Grounded Generation & Citations

You retrieved the right chunks and ranked them first. The model can still ignore them, mix them with what it remembers from training, or state something no chunk supports. Generation is the last stage, and its failure, a confident claim with nothing behind it, is the one users see.

The promise of RAG is faithfulness: every claim in the answer comes from the retrieved context, with nothing added from the model's memory. A faithful answer can still be wrong if the context is wrong, but that's a retrieval problem you can fix. An unfaithful answer asserts something the context doesn't say, and good retrieval can't prevent it.

Three things that make generation faithful

A grounding instruction. Tell the model to answer only from the context and to say it doesn't know when the context doesn't cover the question. Models default to being helpful and filling gaps; you have to give them permission to abstain.

Inline citations. Ask the model to tag each claim with the chunk it came from, like "[c12]." Users can then check the source, and a claim with no valid supporting chunk becomes something you can detect and count.

Context order. Because of lost-in-the-middle, put the highest-ranked chunks at the start and end of the context rather than in the middle.

Measuring faithfulness

Split the answer into individual claims, and for each one, check whether any retrieved chunk supports it. The share of supported claims is the faithfulness score. At scale, a model judge usually does the checking, which is where RAG work meets evaluation work.

Starter code

# Each item: the context the model was given, and the claims its answer made
# (already decomposed). 'key' is the checkable fact substring for each claim.
ITEMS = [
  {"context": "The /export endpoint is rate limited to 20 requests per minute. "
              "Keys rotate every 90 days.",
   "claims": [ {"text":"/export allows 20 req/min", "key":"20 requests per minute"},
               {"text":"keys rotate every 90 days", "key":"90 days"} ]},
  {"context": "The free tier allows 3 projects. Webhooks retry up to 5 times.",
   "claims": [ {"text":"free tier allows 3 projects", "key":"3 projects"},
               {"text":"there is a 30-day money-back guarantee", "key":"30-day money-back"} ]},  # unsupported!
]

def supported(claim, context):
    return claim["key"].lower() in context.lower()

# TODO: per item, compute faithfulness = supported/total; print unsupported claims;
#       print overall faithfulness across all claims.

total = supported_total = 0
for item in ITEMS:
    n = len(item["claims"])
    s = sum(supported(c, item["context"]) for c in item["claims"])
    total += n; supported_total += s
    print(f"faithfulness {s}/{n}")
    for c in item["claims"]:
        if not supported(c, item["context"]):
            print(f"   UNSUPPORTED: {c['text']}")

print(f"overall faithfulness: {supported_total/total:.2f}")