Joeven — Learn Autonomous AI Agents

← LessonCitations and FaithfulnessJoeven

Python loads on first run

import json
import re

SOURCES = {
    "billing.md": "Refunds for INV-* take 5-7 days. Never refund in cash.",
    "runbook.md": "Runner OOM: raise memory limit on the worker, then replay the DLQ.",
}
RETRIEVED = {"billing.md"}

def normalize(s):
    return re.sub(r"\s+", " ", s).strip().lower()

def quote_in_source(quote, source):
    return normalize(quote) in normalize(SOURCES.get(source, ""))

def check_answer(obj):
    problems = []
    for c in obj.get("citations") or []:
        if c["source"] not in RETRIEVED:
            problems.append("not retrieved " + c["source"])
            continue
        if not quote_in_source(c["quote"], c["source"]):
            problems.append("quote not in " + c["source"])
    ans = obj.get("answer", "").lower()
    if "cash" in ans and "never" not in ans and "not" not in ans:
        problems.append("possible unfaithful cash-refund claim")
    return {"ok": not problems, "problems": problems}

good = {
    "answer": "Refunds take 5-7 days and cannot be paid in cash.",
    "citations": [{"source": "billing.md", "quote": "take 5-7 days"}],
}
bad = {
    "answer": "We can refund you in cash today.",
    "citations": [{"source": "billing.md", "quote": "cash today instantly"}],
}
sneak = {
    "answer": "Raise memory.",
    "citations": [{"source": "runbook.md", "quote": "raise memory limit"}],
}
print("good", json.dumps(check_answer(good)))
print("bad ", json.dumps(check_answer(bad)))
print("sneak", json.dumps(check_answer(sneak)))

Run to execute this in your browser. Nothing is sent to a server.