Update app.py

app.py

@@ -2,112 +2,119 @@ import os
 import gradio as gr
 from langchain_community.document_loaders import PyPDFLoader
 from langchain_openai import ChatOpenAI, OpenAIEmbeddings
+from langchain.text_splitter import RecursiveCharacterTextSplitter
+from langchain_community.vectorstores import Chroma
+from langchain.chains import RetrievalQA
+from datasets import Dataset
+from ragas import evaluate
+from ragas.metrics import faithfulness, answer_relevancy
 
-# --- DEBUG & SETUP ---
+# --- 1. KEY LOADER & DIAGNOSTICS ---
+# Try to load the key from Hugging Face Secrets
 api_key = os.getenv("OPENAI_API_KEY")
 
-# Check if Key is missing and warn the logs
-if not api_key:
-    print("⚠️ CRITICAL ERROR: OPENAI_API_KEY is missing from Secrets.")
-    # Fallback: Try to grab it if user named it differently
-    api_key = os.getenv("OPENAI_KEY") or os.getenv("key")
-
-# Force set the variable for LangChain
-os.environ["OPENAI_API_KEY"] = str(api_key)
-from langchain_community.document_loaders import PyPDFLoader
-from langchain_text_splitters import RecursiveCharacterTextSplitter
-from langchain_chroma import Chroma
+# Diagnostic: Determine status without revealing the key
+if api_key:
+    key_status = "✅ ACTIVE (Loaded from Secrets)"
+    # FORCE the environment variable for Ragas (which relies on os.environ)
+    os.environ["OPENAI_API_KEY"] = api_key
+else:
+    key_status = "❌ MISSING (Check Settings -> Secrets)"
 
-# UPDATED IMPORT: This is the stable path for RetrievalQA
-from langchain.chains import RetrievalQA
-from langchain.prompts import PromptTemplate
+def audit_rag(pdf_file, user_question):
+    """
+    1. Reads PDF
+    2. Answers Question (using your Key)
+    3. Audits the Answer (using your Key)
+    """
+    if not api_key:
+        return "ERROR: API Key is missing. Please add OPENAI_API_KEY in Settings -> Secrets.", "ERROR", "0", "0"
 
-# Global variables
-vectorstore = None
-qa_chain = None
+    if not pdf_file or not user_question:
+        return "Please upload a PDF and ask a question.", "Waiting for input...", "0.00", "0.00"
 
-def process_pdf(file_path, api_key):
-    global vectorstore, qa_chain
-    
-    if not api_key:
-        return "⚠️ Error: Please enter your OpenAI API Key first."
-    
-    os.environ["OPENAI_API_KEY"] = api_key
-    
     try:
-        # 1. Load PDF
-        loader = PyPDFLoader(file_path)
-        docs = loader.load()
+        # 1. LOAD & PROCESS DOCUMENT
+        loader = PyPDFLoader(pdf_file.name)
+        documents = loader.load()
         
-        # 2. Split Text
         text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=200)
-        splits = text_splitter.split_documents(docs)
+        texts = text_splitter.split_documents(documents)
         
-        # 3. Embed & Store
-        # We explicitly use the lighter embedding model to save cost/time
-        embeddings = OpenAIEmbeddings()
-        vectorstore = Chroma.from_documents(documents=splits, embedding=embeddings)
+        # 2. CREATE RAG ENGINE (Explicitly passing API Key)
+        embeddings = OpenAIEmbeddings(openai_api_key=api_key)
+        db = Chroma.from_documents(texts, embeddings)
+        retriever = db.as_retriever(search_kwargs={"k": 3})
         
-        # 4. Create Chain
-        llm = ChatOpenAI(model_name="gpt-4o", temperature=0)
+        # Explicitly passing API Key to the LLM
+        llm = ChatOpenAI(model_name="gpt-3.5-turbo", temperature=0, openai_api_key=api_key)
         
-        audit_template = """You are Veritas, an AI Compliance Auditor. 
-        Use the following pieces of context to answer the question at the end. 
-        
-        RULES:
-        1. If the answer is in the text, state it clearly.
-        2. If the answer is NOT in the text, you must explicitly state: "FAIL: Information not found in source document."
-        3. Do not hallucinate.
-        
-        Context: {context}
-        
-        Question: {question}
+        qa_chain = RetrievalQA.from_chain_type(
+            llm=llm, 
+            chain_type="stuff", 
+            retriever=retriever,
+            return_source_documents=True
+        )
         
