localhost:8000/v1",
+    model="meta-llama/Llama-2-7b-chat",
+    api_key=None  # Local models don't need auth
+)
+
+# Use in evaluation
+metric = AnswerRelevancyMetric(model=local_model, threshold=0.7)
+```
+
+**Example: Corporate Internal Model**
+```python
+# Company's internal LLM with authentication
+company_model = InternalLLMClient(
+    endpoint="https://ai-platform.company.internal/api/v1",
+    model="company-gpt-enterprise",
+    api_key="internal-api-key-here"
+)
+
+# Use in evaluation
+metrics = [
+    AnswerRelevancyMetric(model=company_model, threshold=0.7),
+    FaithfulnessMetric(model=company_model, threshold=0.8)
+]
+```
+
+**Key Requirements:**
+
+1. **`async def chat_complete()`** - Must be async and return `(str, Optional[float])`
+2. **`def get_model_name()`** - Return string identifier for logging
+3. **Error Handling** - Handle connection and API errors appropriately
+4. **Cost** - Return `None` for cost if not applicable (e.g., internal/local models)
+
+### Advanced: Custom Authentication
+
+For custom authentication schemes:
+```python
+class CustomAuthLLMClient(CustomLLMClient):
+    """Client with custom authentication"""
+    
+    def __init__(self, endpoint: str, auth_token: str):
+        self.endpoint = endpoint
+        self.headers = {
+            "Authorization": f"Bearer {auth_token}",
+            "X-Custom-Header": "value"
+        }
+        # Use aiohttp or httpx for custom auth
+        import aiohttp
+        self.session = aiohttp.ClientSession(headers=self.headers)
+    
+    async def chat_complete(self, messages, temperature):
+        async with self.session.post(
+            f"{self.endpoint}/chat",
+            json={"messages": messages, "temperature": temperature}
+        ) as response:
+            data = await response.json()
+            return data["content"], None
+    
+    def get_model_name(self):
+        return "custom-auth-model"
+```
+
 ## Test Data Generation
 
 The library includes a powerful test data generator that can create realistic test cases either from scratch or based on your documents.

{eval_ai_library-0.3.1.dist-info → eval_ai_library-0.3.3.dist-info}/RECORD

@@ -1,8 +1,8 @@
-eval_ai_library-0.3.1.dist-info/licenses/LICENSE,sha256=rK9uLDgWNrCHNdp-Zma_XghDE7Fs0u0kDi3WMcmYx6w,1074
-eval_lib/__init__.py,sha256=Jayvtz47_-0POIspT_LJKZ6jmWyf0fQc9fqQ5KvdPRI,3029
+eval_ai_library-0.3.3.dist-info/licenses/LICENSE,sha256=rK9uLDgWNrCHNdp-Zma_XghDE7Fs0u0kDi3WMcmYx6w,1074
+eval_lib/__init__.py,sha256=ySdAQb2DQma2y-ERuFv3VQEAq3S8d8G4vORfo__aqfk,3087
 eval_lib/evaluate.py,sha256=GjlXZb5dnl44LCaJwdkyGCYcC50zoNZn3NrofzNAVJ0,11490
 eval_lib/evaluation_schema.py,sha256=7IDd_uozqewhh7k0p1hKut_20udvRxxkV6thclxKUg0,1904
-eval_lib/llm_client.py,sha256=3eMcarKLkDLDVh4AOxgWbaIzXlzpqsmEfJXNTBonNic,13633
+eval_lib/llm_client.py,sha256=eeTVhCLR1uYbhqOEOSBt3wWPKuzgzA9v8m0F9f-4Gqg,14910
 eval_lib/metric_pattern.py,sha256=wULgMNDeAqJC_Qjglo7bYzY2eGhA_PmY_hA_qGfg0sI,11730
 eval_lib/price.py,sha256=jbmkkUTxPuXrkSHuaJYPl7jSzfDIzQ9p_swWWs26UJ0,1986
 eval_lib/py.typed,sha256=8PjyZ1aVoQpRVvt71muvuq5qE-jTFZkK-GLHkhdebmc,26
@@ -28,7 +28,7 @@ eval_lib/metrics/faithfulness_metric/faithfulness.py,sha256=OqamlhTOps7d-NOStSIK
 eval_lib/metrics/geval/geval.py,sha256=mNciHXnqU2drOJsWlYmbwftGiKM89-Ykw2f6XneIGBM,10629
 eval_lib/metrics/restricted_refusal_metric/restricted_refusal.py,sha256=4QqYgGMcp6W9Lw-v4s0AlUhMSOKvBOEgnLvhqVXaT9I,4286
 eval_lib/metrics/toxicity_metric/toxicity.py,sha256=rBE1_fvpbCRdBpBep1y1LTIhofKR8GD4Eh76EOYzxL0,4076
-eval_ai_library-0.3.1.dist-info/METADATA,sha256=UytyyuWVrL3CuvK7hQC_y-AqoabHEPI0euolxhmfZrQ,37706
-eval_ai_library-0.3.1.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-eval_ai_library-0.3.1.dist-info/top_level.txt,sha256=uQHpEd2XI0oZgq1eCww9zMvVgDJgwXMWkCD45fYUzEg,9
-eval_ai_library-0.3.1.dist-info/RECORD,,
+eval_ai_library-0.3.3.dist-info/METADATA,sha256=S6nodzMnFB5T1Gvtsg19qi1TEwxGtwc9CqLaBWxgPnM,43879
+eval_ai_library-0.3.3.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+eval_ai_library-0.3.3.dist-info/top_level.txt,sha256=uQHpEd2XI0oZgq1eCww9zMvVgDJgwXMWkCD45fYUzEg,9
+eval_ai_library-0.3.3.dist-info/RECORD,,

