Update app.py

app.py

@@ -7,12 +7,10 @@ import time
 import mimetypes
 import pandas as pd
 
-# LangChain & LangGraph
 from langchain_community.chat_models import ChatOpenAI
-from langgraph.graph import StateGraph, END
-from langgraph.prebuilt import create_react_agent
+from langchain.agents import initialize_agent, Tool, AgentType
 
-st.set_page_config(page_title="PDF Tools", layout="wide")
+st.set_page_config(page_title="Accounts Payable AI Agent", layout="wide")
 
 # -------- LLM Model Setup --------
 MODELS = {
@@ -254,7 +252,7 @@ def extract_text_from_unstract(uploaded_file):
     except Exception:
         return r.text
 
-# --------- NEW: UPLOAD PO CSV ---------
+# --------- Upload PO CSV ---------
 st.sidebar.header("Step 1: Upload Active Purchase Orders (POs)")
 po_file = st.sidebar.file_uploader(
     "Upload POs CSV (must include PO number, Supplier, Items, etc.)",
@@ -290,13 +288,16 @@ if st.button("Extract") and inv_file:
 
 extracted_info = extracted_info or st.session_state.get("last_extracted_info", None)
 
-# -------------------------------
-# LANGGRAPH ReAct DECISION AGENT
-# -------------------------------
+# --------- Classic ReAct AGENT ---------
+def po_match_tool_func(input_text):
+    # This tool receives a string "query" (not structured context!) so it must access globals or session
+    # We'll use the extracted_info and po_df from session
+    invoice = st.session_state.get("last_extracted_info")
+    po_df = st.session_state.get("po_df")
+
+    if invoice is None or po_df is None:
+        return "Invoice or PO data not found."
 
-def po_match_tool(query: str, context: dict):
-    invoice = context['invoice']
-    po_df = context['po_df']
     inv_hdr = invoice["invoice_header"]
     inv_po_number = inv_hdr.get("purchase_order_number") or inv_hdr.get("order_number") or inv_hdr.get("our_order_number")
     inv_supplier = inv_hdr.get("supplier_name")
@@ -319,45 +320,44 @@ def po_match_tool(query: str, context: dict):
         return f"PO matched: {matched_po.to_dict()}"
     return "No matching PO found."
 
-def build_decision_agent():
-    openai_api_key = os.getenv("OPENAI_API_KEY")
-    llm = ChatOpenAI(
-        openai_api_key=openai_api_key,
-        model="gpt-4-1106-preview",
-        temperature=0,
-        streaming=False,
-        # No 'tools' param, classic ReAct only!
-    )
-    tools = [
-        {
-            "name": "po_match_tool",
-            "description": "Looks up a PO for a given invoice context.",
-            "func": po_match_tool,
-        }
-    ]
-    agent = create_react_agent(llm, tools)
-    graph_builder = StateGraph(agent)
-    def finish_decision(state, context):
-        return END, state
-    graph_builder.add_node("finish", finish_decision)
-    graph_builder.set_entry_point(agent)
-    graph_builder.add_edge(agent, END)
-    return graph_builder.compile()
+# Save PO df to session for tool access
+if po_df is not None:
+    st.session_state["po_df"] = po_df
 
 if extracted_info is not None and po_df is not None:
     if st.button("Make a decision (AI Agent)"):
-        with st.spinner("Reasoning and making a decision with LangGraph agent..."):
-            agent_graph = build_decision_agent()
-            task = (
-                "Here is an invoice JSON and a list of active POs in context. "
-                "Step by step, reason whether the invoice matches an active PO and can be approved. "
-                "If there is a match, state the matched PO, otherwise explain why not. "
-                "Give a clear final decision: APPROVED or REJECTED."
+        # Define the tool for the agent (takes input string, outputs string)
+        tools = [
+            Tool(
+                name="po_match_tool",
+                func=po_match_tool_func,
+                description="Use this tool to check if the invoice matches any PO in the current PO list.",
             )
-            context = {
-                "invoice": extracted_info,
-                "po_df": po_df,
-            }
-            out = agent_graph.invoke(task, context=context)
+        ]
+        # Use session state for the LLM agent
+        openai_api_key = os.getenv("OPENAI_API_KEY")
+        llm = ChatOpenAI(
+            openai_api_key=openai_api_key,
+            model="gpt-4-1106-preview",
+            temperature=0,
+            streaming=False,
+        )
+        agent = initialize_agent(
+            tools,
+            llm,
+            agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION,
+            verbose=True,
+        )
+        # Build the prompt
+        prompt = (
+            f"Below is an extracted invoice in JSON and a list of active POs is loaded in the system. "
+            f"Use po_match_tool to check if the invoice matches an active PO. "
+            f"Step by step, reason whether the invoice matches an active PO and can be approved. "
+            f"If there is a match, state the matched PO, otherwise explain why not. "
+            f"Give a clear final decision: APPROVED or REJECTED.\n"
+            f"Invoice JSON:\n{json.dumps(extracted_info, indent=2)}"
+        )
+        with st.spinner("AI is reasoning and making a decision..."):
+            result = agent.run(prompt)
         st.subheader("AI Decision")
-        st.write(out)
+        st.write(result)