Update app.py

app.py

@@ -10,6 +10,8 @@ import pandas as pd
 from langchain_community.chat_models import ChatOpenAI
 from langchain.agents import initialize_agent, Tool, AgentType
 
+from fuzzywuzzy import fuzz
+
 st.set_page_config(page_title="Accounts Payable AI Agent", layout="wide")
 
 # -------- LLM Model Setup --------
@@ -256,38 +258,70 @@ def po_match_tool_func(input_text):
     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")
+    inv_total = inv_hdr.get("total_due") or inv_hdr.get("invoice_value") or inv_hdr.get("total_before_tax")
     inv_line_items = invoice.get("line_items", [])
 
-    # Try to match PO by number or supplier name
-    matched_po = None
-    explanation = ""
+    explanation = []
+    best_match = None
+    best_match_type = None
+
     for idx, row in po_df.iterrows():
-        po_number = str(row.get("PO Number", "")).lower().replace(" ", "")
-        supplier = str(row.get("Supplier Name", "")).lower().strip()
-        if inv_po_number and po_number == str(inv_po_number).lower().replace(" ", ""):
-            matched_po = row
-            explanation += f"Matched on PO Number: {inv_po_number}. "
+        po_number = str(row.get("PO Number", "")).strip().lower()
+        supplier = str(row.get("Supplier Name", "")).strip().lower()
+        po_total = str(row.get("Total PO Value", "")).replace(",", "").replace("$", "").strip().lower()
+
+        po_match = po_number and inv_po_number and (po_number == str(inv_po_number).strip().lower())
+        supplier_match = supplier and inv_supplier and (supplier == inv_supplier.strip().lower())
+        total_match = po_total and inv_total and (po_total == str(inv_total).replace(",", "").replace("$", "").strip().lower())
+
+        # Strong approval condition
+        if po_match and supplier_match and total_match:
+            best_match = row
+            best_match_type = "APPROVED"
+            explanation.append(f"PO Number, Supplier Name, and Total Value all matched. PO: {row.to_dict()}")
             break
-        elif inv_supplier and supplier == inv_supplier.lower().strip():
-            matched_po = row
-            explanation += f"Matched on Supplier Name: {inv_supplier}. "
+        # Partial approval
+        elif (po_match or supplier_match) and not total_match:
+            best_match = row
+            best_match_type = "PARTIALLY APPROVED"
+            fields = []
+            if po_match:
+                fields.append("PO Number matched")
+            if supplier_match:
+                fields.append("Supplier Name matched")
+            explanation.append(f"{' and '.join(fields)}, but Total Value did not match. PO: {row.to_dict()}")
             break
 
-    # If no direct match, try to match by line items
-    if matched_po is None and len(inv_line_items) > 0:
+    # If no direct match, try line item fuzzy matching
+    if best_match is None and len(inv_line_items) > 0:
         for idx, row in po_df.iterrows():
             po_desc = str(row.get("Description", "")).lower()
+            po_total = str(row.get("Total PO Value", "")).replace(",", "").replace("$", "").strip().lower()
+            line_item_matched = False
             for line in inv_line_items:
-                if line.get("description") and line["description"].lower() in po_desc:
-                    matched_po = row
-                    explanation += f"Matched on line item description: '{line['description']}'. "
+                desc = line.get("description") or ""
+                qty = str(line.get("quantity") or "").strip()
+                price = str(line.get("price") or "").replace(",", "").replace("$", "").strip()
+                score = fuzz.token_set_ratio(desc.lower(), po_desc)
+                if (desc and score >= 80) or (qty and qty in po_desc) or (price and price in po_desc):
+                    line_item_matched = True
+                    explanation.append(f"Line item '{desc}' (qty: {qty}, price: {price}) matched PO description with score {score}. PO: {row.to_dict()}")
                     break
-            if matched_po is not None:
+            if line_item_matched and po_total and inv_total and po_total == str(inv_total).replace(",", "").replace("$", "").strip().lower():
+                best_match = row
+                best_match_type = "APPROVED"
                 break
 
-    if matched_po is not None:
-        return f"PO matched: {matched_po.to_dict()}. {explanation}"
-    return "No matching PO found based on PO Number, Supplier, or Line Items."
+    if best_match is not None:
+        return json.dumps({
+            "decision": best_match_type,
+            "reason": " | ".join(explanation)
+        })
+    else:
+        return json.dumps({
+            "decision": "REJECTED",
+            "reason": "No match found on PO Number, Supplier Name, Total Value, or any line item (including fuzzy match)."
+        })
 
 if po_df is not None:
     st.session_state["po_df"] = po_df
@@ -306,7 +340,7 @@ if extracted_info is not None and po_df is not None:
             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, including by line items.",
+                description="Check if the invoice matches any PO (headers or fuzzy line items).",
             )
         ]
         decision_llm = ChatOpenAI(
@@ -322,16 +356,13 @@ if extracted_info is not None and po_df is not None:
             verbose=True,
         )
         prompt = (
-            "You are an expert accounts payable decision agent.\n"
-            "You are given an extracted invoice in JSON and have access to a tool called po_match_tool, which can check for matches with all available POs (including matching line items/descriptions between invoice and PO).\n"
-            "To approve an invoice, you must verify at least one of the following:\n"
-            "- The PO number matches a PO\n"
-            "- The supplier name matches a PO\n"
-            "- At least one line item description or quantity/price matches with a PO's item\n"
-            "If you can't match on PO number or supplier, do your best to match using the invoice's line items (description/quantity/unit price/etc) and the PO data, and explain your reasoning step by step."
-            "In your reasoning, list all fields and line items that matched, or say if nothing matched (be specific about what was compared)."
-            "At the end, respond in this JSON format ONLY:\n"
-            '{"decision": "APPROVED or REJECTED", "reason": "<detailed step-by-step explanation for your decision, showing what matched and what did not, including line item checks>"}\n'
+            "You are an expert accounts payable agent. "
+            "Use po_match_tool to check matches based on the following business rules:\n"
+            "- If PO Number AND Supplier Name AND Total Value all match, the invoice is APPROVED.\n"
+            "- If PO Number OR Supplier Name match, but Total Value does not, the invoice is PARTIALLY APPROVED.\n"
+            "- If neither, try matching at least one line item (by fuzzy description, quantity, or price) and require total to match for APPROVED.\n"
+            "- Otherwise, REJECTED.\n"
+            "Call the tool and return its result as-is. Do not invent or guess the answer, do not add any comments outside the JSON.\n"
             f"Invoice JSON:\n{json.dumps(extracted_info, indent=2)}"
         )
         with st.spinner("AI is reasoning and making a decision..."):