Capabilities, Resources & Archetypes

Every provider model card is annotated with three controlled vocabularies that describe what a provider is:

archetype — where it sits on the spectrum from pure retrieval to an autonomous co-scientist.
capabilities — the functional things it can do (verbs).
resources — the data sources / knowledge bases it wraps (nouns).

These vocabularies are defined once, authoritatively, in the LinkML schema at src/deep_research_client/schema/deep_research_client.yaml and generated into Pydantic enums under src/deep_research_client/datamodel/. To change them, edit the schema and run just gen-datamodel — never hand-edit the generated file.

Deep research is a subset of the co-scientist case

It is useful to see the providers as points on a single spectrum. A conventional "deep research" tool searches a corpus and writes a cited report. An autonomous co-scientist does that and more: it forms hypotheses, executes code against biomedical data, and can design experiments. In other words, a deep-research run is essentially a subset of what a co-scientist does — sometimes traded off against more sophisticated literature-scanning algorithms on the retrieval side.

The ProviderArchetype enum makes this ordering explicit (each archetype is broadly a superset of the previous one):

Rank	Archetype	What it does	Examples
1	`retriever`	Returns retrieved evidence (papers, passages, snippets) with no authored synthesis.	Asta
2	`synthesizer`	Searches one or more corpora and writes a cited narrative report — the classic "deep research" tool.	OpenAI Deep Research, Perplexity, Consensus, Edison/Falcon
3	`agentic_researcher`	A multi-step agent that plans, browses, and may run tools/code, still oriented around producing a written answer.	Claude Code, Cyberian, DeepER-Med
4	`co_scientist`	An autonomous scientific agent that forms and tests hypotheses, runs code against biomedical data, and can design experiments.	OpenScientist, Biomni

retriever  ⊂  synthesizer  ⊂  agentic_researcher  ⊂  co_scientist
                (deep research)                       (hypothesis-driven,
                                                        runs code)

Capabilities

ResearchCapability — the functional capabilities a provider exposes. The historical ModelCapability set (web_search, academic_search, scientific_literature, citation_tracking, real_time_data, code_interpretation, visual_analysis, multi_language) is retained as a compatibility enum. Its original UPPER_CASE members and standard enum lookup (for example, ModelCapability.WEB_SEARCH and ModelCapability["WEB_SEARCH"]) keep their historical behavior and string values. New code should use the schema-generated lowercase members such as ResearchCapability.web_search. The remaining values below were added to describe co-scientist behavior.

Capability	Meaning
`web_search`	Searches the open web for current information.
`academic_search`	Searches curated academic / scholarly corpora.
`scientific_literature`	Specialised search / analysis over scientific literature.
`citation_tracking`	Produces and tracks structured citations.
`real_time_data`	Can access recent / real-time information.
`code_interpretation`	Executes code (analysis, computation, plotting).
`visual_analysis`	Interprets or generates figures and charts.
`multi_language`	Operates across multiple natural languages.
`retrieval_only`	Returns retrieved evidence verbatim, no synthesis.
`evidence_synthesis`	Synthesises sources into an authored report.
`hypothesis_generation`	Proposes testable scientific hypotheses.
`experiment_design`	Designs experiments / analysis plans.
`data_analysis`	Analyses structured biomedical / experimental datasets.

retrieval_only and evidence_synthesis are the two terms tied to the archetype axis, and a test enforces the part this page states rather than leaving it to convention: a retriever carries retrieval_only and nothing else does, and a synthesizer — the archetype whose definition is evidence synthesis — carries evidence_synthesis.

Agentic researchers and co-scientists are not required to carry it. Every one shipped today does, and find_models_by_capability(evidence_synthesis) accordingly returns the deep-research tools and the co-scientists. But a co-scientist is defined by forming and testing hypotheses, not by writing a report, so one that returned a ranked hypothesis list instead would be annotated honestly rather than forced to claim the capability.

Resources

ResearchResource — the data sources and knowledge bases a provider wraps. A capability is a verb (what it does); a resource is a noun (what it can reach). pubmed is a resource; code_interpretation is a capability.

Resource	Examples
`general_web`	The open web.
`pubmed`	PubMed / MEDLINE.
`semantic_scholar`	Semantic Scholar corpus.
`arxiv`	arXiv preprints.
`preprint_servers`	bioRxiv / medRxiv.
`clinical_trials`	ClinicalTrials.gov and similar registries.
`biomedical_databases`	Curated disease / pathway / drug knowledge bases.
`genomic_databases`	Ensembl, ClinVar, etc.
`chemical_databases`	PubChem, ChEMBL, etc.
`protein_structure_databases`	UniProt, PDB, etc.

Querying by capability, resource, or archetype

The vocabularies are queryable across all registered providers:

from deep_research_client import (
    find_models_by_capability,
    find_models_by_resource,
    find_models_by_archetype,
    ResearchCapability,
    ResearchResource,
    ProviderArchetype,
)

# Which providers run code?
find_models_by_capability(ResearchCapability.code_interpretation)

# Which providers reach PubMed?
find_models_by_resource(ResearchResource.pubmed)

# Which providers are full co-scientists?
find_models_by_archetype(ProviderArchetype.co_scientist)   # -> openscientist, biomni

The same three axes are available from the CLI:

deep-research-client models --capability code_interpretation
deep-research-client models --resource pubmed
deep-research-client models --archetype co_scientist

Filters combine, so a conjunction asks the question it looks like it asks:

# Co-scientists that reach PubMed, not every co-scientist
deep-research-client models --archetype co_scientist --resource pubmed

Passing an unrecognised term lists the whole vocabulary, which is derived from the enum — so a value added to the schema appears in the help and the error message without anyone editing them. A filter that matches nothing says so rather than printing an empty result.

Each returns a {provider_name: [ModelCard, ...]} mapping.