feat(core): 🌐 implement omni-provider engine (OpenAI, Anthropic, Gemini, Ollama) with max flexibility

Author: Arena AI Agent

epistemic_forge/llm.py

@@ -1,43 +1,63 @@
-"""SOTA LLM Engine using Instructor and Pydantic for Strict Structured Outputs.
-ENFORCES: Reproducibility (Seed 42, Temp 0.0) for scientific benchmarks.
+"""Omni-Provider LLM Engine for Absolute Flexibility.
+Supports: OpenAI, Anthropic, Gemini, Ollama, vLLM, Azure, etc.
+Enforces: Strict Pydantic JSON Schemas.
 """
-import instructor
-from openai import OpenAI
 from pydantic import BaseModel
 from loguru import logger
 from tenacity import retry, stop_after_attempt, wait_exponential
+import instructor
+from litellm import completion
 
-try:
-    client = instructor.from_openai(OpenAI())
-except Exception as e:
-    logger.warning(f"OpenAI client init failed (missing key?): {e}")
-    client = None
+def get_instructor_client(model: str):
+    """Dynamically route the instructor client based on the model provider."""
+    import openai
+    import anthropic
+    import google.generativeai as genai
+    
+    if model.startswith("claude"):
+        return instructor.from_anthropic(anthropic.Anthropic())
+    elif model.startswith("gemini"):
+        return instructor.from_gemini(genai.GenerativeModel(model))
+    else:
+        # Default to OpenAI / LiteLLM proxy / Ollama Local
+        return instructor.from_openai(openai.OpenAI())
 
 @retry(stop=stop_after_attempt(3), wait=wait_exponential(multiplier=1, min=2, max=10))
 def generate_structured(
     messages: list, 
     response_model: type[BaseModel], 
     model: str = "gpt-4o-mini", 
     temperature: float = 0.0,
-    seed: int = 42
+    seed: int = 42,
+    api_base: str = None
 ) -> BaseModel:
     """
-    Research-Grade Extraction.
-    Enforces Temperature=0.0 and Seed=42 to guarantee deterministic, reproducible scientific output.
+    Omni-Provider Structured Extraction.
+    Allows passing ANY model (local Ollama, Claude, GPT, Groq).
     """
-    if not client:
-        raise ValueError("LLM Client is not initialized. Please set OPENAI_API_KEY.")
-        
     try:
-        logger.debug(f"Initiating scientifically rigorous call to {model} [temp={temperature}, seed={seed}] for schema [{response_model.__name__}]...")
-        response = client.chat.completions.create(
-            model=model,
-            messages=messages,
-            response_model=response_model,
-            temperature=temperature,
-            seed=seed
-        )
+        logger.debug(f"Routing neural call to [{model}] for schema [{response_model.__name__}]...")
+        
+        # We use instructor's dynamic client routing
+        client = get_instructor_client(model)
+        
+        kwargs = {
+            "model": model,
+            "messages": messages,
+            "response_model": response_model,
+            "temperature": temperature,
+        }
+        
+        # Only inject seed if the provider supports it (like OpenAI)
+        if "gpt" in model or "llama" in model:
+            kwargs["seed"] = seed
+            
+        if api_base: # For Local Ollama or vLLM routing
+            kwargs["base_url"] = api_base
+            
+        response = client.chat.completions.create(**kwargs)
         return response
+        
     except Exception as e:
-        logger.error(f"SOTA LLM API Critical Failure: {str(e)}")
+        logger.error(f"Omni-Provider API Failure for {model}: {str(e)}")
         raise

pyproject.toml

@@ -31,7 +31,7 @@ classifiers = [
   "Topic :: Scientific/Engineering :: Artificial Intelligence",
   "Topic :: Text Processing :: Linguistic",
 ]
-dependencies = ["instructor>=1.3.0", "openai>=1.30.0", "pydantic>=2.0.0", "loguru>=0.7.0", "tenacity>=8.0.0"]
+dependencies = ["litellm>=1.0.0", "pydantic>=2.0.0", "loguru>=0.7.0", "tenacity>=8.0.0", "instructor>=1.3.0"]
 
 [project.optional-dependencies]
 dev = ["pytest>=7.0"]