Skip to content
Protean

Evidence layer·Retrieval

Evidence retrieval

A local semantic layer over the runtime's scientific context. Retrieves relevant records before candidate explanations, generated papers, failure-aware analysis, and comparative review — without leaving the machine.

What it indexes

The index currently reads from:

  • processed evidence records
  • literature mention records
  • failure records
  • generated assessment context when available
  • patent or raw records when present in the local data layer

The index contract

The index stores text, metadata, and embeddings in a local SQLite file. Sequence-like queries are rejected at the boundary — peptide and protein sequence comparison belongs to the ESM-2 sequence layer, not the text retrieval layer.

data/cache/evidence_index.sqlite
-> documents
-> text embeddings (BAAI/bge-m3)
-> source metadata
-> retrieval scores

How retrieval is used

Retrieved evidence improves context selection. It helps generated papers and candidate explanations speak from the best available local scientific material instead of dumping raw runtime fields. The reranker (BAAI/bge-reranker-v2-m3) sharpens the retrieved set before narrative generation.

Retrieval does not establish assay validation, efficacy, safety, or patentability, and it does not replace deterministic candidate validators. It is a context layer that helps reviewers and local reasoning models focus attention.