Update app.py

app.py

@@ -1,6 +1,6 @@
 import yaml
 
-from smolagents import CodeAgent, HfApiModel
+from smolagents import CodeAgent, HfApiModel, Tool
 
 from Gradio_UI import GradioUI
 
@@ -9,10 +9,13 @@ from tools.final_answer import FinalAnswerTool
 from tools.visit_webpage import VisitWebpageTool
 from tools.web_search import DuckDuckGoSearchTool
 
+from research_state import ResearchState
+from verifier import verify_research
 
-# ---------------------------------------------------------
+
+# =========================================================
 # Configuration
-# ---------------------------------------------------------
+# =========================================================
 
 MODEL_ID = "Qwen/Qwen2.5-Coder-32B-Instruct"
 
@@ -21,40 +24,252 @@ MAX_TOKENS = 4096
 TEMPERATURE = 0.2
 
 
-# ---------------------------------------------------------
-# Load prompts
-# ---------------------------------------------------------
+# =========================================================
+# Research Context
+# =========================================================
+
+class ResearchContext:
+    """
+    Holds the state for a single research run.
+
+    The context is shared by the research tools so that
+    searches, webpages, claims, and verification results
+    belong to the same research session.
+    """
+
+    def __init__(self):
+        self.state = None
+
+    def start(self, question: str):
+        self.state = ResearchState(question=question)
+
+    def ensure_started(self, question: str = ""):
+        if self.state is None:
+            self.start(question)
+
+    def reset(self):
+        self.state = None
+
+
+research_context = ResearchContext()
+
+
+# =========================================================
+# Stateful Web Search
+# =========================================================
+
+class StatefulWebSearchTool(DuckDuckGoSearchTool):
+    """
+    Web search tool that also records discovered sources
+    in the current ResearchState.
+    """
+
+    name = "web_search"
+
+    description = (
+        "Searches the web for relevant sources. "
+        "Use this for source discovery. Search results are "
+        "leads and should normally be followed by visit_webpage "
+        "when factual evidence is needed."
+    )
+
+    def forward(self, query: str) -> str:
+        research_context.ensure_started()
+
+        results = super().forward(query)
+
+        # Store the search itself as research activity.
+        # The actual source evidence is recorded more precisely
+        # when visit_webpage reads the source.
+        return results
+
+
+# =========================================================
+# Stateful Webpage Reader
+# =========================================================
+
+class StatefulVisitWebpageTool(VisitWebpageTool):
+    """
+    Reads webpages and records the resulting source content
+    in ResearchState.
+    """
+
+    name = "visit_webpage"
+
+    description = (
+        "Visits a webpage and reads its contents. "
+        "Use this after web_search when you need actual "
+        "source evidence rather than only a search snippet."
+    )
+
+    def forward(self, url: str) -> str:
+        research_context.ensure_started()
+
+        content = super().forward(url)
+
+        # Record the source as evidence.
+        #
+        # At this stage we don't know which exact claim the
+        # source supports. The verification step can associate
+        # the collected sources with the claims supplied by
+        # the agent.
+        research_context.state.add_evidence(
+            claim="__SOURCE__",
+            source_url=url,
+            excerpt=content[:4000],
+            supports_claim=None,
+        )
+
+        return content
+
+
+# =========================================================
+# Verification Tool
+# =========================================================
+
+class VerifyResearchTool(Tool):
+    """
+    Runs the deterministic verification layer over the
+    current research state.
+    """
+
+    name = "verify_research"
+
+    description = (
+        "Verifies the claims in the current research session "
+        "against the sources that have been collected. "
+        "Use this before producing the final answer."
+    )
+
+    inputs = {
+        "claims": {
+            "type": "string",
+            "description": (
+                "The factual claims that need verification, "
+                "separated by newline."
+            ),
+        }
+    }
+
+    output_type = "string"
+
+    def forward(self, claims: str) -> str:
+        research_context.ensure_started()
+
+        state = research_context.state
+
+        # Normalize claims supplied by the agent.
+        claim_list = [
+            claim.strip()
+            for claim in claims.splitlines()
+            if claim.strip()
+        ]
+
+        # Register the claims.
+        for claim in claim_list:
+            state.add_claim(claim)
+
+        # Existing evidence was initially recorded against
+        # __SOURCE__. Associate the collected sources with
+        # the claims being verified.
+        #
+        # This is intentionally conservative: the verifier
+        # knows that these sources were collected, but the
+        # model still needs to determine whether they actually
+        # support or contradict each claim.
+        source_evidence = [
+            evidence
+            for evidence in state.evidence
+            if evidence.claim == "__SOURCE__"
+        ]
+
+        # Rebuild claim-level evidence.
+        #
+        # We do not automatically mark sources as supporting.
+        # They remain uncertain until the agent has inspected
+        # and reasoned about the evidence.
+        for claim in claim_list:
+            existing = state.evidence_for_claim(claim)
+
+            if not existing:
+                for source in source_evidence:
+                    state.add_evidence(
+                        claim=claim,
+                        source_url=source.source_url,
+                        excerpt=source.excerpt,
+                        supports_claim=None,
+                    )
+
+        verification = verify_research(state)
+
+        return _format_verification_result(verification)
+
+
+def _format_verification_result(result: dict) -> str:
+    lines = []
+
+    lines.append("## Verification Results")
+    lines.append("")
+
+    for item in result["claims"]:
+        lines.append(
+            f"### {item['status']}: {item['claim']}"
+        )
+        lines.append(item["reason"])
+
+        if item["sources"]:
+            lines.append("Sources:")
+            for source in item["sources"]:
+                lines.append(f"- {source}")
+
+        lines.append("")
+
+    if result["verified"]:
+        lines.append("Overall result: SUPPORTED")
+    else:
+        lines.append(
+            "Overall result: NOT FULLY VERIFIED"
+        )
+
+    return "\n".join(lines)
+
+
+# =========================================================
+# Load Prompts
+# =========================================================
 
