Hallucination Detector

Verify academic references in LLM-generated text against Semantic Scholar. Extracts references using LLM-based structured extraction (with regex fallback), then checks each title against the Semantic Scholar API with composite scoring: 60% title fuzzy match + 25% author match + 15% journal match. Supports text input or PDF upload. Results stream progressively as each reference is verified.

Open App

---

Overview

The Hallucination Detector addresses fabricated academic references in LLM-generated text. The pipeline is: (1) extract references from the bibliography section using an LLM (OpenAI SDK with structured JSON extraction), falling back to regex parsing if LLM fails; (2) for each reference, search Semantic Scholar's paper search API; (3) compute a composite verification score using rapidfuzz token_set_ratio on title, author last-name matching, and journal/venue matching; (4) classify as "verified" (≥threshold), "partial" (≥60), or "likely hallucinated" (<60).

Data Source

Semantic Scholar API — api.semanticscholar.org/graph/v1/paper/search for paper verification. LLM — OpenAI-compatible endpoint for structured reference extraction. rapidfuzz — token_set_ratio for fuzzy title/journal matching.

Endpoints

POST /hallucination/check_stream — Verify references in text

Upload text or a PDF containing academic references. The detector extracts references, then verifies each against Semantic Scholar. Results stream as NDJSON with a configurable delay between API calls.

curl -X POST "https://atomgpt.org/hallucination/check_stream" \
  -H "Authorization: Bearer sk-XYZ" \
  -F 'text=This paper discusses materials discovery using AI.

References
[1] Kamal Choudhary, Brian DeCost, and Francesca Tavazza. Benchmarking graph neural networks for materials chemistry. npj Computational Materials, 6(1), 2020.
[2] Kamal Choudhary. AtomGPT: Atomistic generative pretrained transformer for forward and inverse materials design. The Journal of Physical Chemistry Letters, 15(27):6909-6917, 2024.
[3] Fake Q. Author. Quantum blockchain neural networks for perpetual motion machines. Nature of Nonsense, 99:1-99, 2099.' \
  -F "top_k=5" \
  -F "threshold=80" \
  -F "sleep_sec=2.0"

Verify a PDF:

curl -X POST "https://atomgpt.org/hallucination/check_stream" \
  -H "Authorization: Bearer sk-XYZ" \
  -F "file=@paper.pdf" \
  -F "top_k=5" \
  -F "threshold=80"

Field	Type	Default	Description
text	string	—	Text with references section (provide text or file)
file	file	—	PDF file to extract and verify
ss_api_key	string	""	Optional Semantic Scholar API key (built-in key used by default)
top_k	int	5	Number of Semantic Scholar results to check per reference (1–20)
threshold	int	80	Composite score threshold for "verified" status (30–100)
sleep_sec	float	2.0	Delay between Semantic Scholar queries to avoid rate limiting (1–60)

Response is streamed as NDJSON (application/x-ndjson) with progressive updates:

1. {"type": "extracting", "message": "Extracting references..."}
2. {"type": "total", "total": N, "method_used": "llm"} — extraction complete
3. {"type": "checking", "index": i, "total": N, "title_guess": "..."} — per-reference progress
4. {"type": "result", "index": i, ...} — per-reference verification result
5. {"type": "done", "total": N, "method_used": "llm"} — all checks complete

Per-reference result fields:

Field	Description
status	"verified" (score ≥ threshold), "partial" (≥ 60), or "likely hallucinated" (< 60)
score	Composite score (0–100): 60% title + 25% author + 15% journal
title_score	Fuzzy title match score alone
author_score	Author last-name match percentage (null if no authors extracted)
journal_score	Journal/venue fuzzy match (null if no journal extracted)
year_match	true/false/null — whether extracted year matches
best_match	Best matching paper title from Semantic Scholar
matched_authors	Author names from the matched paper
matched_venue	Journal/venue from the matched paper
matched_year	Publication year from the matched paper
total_hits	Total Semantic Scholar search results
query	Cleaned search query sent to Semantic Scholar
reference	Original extracted reference text

---

Python Examples

Verify referencesVerify PDFBatch check files

import requests
import json

TEXT = """
This paper builds on prior work in materials AI.

References
[1] Kamal Choudhary et al. The joint automated repository for various
    integrated simulations (JARVIS). npj Computational Materials, 6, 173, 2020.
[2] Fake Q. Author. Quantum blockchain for perpetual motion. Nature of
    Nonsense, 99:1-99, 2099.
"""

response = requests.post(
    "https://atomgpt.org/hallucination/check_stream",
    headers={"Authorization": "Bearer sk-XYZ"},
    data={
        "text": TEXT,
        "top_k": "5",
        "threshold": "80",
        "sleep_sec": "2.0",
    },
    stream=True,
)

verified, flagged = 0, 0
for line in response.iter_lines():
    if not line:
        continue
    chunk = json.loads(line)
    if chunk["type"] == "result":
        status = chunk["status"]
        icon = "✓" if status == "verified" else "~" if status == "partial" else "✗"
        print(f"  {icon} [{chunk['score']:.0f}%] {chunk['reference'][:60]}")
        print(f"     → {chunk['best_match'][:60]}")
        if status == "verified":
            verified += 1
        else:
            flagged += 1
    elif chunk["type"] == "done":
        print(f"\nTotal: {chunk['total']}, Verified: {verified}, Flagged: {flagged}")

import requests
import json

with open("paper.pdf", "rb") as f:
    response = requests.post(
        "https://atomgpt.org/hallucination/check_stream",
        headers={"Authorization": "Bearer sk-XYZ"},
        files={"file": ("paper.pdf", f, "application/pdf")},
        data={"threshold": "75"},
        stream=True,
    )

hallucinated = []
for line in response.iter_lines():
    if not line:
        continue
    chunk = json.loads(line)
    if chunk["type"] == "result" and chunk["status"] == "likely hallucinated":
        hallucinated.append({
            "ref": chunk["reference"][:80],
            "score": chunk["score"],
            "best_match": chunk["best_match"][:60],
        })

if hallucinated:
    print(f"Found {len(hallucinated)} potentially hallucinated references:")
    for h in hallucinated:
        print(f"  ✗ [{h['score']:.0f}%] {h['ref']}")
        print(f"     Best match: {h['best_match']}")
else:
    print("All references verified!")

import requests
import json
import glob

for pdf_path in glob.glob("papers/*.pdf"):
    with open(pdf_path, "rb") as f:
        response = requests.post(
            "https://atomgpt.org/hallucination/check_stream",
            headers={"Authorization": "Bearer sk-XYZ"},
            files={"file": (pdf_path, f, "application/pdf")},
            data={"threshold": "80"},
            stream=True,
        )

    results = {"verified": 0, "partial": 0, "hallucinated": 0}
    for line in response.iter_lines():
        if line:
            chunk = json.loads(line)
            if chunk["type"] == "result":
                if chunk["status"] == "verified":
                    results["verified"] += 1
                elif chunk["status"] == "partial":
                    results["partial"] += 1
                else:
                    results["hallucinated"] += 1

    total = sum(results.values())
    print(f"{pdf_path}: {total} refs — {results}")

AGAPI Agent [WIP]

from agapi.agents import AGAPIAgent
import os

agent = AGAPIAgent(api_key=os.environ.get("AGAPI_KEY"))
response = agent.query_sync("Verify the references in this text for hallucinations")
print(response)

References

• atomgptlab