From 7e73c4669c2653fee7bf8f2f55bee1cad9638430 Mon Sep 17 00:00:00 2001
From: Conan Scott <conanscott@gmail.com>
Date: Wed, 1 Jul 2026 04:53:18 +0000
Subject: [PATCH] Add RAG tool Lambda: token resolve + de-identified evidence
 package

---
 lambda_rag/handler.py | 60 +++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 60 insertions(+)
 create mode 100644 lambda_rag/handler.py

diff --git a/lambda_rag/handler.py b/lambda_rag/handler.py
new file mode 100644
index 0000000..0b01f1b
--- /dev/null
+++ b/lambda_rag/handler.py
@@ -0,0 +1,60 @@
+"""
+RAG tool Lambda  --  HNCB de-identification demo (steps 4-6).
+
+The cloud agent calls this tool via AgentCore Gateway with a TOKEN, never a name.
+This function:
+  1. resolves the token against the on-prem vault  ->  real customer_id
+  2. reads the raw customer record (stays here, never returned)
+  3. returns a token-keyed, de-identified evidence package
+
+If the token can't be resolved, it refuses -- it never guesses or leaks.
+"""
+import os
+import boto3
+
+ddb = boto3.resource("dynamodb")
+VAULT = ddb.Table(os.environ["VAULT_TABLE"])
+CUSTOMERS = ddb.Table(os.environ["CUSTOMERS_TABLE"])
+
+
+def _tool_input(event):
+    # AgentCore Gateway passes tool arguments in the event. Accept a few shapes
+    # so this is robust across Gateway target styles during the demo.
+    for key in ("input", "arguments", "parameters", "body"):
+        if isinstance(event.get(key), dict):
+            return event[key]
+    return event
+
+
+def lambda_handler(event, _context):
+    args = _tool_input(event)
+    token = args.get("customer_token")
+    period_days = int(args.get("period_days", 90))
+
+    if not token:
+        return {"error": "customer_token is required"}
+
+    # 1. resolve token -> customer_id (the reversible map lives only here)
+    entry = VAULT.get_item(Key={"token": token}).get("Item")
+    if not entry:
+        return {"error": f"unknown token {token}; refusing to proceed"}
+    customer_id = entry["value"] if entry.get("type") == "CUSTOMER" else entry.get("customer_id")
+    if not customer_id:
+        return {"error": "token did not resolve to a customer record"}
+
+    # 2. read the raw record (NEVER returned to the caller)
+    cust = CUSTOMERS.get_item(Key={"customer_id": customer_id}).get("Item")
+    if not cust:
+        return {"error": "customer record not found"}
+
+    # 3. build a de-identified evidence package (summaries + categories only)
+    return {
+        "customer_token": token,
+        "period": f"last {period_days} days",
+        "activity_summary": cust.get("activity_summary", "n/a"),
+        "amount_band": cust.get("amount_band", "n/a"),
+        "product_type": cust.get("product_type", "n/a"),
+        "risk_profile": cust.get("risk_profile", "n/a"),
+        "sales_constraints": cust.get("sales_constraints", "n/a"),
+        "note": "de-identified; no raw PII in this payload",
+    }