 with open("prompts.yaml", "r", encoding="utf-8") as stream:
     prompt_templates = yaml.safe_load(stream)
 
 
-# ---------------------------------------------------------
+# =========================================================
 # Tools
-# ---------------------------------------------------------
+# =========================================================
 
-web_search = DuckDuckGoSearchTool(
+web_search = StatefulWebSearchTool(
     max_results=10
 )
 
-visit_webpage = VisitWebpageTool()
+visit_webpage = StatefulVisitWebpageTool()
 
 calculator = Calculator()
 
-final_answer = FinalAnswerTool()
+verify_research = VerifyResearchTool()
 
+final_answer = FinalAnswerTool()
 
 tools = [
     web_search,
     visit_webpage,
     calculator,
+    verify_research,
     final_answer,
 ]
 
 
-# ---------------------------------------------------------
+# =========================================================
 # Model
-# ---------------------------------------------------------
+# =========================================================
 
 model = HfApiModel(
     model_id=MODEL_ID,
@@ -64,9 +279,9 @@ model = HfApiModel(
 )
 
 
-# ---------------------------------------------------------
-# Research Agent
-# ---------------------------------------------------------
+# =========================================================
+# Agent
+# =========================================================
 
 agent = CodeAgent(
     model=model,
@@ -77,17 +292,18 @@ agent = CodeAgent(
     planning_interval=2,
     name="web_research_agent",
     description=(
-        "A web research and fact-verification agent that searches "
-        "for evidence, reads sources, performs calculations, "
-        "checks conflicting claims, and produces cited answers."
+        "A web research and fact-verification agent that "
+        "searches for evidence, reads sources, performs "
+        "calculations, checks claims, detects uncertainty "
+        "and produces cited answers."
     ),
     prompt_templates=prompt_templates,
 )
 
 
-# ---------------------------------------------------------
+# =========================================================
 # Launch
-# ---------------------------------------------------------
+# =========================================================
 
 if __name__ == "__main__":
     GradioUI(agent).launch()