Configuration

Configure schema loading, retrieval and optional credentials in metasalmonpy.

Settings are read at call time. Importing the package does not set environment variables or supply credentials. Environment settings do not request a model call. Model assessment of semantic suggestions requires llm_assess=True (and seed_semantics=True in infer_dictionary()). chat_decomposition() has a separate opt-in path through chat_provider= or chat_request_fn=. The S16 plan removes these in-package provider paths in 0.7.0; this page describes their current controls.

SDP schema

Setting Meaning and precedence
METASALMONPY_SDP_SCHEMA_SOURCE The loader’s auto default tries the pinned remote bundle then the vendored copy; remote requires remote loading; vendored stays offline. Review and setter defaults are described below.
METASALMONPY_SDP_SCHEMA_BASE_URL Select a different schema bundle base URL; unset uses the package’s current pinned release.
set_sdp_schema_source() Process override ahead of the source environment variable. Pass None to restore environment/default resolution.
set_sdp_schema_base_url() Process override ahead of the base-URL environment variable. Pass None to restore environment/default resolution. 
from metasalmonpy.sdp_schema import set_sdp_schema_source

set_sdp_schema_source("vendored")  # Read the bundled schema offline.
set_sdp_schema_source(None)        # Restore environment/default resolution.

Changing either process override clears the schema cache. Under the shipped auto source and pinned base URL, package creation tries the remote bundle before falling back to the vendored copy. review_metadata() and the four set_sdp_*() setters instead read the vendored copy offline; this is still true if auto was selected explicitly with the pinned URL. A different source or base URL makes review and setters use the selected loader schema, which may contact the network. The default schema pin is maintained with the vendored bundle when a later spec release is implemented.

Retrieval

Environment variable Meaning
METASALMONPY_CACHE Enable the process-local term-result cache with 1, true or yes; unset/empty disables it.
METASALMONPY_DEBUG_FETCH Enable retrieval diagnostics with 1, true or yes; unset/empty disables them.
SALMONPY_CACHE, SALMONPY_DEBUG_FETCH Deprecated pre-rename fallbacks, consulted only when the corresponding current spelling is unset/empty. Each emits one DeprecationWarning per process (which Python may filter by default); both remain supported here.
BIOPORTAL_APIKEY BioPortal credential. The first BioPortal search without a key emits one RuntimeWarning per process and returns an empty BioPortal result.

BioPortal warning bookkeeping is private module state. Session identifier generation uses UUIDs without consuming Python’s statistical random stream. The R companion’s configuration topic is ?metasalmon_configuration; its metasalmon.* options and METASALMON_* retrieval knobs are R controls.

GitHub and DataONE

A nonempty token= argument takes precedence for GitHub readers. Otherwise, token discovery checks GITHUB_PAT, then GH_TOKEN, then the Git credential helper. See GitHub data access for the existing setup and read helpers.

DataONE credentials are process-local, separate for each publication environment:

from metasalmonpy.knb_publication import (
    set_dataone_token,
    set_dataone_test_token,
)

Supply a short-lived token through the appropriate setter when making an explicitly requested deposit. Production and test credentials do not substitute for each other, and tokens are never written to a package manifest. Passing None clears that environment’s process-local credential.

Optional in-package model provider path

Model assessment of semantic suggestions selects a provider with llm_provider (openai by default); explicit llm_model, llm_base_url and llm_api_key arguments take precedence. chat_decomposition() uses the corresponding chat_provider, chat_model, chat_base_url and chat_api_key arguments. Both paths use the provider-specific fallbacks below when a call is explicitly requested. Setting an environment variable alone does not opt in.

Provider Model fallback Endpoint fallback Credential fallback
openai OPENAI_MODEL OPENAI_BASE_URL OPENAI_API_KEY
openrouter OPENROUTER_MODEL OPENROUTER_BASE_URL OPENROUTER_API_KEY
openai_compatible OPENAI_COMPATIBLE_MODEL OPENAI_COMPATIBLE_BASE_URL METASALMON_LLM_API_KEY
chapi CHAPI_MODEL CHAPI_BASE_URL CHAPI_API_KEY

An unset model or endpoint falls back to that provider’s package preset where one exists. openai_compatible has no preset: a network call requires a model, endpoint and credential. An injected llm_request_fn or chat_request_fn replaces the network request and does not require an endpoint or credential. Semantic review accepts an explicit llm_reasoning_effort; chat_decomposition() does not have a reasoning-effort argument.

Follow the semantic review guide for the explicit review flow. The configuration above supplies inputs to existing calls; it never applies or approves an ontology term.

Back to top