Literature search with NASA ADS and a language model

Build the query yourself, export abstracts from ADS, and let the model map the field with every claim tied to a bibcode you can check.

Author

Affiliation

Francesco De Gasperin

INAF

AUTHOR = 'Francesco De Gasperin' MODEL = 'to be tested' TOOL = 'Any chat model + ADS API' TESTED = 'not yet' QFLAG = 'L0' TIMESAVE= 'to be measured' PHASE = 'Literature' END

What is this header?

NoteDraft

This page shows the structure of a workflow. The steps are a proposal and have not yet been tested and timed. Once tested, the header above will record the model, the date and the verification level.

Goal

Get a reliable first map of a topic (main results, open questions, who works on what) from a set of papers you chose, without the model inventing references.

Use it when you are entering a new subfield or writing the introduction of a paper or proposal. Don’t use it as a substitute for reading the key papers, or when the relevant results are only in the body of papers rather than in abstracts.

What you need

• A free NASA ADS account and an API token
• Python with requests
• A chat model that accepts file uploads or long input

Steps

Write the query yourself

The model should not decide which papers count. Build and refine the search in the ADS interface first, for example:

abs:"radio halo" year:2018-2026 property:refereed

Sort the results by citation count, as the export script below does, and check the first results by eye. Adjust the query until most of the top 50 are relevant.

Export titles and abstracts

import json, os, requests

token = os.environ["ADS_TOKEN"]
query = 'abs:"radio halo" year:2018-2026 property:refereed'
r = requests.get(
    "https://api.adsabs.harvard.edu/v1/search/query",
    headers={"Authorization": f"Bearer {token}"},
    params={"q": query, "fl": "bibcode,title,author,year,abstract",
            "rows": 50, "sort": "citation_count desc"},
)
r.raise_for_status()
docs = r.json()["response"]["docs"]
json.dump(docs, open("papers.json", "w"), indent=1)
print(len(docs), "papers saved")

Ask for a map, constrained to the file

Upload papers.json and use this prompt:

You are helping me review the literature. Use ONLY the papers in the attached
file. Do not use outside knowledge and do not add references.

1. Group the papers into 4-7 themes. For each theme write 3-5 sentences.
2. Every factual statement must end with the bibcode(s) it comes from,
   in square brackets, exactly as written in the file.
3. List open questions or disagreements between papers, with bibcodes.
4. List what the abstracts do not let you judge (methods, caveats, numbers).

If something is not supported by the abstracts, say so instead of guessing.

How to check the result

1. Every bibcode exists in your file. Run this on the model’s answer saved as answer.md:

   import json, re
   known = {d["bibcode"] for d in json.load(open("papers.json"))}
   cited = set(re.findall(r"\b\d{4}[A-Za-z&.][\w&.]{13}[A-Z.]\b",
                          open("answer.md").read()))
   print("not in file:", sorted(cited - known) or "none")
   print("never cited:", len(known - cited), "papers")

2. Spot-check five claims against the abstracts they cite. If one is wrong, check them all.

3. Look at the papers that were never cited. The model often drops papers that do not fit its themes.

4. Read the key papers yourself before relying on any statement in your own writing.

Where it fails

• Results that are only in the body of a paper are invisible: the model sees abstracts only.
• Themes can be too tidy. Real disagreements get smoothed into consensus.
• The query decides everything. A narrow query gives a confident but partial map.

Variations

• Ask for a table instead of prose: theme, key result, bibcodes, open question.
• Repeat the same prompt with a second model and compare the themes.