build-study-context

name	build-study-context
description	Use this skill to assemble an evidence bundle (context texts) from a parsed submission or ETL study context, including DOI abstracts and PDF content.

Build Study Context Skill

Use this skill to assemble an evidence bundle from a parsed submission or ETL study context. Called by both metadata-suggestion-pipeline and env-triad before making suggestions. Requires the doi-ingestion and pdf-ingestion skills.

---

Inputs

Provide one of:

• parsed_submission — output of the submission-parser skill (has description, notes, study_name, dois, protocol_descs, protocol_names)
• study_context — list of strings or dicts (ETL path — use when no submission_object is available)

---

Steps

1 — Assemble base text

If parsed_submission is provided, build the base context string:

Submission description: <description>
Notes: <notes>
Study name: <study_name>
Protocol descriptions: <protocol_descs joined by '; '>
Protocol names: <protocol_names joined by '; '>

If only study_context is provided, use those strings/dicts directly as the base text. Skip to output.

2 — Fetch DOI abstracts

For each entry in parsed_submission.dois, use the doi-ingestion skill:

result = get_doi_description_or_abstract(
    doi=entry["value"],
    sources=[entry["provider"]] if entry["provider"] else None,
    skip_classification=True,
)

• Add result.context to the evidence texts (label it with the DOI value)
• Accumulate all result.publication_urls into a list

3 — Fetch PDF content

Utilize download_pdf_to_tempfile (from pdf-ingestion skill).

4 — (Optional) Fetch supplementary materials

For publication DOIs, optionally enrich the evidence with high-value supplements (sample-metadata tables, supplementary methods) using the supplement-retrieval skill. This is performance-first — make a single attempt per DOI and skip on failure:

from nmdc_metadata_suggestor_ai_tool.publication_ingestion.supplements import (
    retrieve_supplements,
)

# Pass the DOI's abstract text (from step 2) so linked-dataset DOIs mentioned in
# the prose can also be resolved. The skill routes by DOI type and merges hosted
# supplements with any linked Dryad/Zenodo/Figshare datasets.
supp = retrieve_supplements(doi=entry["value"], text=result.context)
for f in supp.files:
    if f.text:
        context_texts.append(
            f"Supplement {f.filename} [{f.source}] (for {entry['value']}):\n{f.text}"
        )
    # non-text files are at f.saved_path for downstream readers

Only tabular and document supplements are kept by default. Do not retry when supplements are unavailable.

Non-text supplements (xlsx/pdf/docx) are written to temp files, so delete them once you have read what you need — otherwise every run leaves up to max_total_bytes of them behind:

from nmdc_metadata_suggestor_ai_tool.publication_ingestion.download_pdf import (
    remove_temp_files,
)

remove_temp_files([f.saved_path for f in supp.files if f.saved_path])

Pass save_dir=... to retrieve_supplements instead if you want to keep the files; those are yours to manage and are never removed for you.

---

Output

An evidence bundle — pass to whatever pipeline skill calls this:

Part	Description
context_texts	List of strings: base text + one entry per DOI abstract + fetched PDF content. Use as evidence when reasoning.

No cleanup required in the agentic path — web_fetch reads content inline without creating temp files. The programmatic supplement path in step 4 is the exception: it writes temp files, which that step's remove_temp_files call cleans up.