-        Verdict:"""
+        # 3. GENERATE ANSWER
+        result = qa_chain.invoke({"query": user_question})
+        generated_answer = result['result']
+        source_docs = [doc.page_content for doc in result['source_documents']]
         
-        QA_CHAIN_PROMPT = PromptTemplate.from_template(audit_template)
+        # 4. RUN THE AUDIT (RAGAS)
+        # Ragas requires the 'llm' and 'embeddings' to be passed explicitly to avoid config errors
+        data = {
+            'question': [user_question],
+            'answer': [generated_answer],
+            'contexts': [source_docs],
+            'ground_truth': [""] 
+        }
+        dataset = Dataset.from_dict(data)
         
-        qa_chain = RetrievalQA.from_chain_type(
-            llm=llm,
-            chain_type="stuff",
-            retriever=vectorstore.as_retriever(),
-            chain_type_kwargs={"prompt": QA_CHAIN_PROMPT}
+        # Evaluate using the explicitly configured LLM/Embeddings
+        score = evaluate(
+            dataset=dataset,
+            metrics=[faithfulness, answer_relevancy],
+            llm=llm,       # Force Ragas to use our authenticated LLM
+            embeddings=embeddings # Force Ragas to use our authenticated Embeddings
         )
         
-        return "✅ Document Processed. Veritas is ready."
+        audit_results = score.to_pandas()
+        faith_score = audit_results.iloc[0]['faithfulness']
+        relevancy_score = audit_results.iloc[0]['answer_relevancy']
+
+        # 5. GENERATE VERDICT
+        verdict = "✅ PASS" if faith_score > 0.8 else "❌ FAIL (Hallucination Detected)"
         
-    except Exception as e:
-        return f"❌ Error: {str(e)}"
+        return generated_answer, verdict, f"{faith_score:.2f}", f"{relevancy_score:.2f}"
 
-def audit_query(query):
-    global qa_chain
-    if not qa_chain:
-        return "⚠️ Please upload a document first."
-    
-    try:
-        response = qa_chain.invoke(query)
-        return response['result']
     except Exception as e:
-        return f"Error: {str(e)}"
+        return f"System Error: {str(e)}", "ERROR", "0", "0"
 
-# --- INTERFACE ---
+# --- USER INTERFACE ---
 with gr.Blocks(theme=gr.themes.Soft()) as demo:
-    gr.Markdown("# 🛡️ Veritas: AI Compliance Auditor")
+    gr.Markdown("# ⚖️ Veritas: AI Hallucination Auditor")
+    gr.Markdown(f"**System Status:** {key_status}") # Display key status clearly
+    gr.Markdown("Upload a document (e.g., Financial Report) and ask a question. This tool will answer AND verify if the AI stuck to the facts.")
     
     with gr.Row():
         with gr.Column():
-            api_input = gr.Textbox(label="OpenAI API Key", type="password")
-            file_input = gr.File(label="Upload PDF", file_types=[".pdf"])
-            upload_btn = gr.Button("Initialize", variant="primary")
-            status = gr.Textbox(label="Status", interactive=False)
-            
+            file_input = gr.File(label="Upload PDF Evidence", file_types=[".pdf"])
+            question_input = gr.Textbox(label="Cross-Examination Question", placeholder="e.g., What was the net profit in Q3?")
+            submit_btn = gr.Button("Run Audit", variant="primary")
+        
         with gr.Column():
-            query_input = gr.Textbox(label="Audit Query")
-            audit_btn = gr.Button("Run Audit")
-            output = gr.Textbox(label="Verdict", lines=10)
+            answer_output = gr.Textbox(label="AI Witness Testimony (Answer)")
+            with gr.Row():
+                verdict_output = gr.Textbox(label="Verdict")
+                faith_output = gr.Textbox(label="Faithfulness Score (0-1)")
+            relevance_output = gr.Textbox(label="Relevancy Score")
 
-    upload_btn.click(process_pdf, inputs=[file_input, api_input], outputs=status)
-    audit_btn.click(audit_query, inputs=query_input, outputs=output)
+    submit_btn.click(
+        audit_rag, 
+        inputs=[file_input, question_input], 
+        outputs=[answer_output, verdict_output, faith_output, relevance_output]
+    )
 
 if __name__ == "__main__":
     demo.launch()