Terpedia: an evidence-first terpene knowledge platform

Terpedia, LLC is a Colorado company founded by a terpene biochemist and a knowledge-systems engineer to build the reference infrastructure the terpene economy lacks. The platform has three layers: a knowledge base that ingests and reconciles sources with provenance intact; a set of access surfaces (encyclopedia, API, agents, product catalog) that put that knowledge where people and systems work; and a research program that tests the platform’s own claims and publishes the results.

Design principles¶

Four principles govern every part of the system.

1. Identity is structural. Compounds are keyed on standard InChI and InChIKey; names, synonyms, and source identifiers are attributes of the identity, never the key.

2. Provenance travels with the record. Every promoted row carries, where applicable, its source release, source file URI, manifest URI, content hash, ingestion run identifier, load timestamp, and license. Every statement in the semantic graph carries its dataset, retrieval time, evidence type, and (where available) PubMed identifier or DOI.

3. Evidence types are explicit. A source-curated association, a reviewed ontology term, a measured concentration, an occurrence report, and a computational prediction are different kinds of statement and are stored and displayed as such.

4. Claims stay smaller than the data. The platform does not assert effects. It reports what sources say, at the evidence grade they say it, and it says so in every API response.

Architecture¶

The shared knowledge base runs on Google Cloud (project terpedia-489015) as five deliberately separate layers, each chosen for what it is good at.

A synchronization layer (“tonsync”) manages immutable snapshots, manifests, validation, and promotion between layers, so that any served record can be traced to the exact source bytes it was derived from. Scheduled jobs refresh sources on their natural cadence; relation sets from CannabisDatabase.ca, for example, refresh weekly.

Sources¶

The source catalog currently covers more than thirty public and licensed databases across chemistry, biology, composition, regulation, and literature. Appendix A: Source inventory gives the full inventory with ingestion status. The principal groups are:

• Chemical identity and classification: PubChem, ChEBI, the FDA UNII registry (159,376 records, 986,598 names), COCONUT Sorokina et al., 2021, TeroKit Zeng et al., 2020, SuperNatural II, KNApSAcK.

• Occurrence: LOTUS Rutz et al., 2022, Dr. Duke’s phytochemical and ethnobotanical databases, the Native American Ethnobotany Database (NAEB), TCMID.

• Biochemistry: Rhea, BRENDA, UniProt (adapter planned), and terpene–protein interaction sets (SAIR).

• Composition: EssoilDB essential-oil compositions; CannabisDatabase.ca (12,039 source records, 54,251 relations); normalized cannabis laboratory profiles (43,018 source rows, 409,655 measurements).

• Assays: PubChem BioAssay.

• Regulatory and pharmacopoeial: FDA and SCOGS records, HPUS, and monograph collections.

• Literature and IP: PubMed (via an automated terpene-focused collection pipeline), patent indexes and documents.

• Ontologies: MeSH, MONDO, Cellosaurus.

Scale, stated carefully¶

Terpedia publishes its counts as versioned observations with an explicit identity key and classification rule Terpedia, LLC, 2026. The following figures were retrieved from the production counting endpoint in September 2026.

These figures are not additive; records overlap across datasets, and the general COCONUT, UNII, and SuperNatural tables are deliberately not presented as terpene totals because they have not been through a classification pass. That restraint is the point. A number with a scope can be reproduced, audited, and updated; a number without one cannot.

Access surfaces¶

0.1kb.terpedia.com: the open biochemical graph¶

The encyclopedia site is a browsable, searchable view of the resolved graph. Users can search any name and browse by category: terpenes, molecules, proteins, enzymes, reactions, pathways, products, ingredients, diseases, and organisms. Each entity page shows identity, structure, cross-references, classification, related entities, and every statement with its source and license. The site is server-rendered from the same API described next, so what a person sees and what a program receives are the same data.

0.2The Terpedia Knowledge API¶

The Knowledge API is the programmatic surface. Its open encyclopedia resolver, GET /v1/cyc/{keyword}, accepts a compound, cultivar, or protein name and returns the resolved entry, related entities, typed statements with sources, citations, and a plain-text evidence context. It requires no key and permits cross-origin requests, which is why the live examples in this paper can run in a browser. Keyed partner endpoints add:

• POST /v1/tabular/search: search a named source table or the unified source union (for example, all records mentioning “limonene” across sources).

• POST /v1/tabular/count: scoped, classified counts with identity keys, suitable for audit.

• POST /v1/semantic/search: semantic search with source citations and graph context.

• GET /v1/entities/{id} and /v1/entities/{id}/neighbors: entity and graph-neighborhood retrieval.

Partners with analytical needs can also be granted direct BigQuery access to the promoted tables and SPARQL access to the semantic store. A documented strain-profile endpoint distinguishes normalized measurements from wide source rows for structured consumers such as Colab notebooks.

0.3chat.terpedia.com: domain agents with a knowledge boundary¶

The chat surface runs a hub-and-spoke multi-agent system over the knowledge base. As of this writing the live roster includes twenty-two specialist agents and eleven Tersona personas. Specialists include a Terpene Scientist (chemistry, sensory profiles, formulation implications, analytical interpretation), a COA Analyst (certificate interpretation, lab-result review, specifications), a Compliance Agent backed by state-specific regulatory datasets, Formulation, Sourcing, Patent, Spectra, Cultivation, Medical Literature, Functional Products, Product Intelligence, and a Publication Editor for template-governed documents. The backend also exposes OpenAI-compatible /v1 endpoints so that existing tools can point at Terpedia as a drop-in model provider with domain retrieval behind it.

