Search Tips

Search in the source language, in period spelling

The documents are early-modern Swedish, Latin and German — not Finnish. Orthography is pre-reform and unstandardised.

And in Latin: ecclesia 602, littera 356, dominus 244.

Swedish stemming is applied, so konungen finds konung, konungs and konungavalit. Accents are folded, so Abo finds Åbo. Stop words are kept, unlike a normal Swedish index: they are Latin content words here, and removing them cost de 1,735 documents, den 846 and om 1,133. Quoted phrases work — "de ecclesia" is narrower than de ecclesia.

The edition's apparatus is stripped before indexing

Diplomatarium Fennicum is a scholarly edition, so its transcripts carry editorial apparatus inline: footnote markers fused to the word they annotate (Hundæbæth⁶, Karulj²) and editorial insertions in square brackets, sometimes mid-word (Fi[n]llandh, eccl[esi]a). A tokeniser has no reason to treat either as punctuation, so Hundæbæth⁶ indexed as a single token and the plain word matched nothing — across 2,266 occurrences in 621 records, plus 4,801 bracketed forms in another 1,969.

The apparatus is now removed from the search text only. transcript still shows the edition verbatim, because the apparatus is part of what a researcher is reading — so you may well see a ⁶ in a result whose word you found without one.

Several words mean all of them

A multi-word keyword requires every word; "quoted words" require that exact phrase. AND, OR and NOT are not operators and are searched for literally, so bref OR littera also matches the 51 charters containing the word or. Widen by dropping a word or passing match_all=false, never by writing OR.

Spelling is the main reason a search looks empty

Orthography was never standardised, so the same word appears in many forms and a single spelling finds one scribe's usage, not the word. Measured on this corpus:

Two different mechanisms address this and the engine makes them mutually exclusive:

• Stemming (always on) handles inflection: konungen finds all 279 charters that stem to konung.
• fuzzy=1 handles orthography: it takes bref from 257 charters to 2,212, and 96.9% of those still contain a real variant. fuzzy=2 was tried and rejected — recall barely moves while precision falls to 67%.

A fuzzy term skips the analysis pipeline, so it is matched raw against stemmed index terms: konungen with fuzzy=1 collapses from 279 hits to 6. So exact-plus-stemming is the default, and when a search looks thin the right second attempt is fuzzy=1 on a base form — konung, not konungen.

Latin and German are not stemmed — use a prefix

The stemmer is Swedish, so ecclesia, ecclesie and ecclesiam are three unrelated tokens, and fuzzy does not bridge them: it is whole-word edit distance, and lepros is four edits from leprosorum. On the full corpus lepros with fuzzy=2 returned 19 charters, none of them the one about the leper house at Reval, and most of them noise.

A trailing * is a prefix, expanded against the corpus's own vocabulary: lepros* matches leprosi and leprosorum and finds all three leprosy charters. It needs at least 3 characters and expands to the 300 most frequent forms (kon* begins 347, and a capped search says so). | lists alternatives within a term — bref|breff, or reval*|reual*|revel* — which is what to write instead of OR.

A place is two different questions

issuingplace="Tallinn" finds the 197 charters issued at Tallinn. It does not find the ones about Tallinn: a charter issued elsewhere is issued elsewhere, and 2,314 charters — a third of the corpus — record no place at all and are excluded by any place filter. DF 173, the Reval chapter's 1279 appeal for its leper house, is one of those; no keyword can reach it through the filter.

The text says the period name, so search that, as a prefix, with the topic as another word:

df_search(keyword="lepros* reval*|reual*|revel*|reuel*|reffl*")

Measured on the corpus, Tallinn appears as revele (102 charters), reuel (93), revel (77), revall (69), reval (60), reuall (59), reffle (38), reualie (22), ræffla (10) and a dozen rarer forms. Gdansk is dantz*|dantsk*, Tartu darpt*|darbt*|dorpt*|tarbat*. A filtered search ends with a note that says all this, so a narrow result does not read as a complete one.

The DF number is searchable

DF 1451 — or just 1451 — finds that charter. This is also what makes the four charters with no transcript, place, index term or language reachable at all.

A zero result means the term is absent

kirkko — the Finnish for "church" — appears in zero of the 6,876 charters. That is not a broken index; the Finnish index vocabulary uses Kirkolliset and Piispat, and the documents themselves say ecclesia or kyrkia. Do not smoke-test with a modern Finnish word.

Filter values are Finnish, and unaccented

language takes an exact label: ruotsi (Swedish, 3,045), latina (1,638), saksa (German, 1,232), venaja (Russian, 89). Note venaja, not venäjä — the accented form matches nothing. 27 charters carry more than one language, as compounds like latina, ruotsi.

country takes a Finnish country label as a substring: Suomi (2,249), Ruotsi (1,207), Italia (467). 2,321 charters record no country.

issuingplace takes a place name as a substring, over a vocabulary of 499 values. 2,314 charters record no place. These are substring counts, which is what the filter does: Rom (364) also matches the five charters issued at Magliano Romano.

The naming is mixed, and this is the trap. Finnish and Swedish places keep their historical Swedish form — Åbo (815, not Turku), Viborg (311, not Viipuri), Nådendal (134), Raseborg (120), Tavastehus (43). But places outside that realm are recorded under their modern name:

Reval, Danzig and Dorpat return nothing at all. So the "use the period form" rule that governs the text does not govern this filter: the text is medieval, the place vocabulary is a modern cataloguer's.

Untranscribed does not mean absent

2,464 of 6,876 charters — 36% — are catalogued but never transcribed. They are still returned, marked (catalogued but not transcribed), and remain findable by place, index term and language, because the full-text index is built over those fields as well as the transcript. An index over the transcript alone would silently lose a third of the corpus.

Dates are intervals, and mostly open-ended

6,843 of 6,876 charters have no end year. year_min / year_max therefore test interval overlap against a derived closed interval, not a comparison against the raw end year — a filter written the obvious way would match 33 charters out of the whole corpus.

22 charters carry no year at all. A year-bounded search excludes them, as SQL excludes NULLs.

The text is machine-recognised

These are OCR and handwritten-text-recognition outputs over a millennium of handwriting, not proofread transcriptions. Expect misrecognised characters and broken word boundaries, and check any quotation against the archive before presenting it as exact.