wikidata-guide

Scope

Use Wikidata for structured relationships, identifiers, multilingual entity names, and links to Wikimedia articles. No API key is required. Use an article or source-document service when the task needs full text; a Wikidata statement or sitelink does not supply that text. Missing claims and links do not establish that a real-world fact or article is absent.

Identifiers

• Q-ID: an item, obtained from wikidata_search_entities, wikidata_resolve_external_id, a returned statement value, or a SPARQL entity URI. Keep the Q-ID through joins; labels can be shared by unrelated items.
• P-ID: a property, discovered with property search in wikidata_search_entities or read from returned statements. Inspect an unfamiliar property's entity with wikidata_get_entity to check its meaning and datatype before using it.
• External identifier: the DOI, PMID, ORCID, or other identifier supplied by a source record. Pair it with that identifier scheme's P-ID for wikidata_resolve_external_id; a name or title is not an external identifier.
• Statement ID: returned with a claim by wikidata_get_statements; retain it with the evidence it identifies. It is not the entity's Q-ID or the property's P-ID.
• Site code: a key such as enwiki in returned sitelinks, identifying a Wikimedia project and language. Obtain article URLs with wikidata_get_sitelinks, rather than constructing URLs from entity labels.

Gates

1. Settle identity before joining facts. Compare candidate descriptions and relevant statements with the known person, place, or work. Search rank and matching labels alone are insufficient. If candidates remain indistinguishable, report the ambiguity or ask for the missing discriminator.
2. Separate resolver outcomes. Inspect multipleMatches before interpreting a null match as a miss; fetch candidates to disambiguate them. An ordinary null means no indexed match, not proof of absence: the Query Service can lag the live wiki. Inspect known candidate entities directly when recent edits matter.
3. Retrieve claims before citing them. A kind: "outline" response contains an inventory, not the statements or entity fields. Follow retrieval_notice; do not turn section names or counts into evidence.
4. Read statement context before making a factual conclusion. Inspect rank, qualifiers, and references, keeping each reference's parts together. Preferred rank can express recency or consensus; it does not alone establish that a value is current. Check time qualifiers for time-sensitive claims, and state when Wikidata supplies no supporting reference.
5. Preserve the value's meaning. somevalue means an unknown value, novalue an explicit absence; neither means a missing property. Preserve time precision, quantity units and bounds, coordinate precision, and language tags when present. Do not convert a year-precision time into a known day, or an unresolved Q-ID into a guessed label.
6. Check coverage before saying “all” or “none.” Inspect search pagination, SPARQL counts and notices, and label lookup misses. SPARQL truncated signals an inferred upstream row cap; response-budget cuts are reported separately. A false flag, a LIMIT-bounded query, or a page of results does not prove completeness. Describe an empty query as no matching indexed rows under that query's conditions.