eval_lib/__init__.py

@@ -7,7 +7,7 @@ A powerful library for evaluating AI models with support for multiple LLM provid
 and a wide range of evaluation metrics for RAG systems and AI agents.
 """
 
-__version__ = "0.3.1"
+__version__ = "0.3.3"
 __author__ = "Aleksandr Meshkov"
 
 # Core evaluation functions
@@ -39,6 +39,7 @@ from eval_lib.llm_client import (
     chat_complete,
     get_embeddings,
     LLMDescriptor,
+    CustomLLMClient,
     Provider
 )
 
@@ -68,12 +69,14 @@ from eval_lib.agent_metrics import (
 
 def __getattr__(name):
     """
-    Ленивый импорт для модулей с тяжёлыми зависимостями.
-    DataGenerator импортируется только когда реально используется.
+    Lazy loading for data generation components.
     """
-    if name == "DataGenerator":
-        from eval_lib.datagenerator.datagenerator import DataGenerator
-        return DataGenerator
+    if name == "DatasetGenerator":
+        from eval_lib.datagenerator.datagenerator import DatasetGenerator
+        return DatasetGenerator
+    if name == "DataGenerator":  # Alias for DatasetGenerator
+        from eval_lib.datagenerator.datagenerator import DatasetGenerator
+        return DatasetGenerator
     if name == "DocumentLoader":
         from eval_lib.datagenerator.document_loader import DocumentLoader
         return DocumentLoader
@@ -104,6 +107,7 @@ __all__ = [
     "chat_complete",
     "get_embeddings",
     "LLMDescriptor",
+    "CustomLLMClient",
     "Provider",
 
     # RAG Metrics

eval_lib/llm_client.py

@@ -2,6 +2,7 @@
 import openai
 import functools
 import anthropic
+from abc import ABC, abstractmethod
 from openai import AsyncAzureOpenAI
 from google import genai
 from google.genai.types import GenerateContentConfig
@@ -13,6 +14,45 @@ from types import SimpleNamespace
 from .price import model_pricing
 
 
+class CustomLLMClient(ABC):
+    """
+    Base class for custom LLM clients.
+    Inherit from this to create your own model implementations.
+
+    Example:
+        class MyCustomLLM(CustomLLMClient):
+            async def chat_complete(self, messages, temperature):
+                # Your implementation
+                return response_text, cost
+
+            def get_model_name(self):
+                return "my-custom-model"
+    """
+
+    @abstractmethod
+    async def chat_complete(
+        self,
+        messages: list[dict[str, str]],
+        temperature: float
+    ) -> tuple[str, Optional[float]]:
+        """
+        Generate a response for the given messages.
+
+        Args:
+            messages: List of message dicts [{"role": "user", "content": "..."}]
+            temperature: Sampling temperature
+
+        Returns:
+            Tuple of (response_text, cost_in_usd)
+        """
+        pass
+
+    @abstractmethod
+    def get_model_name(self) -> str:
+        """Return the model name for logging/tracking purposes."""
+        pass
+
+
 class LLMConfigurationError(Exception):
     """Raised when LLM client configuration is missing or invalid."""
     pass
@@ -24,6 +64,7 @@ class Provider(str, Enum):
     GOOGLE = "google"
     OLLAMA = "ollama"
     ANTHROPIC = "anthropic"
+    CUSTOM = "custom"
 
 
 @dataclass(frozen=True, slots=True)
@@ -308,7 +349,7 @@ _HELPERS = {
 
 
 async def chat_complete(
-    llm: str | tuple[str, str] | LLMDescriptor,
+    llm: str | tuple[str, str] | LLMDescriptor | CustomLLMClient,
     messages: list[dict[str, str]],
     temperature: float = 0.0,
 ):
@@ -327,6 +368,11 @@ async def chat_complete(
         LLMConfigurationError: If required API keys or configuration are missing
         ValueError: If provider is not supported
     """
+    # Handle custom LLM clients
+    if isinstance(llm, CustomLLMClient):
+        return await llm.chat_complete(messages, temperature)
+
+    # Standard providers
     llm = LLMDescriptor.parse(llm)
     helper = _HELPERS.get(llm.provider)