Purpose: classify structured papers, map study/experiment scopes, build packets,
extract canonical evidence, and derive traceable general-evidence/RAG handoff
artifacts.
Commands:
uv run victus-processing evidence-extraction run
uv run victus-processing evidence-extraction run --skip-existing --limit 20
uv run victus-processing evidence-derivation build
uv run victus-processing evidence-derivation build --input data/runtime/04-evidence/{paper_id}/canonical_evidence.json --llm-conclusions --language es
Inputs:
Outputs:
data/runtime/04-evidence/{paper_id}/;paper_classifications PostgreSQL rows;experiment_maps PostgreSQL rows for primary-research papers;canonical_evidence PostgreSQL rows after extraction;general_evidence_artifacts.json, containing exposure/outcome registries,rag_export.json, containing only general_evidence and evidence_supportpaper_processing_state refresh shows the next missing stage.evidence-derivation build is the post-canonical command. It reads
canonical_evidence.json plus sibling experiment_map.json, then writes
general_evidence_artifacts.json and rag_export.json. Use
--llm-conclusions only when conclusion text should be generated with the
controlled GeneralEvidence prompt; counts, ranks, consensus, and recommendation
use remain deterministic.
This repository does not implement vector indexing, retrieval, Qdrant, or RAG
serving. It only prepares the JSON handoff payloads.
Design notes:
study_design because design describes the study/scopeCanonicalEvidence remains paper-level extracted evidence and does not storeGeneralEvidence aggregates by study/paper support units so one paper withMigration notes:
intervention_or_exposure maps to raw_exposure.outcomes maps to raw_outcomes.direction: no_change maps to effect_direction: no_effect;canonical_evidence.effect_direction.Validation:
uv run victus-processing evidence-extraction --help
uv run victus-processing evidence-extraction run --help