How HerbaGraph turns labs into evidence-graded recommendations
Seven deterministic stages connect biomarkers to biological systems, pathways,
a knowledge graph, and literature-backed intervention classes — with dual
knowledge paths you can A/B test.
7pipeline stages
9recommendation trees
28biological pathways
2knowledge paths
The 7-stage analysis pipeline
Click a stage to see what happens, what data moves, and where recommendations come from.
Two knowledge paths — same labs, different substrate
Legacy catalogs
Production default. Curated intervention catalogs, Tier A evidence claims with PMIDs, and live literature retrieval (PubMed / ClinicalTrials / Europe PMC).
Routable claims on 28 priority pathways
Nine test-type recommendation trees
Deterministic catalog fallback when LLM is unstable
Recommended for clinical review
Canonical graph engine
Experimental A/B path. Pathway activations traverse MODULATES / TARGETS / ACTIVATES / INHIBITS edges on canonical_entities, ranked by evidence type and confidence.
PMID-backed edges from evidence claims
Predicted edges from catalog mechanisms (~700 long-tail interventions)
CONTAINS composition for food sources
Graph-backed · hybrid fill
Knowledge graph anatomy
Recommendations are not a black box. They are walks on a typed graph with provenance.
Recommendation trees & pathway density
Nine recommendation treesExample pathway claim density (illustrative)
Charts are schematic for education. Live counts evolve as catalogs and graph edges expand.
From graph hits to the clinical decision map
Safety layer — contraindications, drug interactions, regulated flags strip or demote items.
Evidence confidence — study type, quality score, catalog connectivity, and uncertainty.
Four lanes — direct biomarker, lifestyle, supportive adjunct, regulated context.
Explainability — supporting literature, pathway drivers, and limitations stay attached to each recommendation.
1Direct biomarker
2Lifestyle
3Supportive adjunct
4Regulated context
What HerbaGraph does not do
Not a diagnosis
Outputs are evidence synthesis and biological reasoning for research and education. They do not replace licensed medical judgment.
Not invented literature
Tier A claims carry PMIDs. Predicted graph edges are labeled PREDICTED and scored lower than clinical or meta-analytic edges.
See it on a real-shaped report
Walk a sample clinical report, or upload your own labs in the workspace.