Assign EDAM ontology terms to text using a live SemanticSQL database and an LLM

Usage

edamize(
  content_for_edam,
  provider = "anthropic",
  model = "claude-sonnet-4-5",
  nterms = 20L,
  prescrub = TRUE,
  prompt = read_prompt("edamize.txt"),
  retrieve_k = NULL,
  sim_threshold = 0.3,
  embed_model = "text-embedding-3-small",
  ...
)

Arguments

content_for_edam: character(1) text describing a bioinformatics resource
provider: character(1) LLM provider; see llm_env_var. Defaults to "anthropic".
model: character(1) model identifier for the selected provider. Defaults to "claude-sonnet-4-5".
nterms: integer(1) approximate number of EDAM terms to select. Defaults to 20.
prescrub: logical(1) if TRUE apply cleantxt before processing. Defaults to TRUE.
prompt: character(1) instruction text sent to the LLM before the EDAM vocabulary tables and content. The string must contain one %d placeholder that will be replaced by nterms. Defaults to the contents of inst/prompts/edamize.txt; supply your own string to customise curation behaviour without editing package files.
retrieve_k: integer(1) or NULL. When not NULL, use embedding-based retrieval (via retrieve_edam_candidates) to pre-filter the EDAM vocabulary to the top retrieve_k candidates per type before LLM selection. Requires the API key for the embedding provider recorded in the artifact (see get_edam_embeddings and llm_env_var). Set to NULL to pass the full vocabulary directly to the LLM. Defaults to 75L.
sim_threshold: numeric(1) minimum cosine similarity for a candidate term to be passed to the LLM. Terms below this threshold are dropped before the LLM selection step, reducing irrelevant tags. Only used when retrieve_k is not NULL. Defaults to 0.3.
embed_model: character(1) embedding model used for retrieval; must match the model used to build the artifact. Defaults to "text-embedding-3-small".
...: passed to llm_chat

Value

a data.frame with columns uri (full EDAM URI) and tm (term label), restricted to confirmed vocabulary entries and deduplicated. Compatible with mkdf, toline, and edam_graph.

Note

This function replaces the former Python/curbioc.py implementation. It connects to the current EDAM release via ontoProc2::semsql_connect() and selects terms using chat_structured() via ellmer, so no JSON schema validation loop is needed and hallucinated term labels are eliminated by post-filtering against the actual vocabulary.

Examples

# Input validation fires without any API key
tryCatch(edamize(list(a=1)), error = function(e) conditionMessage(e))
#> [1] "content_for_edam must be a single character string; did you mean to pass e.g. tst$focused?"

if (interactive() && nchar(Sys.getenv("ANTHROPIC_API_KEY")) > 0) {
  content <- readRDS(system.file("rds/tximetaFocused.rds", package="biocEDAM"))
  lk <- edamize(content$focused, retrieve_k = NULL)
  print(lk)
}