C4-EvidenceRAG

Condition-aware, evidence-traceable retrieval and abstention baseline for maize C4 photosynthesis.
中文说明 · Data card · Methods · Results

What this repository is

C4-EvidenceRAG is a reproducible research prototype that turns a curated maize C4 evidence table into an inspectable retrieval pipeline. It returns claim-level evidence with a claim_id → document_id → DOI → PDF locator trail and applies explicit abstention guardrails when a question exceeds the corpus boundary.

Public release scope: 45 document records, 192 curated claims, 192 evidence-rooted graph edges, and 100 benchmark questions (40 dev / 60 test / 4 abstention items).

Why it is different

• Condition-aware: temperature, water, CO2, light, nitrogen, tissue/cell type, developmental stage, and treatment remain explicit fields.
• Evidence-bound: each result includes a claim, DOI, page/section locator, evidence grade, and context.
• Graph-safe: condition-dependent results are represented as contrasts instead of universal contradictions.
• Abstention-first: unsupported intervention optimization, future-yield extrapolation, cross-species transfer, and raw-image measurement requests are declined.
• Reproducible: Python standard-library baseline; no PDFs, model weights, APIs, or paper text are needed.

Quick start

git clone https://github.com/HZ-HZAU/c4-evidence-rag.git
cd c4-evidence-rag
python -m unittest discover -s tests -v
python scripts/query.py --question "How does drought change maize photosynthesis under elevated CO2?"
python scripts/evaluate.py

The query command prints JSON with an answerability status and citation-bound evidence summaries. Results are written to artifacts/.

Layout

src/c4_evidence_rag/   retrieval and citation-bound answer compiler
data/                   metadata-only claims, document registry, graph, benchmark
docs/                   methods, governance, results, release guidance
scripts/                query, evaluation, and release validation
tests/                  regression tests
artifacts/              reproducible public-baseline outputs

Deliberately excluded

No source PDFs, figures, tables, full text, verbatim evidence quotations, local paths, reviewer identifiers, model weights, or hidden API calls are included. confirmed is not independent expert consensus, a breeding recommendation, or causal proof beyond the cited study context.

Status and citation

This is a research prototype, not a production agronomy or breeding-decision system. High-priority benchmark review and prior expert adjudication remain pending; see docs/RESULTS.md. Cite this release and each original paper DOI used. See CITATION.cff.

Code: MIT. Project-authored metadata annotations: DATA_LICENSE.md. Third-party publications retain their own licenses.