Skip to content
Notifly Notifly

RAG answer with no sources (grounding lost)

The RAG pipeline finds relevant chunks, puts them in the context — and the model answers off the mark: no links, no [source] markers, inventing on top of nothing. Retrieval did its job: the chunks were there. The answer looks confident and plausible, so the regression is quiet — the user simply gets “facts” that aren’t in your database.

This is a signal distinct from hallucinations: there the model invents when the context is empty, here it ignores a non-empty one. The cause is almost always in three places: the retriever broke, the embeddings version changed and the old chunks are no longer found, or someone edited the system prompt and dropped the “cite your sources” line.

1. A cheap structural check

Section titled “1. A cheap structural check”

The cheapest option — without a single extra model call. If retrieval returned chunks but the answer cites none of their ids (no [source], [1], doc_id in the text), we treat the answer as “ungrounded”. We accumulate the share of such answers in a rolling window and send a push when it crosses a threshold.

import os, re, json, time, statistics, requests


WIN = []                       # 1 = ungrounded, 0 = ok
FLAG = "/tmp/citation_spike.flag"
THRESHOLD = 0.30               # >30% of answers with no sources over the window
MIN_WINDOW = 40


def is_ungrounded(answer_text, retrieved_ids):
    # retrieval found something, but the answer cited none of the ids?
    if not retrieved_ids:
        return False           # empty retrieval — that's a different scenario
    cited = set(re.findall(r"\[(?:source|src|doc)[:\s]*([\w\-]+)\]", answer_text))
    cited |= set(re.findall(r"\[(\d+)\]", answer_text))   # footnotes like [1]
    return len(cited & {str(i) for i in retrieved_ids}) == 0


def observe(answer_text, retrieved_ids):
    WIN.append(1 if is_ungrounded(answer_text, retrieved_ids) else 0)
    if len(WIN) > 200:
        WIN.pop(0)
    if len(WIN) < MIN_WINDOW:
        return
    rate = statistics.mean(WIN)
    if rate >= THRESHOLD and not os.path.exists(FLAG):
        open(FLAG, "w").write(str(time.time()))
        notify("📄❓ Ответы без источников",
               f"{int(rate*100)}% ответов не цитируют найденные чанки "
               f"(окно {len(WIN)}). Проверьте retriever / промпт.",
               priority=8)
    elif rate < THRESHOLD * 0.5 and os.path.exists(FLAG):
        os.remove(FLAG)        # the spike has passed — clear the flag


def notify(title, message, priority):
    requests.post(f"{os.environ['NOTIFLY_URL']}/message",
                  params={"token": os.environ["NOTIFLY_TOKEN"]},
                  json={"title": title, "message": message, "priority": priority},
                  timeout=5)

The flag file guarantees one push per spike — otherwise every answer in the window would fire a new alert. The flag is cleared once the share returns to normal.

2. A grounding score via an LLM judge

Section titled “2. A grounding score via an LLM judge”

The structural check catches “no markers at all” but misses an answer that cites a source formally while actually paraphrasing from the model’s own head. A more precise measure is a grounding score: do the answer’s claims trace back to the retrieved context (0..1)? This is more expensive, so we run it on a sample (say, every 10th answer) and watch the window mean.

import anthropic


client = anthropic.Anthropic()
GWIN = []
GTHRESHOLD = 0.6               # mean grounding dropped below 0.6 → alert


def grounding_score(answer, context_chunks):
    ctx = "\n---\n".join(context_chunks)
    judge = client.messages.create(
        model="claude-haiku-4-5",
        max_tokens=8,
        messages=[{"role": "user", "content":
            f"Context:\n{ctx}\n\nAnswer:\n{answer}\n\n"
            "What fraction of the answer's claims follows directly from the "
            "context? Return only a number from 0.0 to 1.0."}],
    )
    try:
        return max(0.0, min(1.0, float(judge.content[0].text.strip())))
    except ValueError:
        return 0.0


def observe_grounding(answer, context_chunks):
    GWIN.append(grounding_score(answer, context_chunks))
    if len(GWIN) > 100:
        GWIN.pop(0)
    if len(GWIN) >= 20:
        mean = statistics.mean(GWIN)
        if mean < GTHRESHOLD and not os.path.exists("/tmp/grounding.flag"):
            open("/tmp/grounding.flag", "w").write("1")
            notify("🧷 Grounding упал",
                   f"Средний grounding {mean:.2f} за {len(GWIN)} ответов "
                   f"(порог {GTHRESHOLD}). Ответы плохо опираются на контекст.",
                   priority=8)

Keep the judge model cheap — the task is binary-simple, no reason to pay for a flagship here.

3. A hard per-answer alert for high-stakes

Section titled “3. A hard per-answer alert for high-stakes”

For answers where the cost of a mistake is high — legal, medical, financial — we don’t wait to accumulate statistics. A single ungrounded answer in such a category is an immediate loud push, so a human looks before the answer reaches the user.

HIGH_STAKES = {"legal", "medical", "finance"}


def guard_high_stakes(answer_text, retrieved_ids, category):
    if category in HIGH_STAKES and is_ungrounded(answer_text, retrieved_ids):
        notify("🚨 High-stakes ответ без источника",
               f"Категория «{category}»: ответ без цитат при "
               f"{len(retrieved_ids)} найденных чанках. Нужен ревью.",
               priority=10)
        return False           # block the answer until it's reviewed
    return True

What to put in the push

Section titled “What to put in the push”
  • the ungrounded share / mean grounding score over the window;
  • how many chunks retrieval returned (to tell “empty” from “ignored”);
  • the category, if it’s high-stakes;
  • a hint about the cause: embeddings version changed?, system prompt edited?, retriever returns empty?.

If the spike coincides in time with a deploy, the culprit is almost certainly an embeddings update or a prompt edit. Split the checks into layers and run the counters in shared storage (Redis) so the window survives restarts and covers all instances.

Section titled “Related recipes”