What distinguishes these agents from a general chatbot with a good prompt is where the boundary lives. Retrieval returns statements with their evidence type and source, and the knowledge API’s context block begins with an explicit instruction not to extend claims beyond the cited evidence. The knowledge layer sets the boundary; the language model works inside it.

0.4Tersona: the terpenes speak for themselves¶

Tersonae are entity personas, one per terpene (Limonene, Myrcene, α-Pinene, Linalool, β-Caryophyllene, Humulene, Terpinolene, Ocimene, Bisabolol, Geraniol), hosted by TerpeneQueen, a persona voiced by Terpedia’s founding scientist. Each Tersona is grounded in its molecule’s record and evidence, and the system supports speech in and out, so a consumer, a retail associate, or a student can ask limonene about itself and get an answer that stays inside what the graph can support. Tersonae also subscribe to literature and news feeds for their own name, which is the basis of the planned feed-first terpedia.com experience.

0.5Terproduct: products, ingredients, certificates, compounds¶

Terproduct is the system of record for product-level chemistry: products → ingredients → certificates of analysis → compound results. A product’s label lines are parsed into ingredients; ingredients are linked to organisms and to compounds; certificates are stored as documents with their parsed compound results; and every compound is joined to the knowledge base by InChIKey so that identity, bioactivity, disease associations, literature, and LOTUS occurrence records attach automatically. An installable web app supports barcode scanning to look up or ingest products in the field.

Two data-model decisions in Terproduct show the evidence rules at work. Disease associations carry a kind: a reported association (the compound was detected or studied in that condition) is distinguished from an occupational exposure hazard, and the molecule page must render the kind alongside the condition, because a bare disease name next to a consumer product reads as a health claim regardless of what the record says. And LOTUS occurrences are described in the interface as reports of detection, not as concentrations or sourcing recommendations.

0.6Terports and the encyclopedia¶

Terports are Terpedia’s short-form research reports: structured, cited summaries on a compound, a question, or a product class, generated from the knowledge base and reviewed before publication. The terpedia.com encyclopedia exposes compound lookups (/cyc/{compound}) and text-processing endpoints that automatically link scientific keywords in content to their entries, so that a partner’s own site can carry Terpedia definitions without leaving the page.

Partnership example: MONDAYS¶

MONDAYS, a hemp-derived terpene chew brand, publishes a measured terpene profile on each product page: every compound at or above 0.05% of total volatiles, the aggregate trace fraction, and the milligrams each compound contributes to a 20 mg terpene serving. Terpedia builds the co-branded catalog at mondays.terpedia.com from that data and maintains the molecule reference pages that each compound links to. A product marked as measured carries laboratory-quantified compounds from its certificate; the catalog shows composition only. The protein-assay, disease-association, and literature records for each molecule live on Terproduct, where they are labeled as laboratory-dose findings, and each molecule page links out to them rather than restating them next to the product.

The arrangement is a template for any brand that wants to talk about its terpenes truthfully: publish the measurement, link to the science, keep the two visibly separate.

Research portfolio¶

Terpedia runs a source-auditable research program whose purpose is to test the platform’s own assumptions and to publish reusable methods. Fourteen projects are active as of September 2026; the index at terpedia.github.io/research tracks status and target journals. Several are directly relevant to industry:

• Census Terpedia, LLC, 2026: how many terpenes there are, and why the answer depends on the counting rule.

• Claims Terpedia, LLC, 2026: promotional terpene effects recast as hypotheses and joined to receptor evidence without confusing mechanism with efficacy; includes a prioritized register of falsifiable tests.

• Cannabis terpene chemistry: the Terpotype classifier McShan & Trapp, 2026 and the provenance audit that demoted it McShan & Trapp, 2026, together with a record-level variability analysis and a causal-network model of terpene profile variation.

• Classifier and TerpNet: auditable molecular classification using canonical identity and fingerprints, and a general terpene pathway graph framework for genome, transcript, and docking evidence.

• Absinthe Terpedia, LLC, 2026: a critical evidence map testing whether absinthe should be described as psychedelic, mapping thujone, fenchone, and trans-anethole against 5-HT2A and GABA_A pharmacology; the companion public site is at nowhere.terpedia.com.

• Artemisia Terpedia, LLC, 2026: a frozen scoping review of terpene diversity, biosynthetic evolution, and bounded antiparasitic evidence across the genus that gave the world artemisinin.

• Terpene PubMed: a nightly search matrix that makes literature density visible across terpene and topic combinations.

• NAuTILUS: a GPU-accelerated design for large-scale terpene–protein interaction ranking and docking, positioned explicitly as a hypothesis-ranking tool rather than a binding oracle.

• Functional Flavors and ADHD attention: research surfaces for flavor chemistry and for the attention literature in which terpenes are sometimes invoked.

The portfolio’s stated evidence rule is worth quoting in full, because it is also the platform’s:

A database record is not a validated natural product. A receptor association is not efficacy. A docking score is not affinity. A graph path is not proof of in-vivo production.

The next section shows the platform doing what this section describes, against the live service.