feat(tools): 🌍 implement live web-search grounding (RAG via DuckDuckGo) to eradicate hallucinations and enforce empirical truth

Author: Arena AI Agent

epistemic_forge/experts/claim_expert.py

@@ -1,36 +1,43 @@
-"""Claim Lattice Expert Implementation."""
+"""Claim Lattice Expert Implementation (Grounded with Real-World Search)."""
 from typing import Dict, Any
 from epistemic_forge.experts.base import EpistemicExpert
 from epistemic_forge.models import ProjectSpec, ClaimLatticeOutput
 from epistemic_forge.llm import generate_structured
+from epistemic_forge.tools.search import search_web
 from loguru import logger
 
 class ClaimLatticeExpert(EpistemicExpert):
     """Deconstructs the question into a structured, epistemically grounded claim lattice."""
 
     @property
     def expert_name(self) -> str:
-        return "Epistemic_Claim_Lattice_Generator"
+        return "Grounded_Claim_Lattice_Generator"
 
     def analyze(self, spec: ProjectSpec, context: Dict[str, Any]) -> ClaimLatticeOutput:
-        """Forces the LLM to generate claims ONLY if it can provide a warrant/explanation."""
+        """Uses Live Web Search to ground the LLM's claims in reality."""
+        
+        # 1. Fetch real-world context before asking the LLM to build claims
+        logger.debug("Gathering live empirical data to prevent hallucination...")
+        search_query = f"{spec.question} scientific consensus"
+        live_evidence = search_web(search_query, max_results=3)
 
         messages = [
             {
                 "role": "system", 
                 "content": (
-                    "You are a rigorous analytical philosopher and scientist. Your task is to break down the user's premise into a 'Claim Lattice'. "
-                    "CRITICAL RULE: You are strictly forbidden from making ANY claim without providing an 'epistemic_warrant' (a clear logical explanation or evidence) "
-                    "and a 'potential_falsifier' (what would prove it wrong). No confident mush allowed."
+                    "You are a rigorous analytical philosopher and empirical scientist. Your task is to break down the user's premise into a 'Claim Lattice'. "
+                    "CRITICAL RULE: You MUST ground your claims using the 'Live Evidence' provided. Do not hallucinate. "
+                    "You are strictly forbidden from making ANY claim without providing an 'epistemic_warrant' (a clear logical explanation) "
+                    "and a 'potential_falsifier'. No confident mush allowed."
                 )
             },
             {
                 "role": "user", 
-                "content": f"Core Premise: {spec.question}\nKeywords: {spec.keywords}\nDeconstruct this into rigorously grounded claims."
+                "content": f"Core Premise: {spec.question}\nKeywords: {spec.keywords}\n\n=== LIVE EMPIRICAL EVIDENCE ===\n{live_evidence}\n=====================\n\nDeconstruct this into rigorously grounded claims, citing the evidence where applicable."
             }
         ]
 
-        logger.debug("Dispatching to LLM for Claim Lattice Generation (Enforcing Truthfulness)...")
+        logger.debug("Dispatching to LLM for Grounded Claim Lattice Generation...")
         return generate_structured(
             messages=messages,
             response_model=ClaimLatticeOutput,

epistemic_forge/tools/search.py

@@ -0,0 +1,25 @@
+"""External Knowledge Retrieval Tool (Web Search).
+
+Provides live grounding for Claim Lattices without requiring paid API keys 
+(using DuckDuckGo Search).
+"""
+from duckduckgo_search import DDGS
+from loguru import logger
+from typing import List, Dict
+
+def search_web(query: str, max_results: int = 3) -> str:
+    """Performs a web search and returns concatenated evidence."""
+    logger.info(f"🔍 Executing Live Web Search for: '{query}'")
+    try:
+        results = DDGS().text(query, max_results=max_results)
+        if not results:
+            return "No external evidence found."
+            
+        evidence = []
+        for r in results:
+            evidence.append(f"[Source: {r.get('title')}]\nSnippet: {r.get('body')}\nURL: {r.get('href')}")
+            
+        return "\n\n".join(evidence)
+    except Exception as e:
+        logger.warning(f"Web search failed: {e}")
+        return "Search tool temporarily unavailable."

pyproject.toml

@@ -31,7 +31,7 @@ classifiers = [
   "Topic :: Scientific/Engineering :: Artificial Intelligence",
   "Topic :: Text Processing :: Linguistic",
 ]
-dependencies = ["litellm>=1.40.0", "pydantic>=2.0.0", "loguru>=0.7.0", "tenacity>=8.0.0", "instructor>=1.3.0", "rich>=13.0.0", "streamlit>=1.35.0"]
+dependencies = ["litellm>=1.40.0", "duckduckgo-search>=5.0.0", "pydantic>=2.0.0", "loguru>=0.7.0", "tenacity>=8.0.0", "instructor>=1.3.0", "rich>=13.0.0", "streamlit>=1.35.0"]
 
 [project.optional-dependencies]
 dev = ["pytest>=7.0"]