Recall 2.6: citations that survive translation

Recall 2.6 is rolling out to all regions this week. The headline change is narrow and took most of the summer: citations are now anchored to the source document’s own text, in its own language, rather than to the translated rendering a user happens to be reading.

The problem

A lawyer in Zurich asks Recall, in German, for precedents on exclusivity carve-outs in distribution agreements. Recall finds a strong match in a Mandarin-language agreement from a 2022 matter and shows the relevant clause, rendered in German, with a citation to “clause 7.3, lines 4–9”.

In 2.5, that pinpoint was computed against the German rendering. Open the original Mandarin document and lines 4–9 of 7.3 were not where the sentence sat; translation changes sentence boundaries and length. The citation was correct in spirit and wrong in fact. A citation that is wrong in fact is not one a lawyer can put in a memo.

Across accounts using Recall in more than one language, roughly one query in six returned at least one result where the rendered language differed from the source language. That was enough to make the problem routine rather than an edge case.

What changed

Citations anchor to the source. Every citation now refers to a span in the source document as indexed: page, clause number where the document has them, and a character offset range in the original text. The rendered translation is a view on that span, not the thing cited.

Sentence-level alignment. When Recall renders a source span in another language, it keeps an alignment map between source sentences and rendered sentences. Hover a rendered sentence and the corresponding source sentence highlights. Click and the source opens at that position. The alignment confidence is shown; below a threshold, Recall shows the source alongside by default rather than letting a weak rendering stand alone.

Trace view shows both. The trace, introduced in 2.5 to show how an answer was assembled, now shows the source span and the rendered span side by side for every citation, with the alignment map visible. A reviewer can check that the translation did not drift without leaving the trace.

Export carries the source. Citations exported to a memo or to Draft include the source-language span as well as the rendering, with the source marked as authoritative. This mirrors how a prevailing-language clause works in a bilingual contract, which is not a coincidence.

Query and display in all 38 languages, independently. Ask in English, read in French, cite the Mandarin original.

Why this took a summer

Anchoring to source offsets sounds simple. It required re-indexing every document in every tenant to store stable offsets; building alignment at index time rather than at query time (query-time alignment cost around 300 ms per result and we were not willing to spend it); and handling the long tail of documents where “the source” is itself ambiguous: scanned PDFs with OCR, documents that are bilingual in two columns, documents where a translation was filed alongside an original and both were indexed.

Re-indexing runs per tenant, in region, in the background. For most accounts it completed before this release. Larger indices may show a “legacy citation” marker on older results for a few more days; the marker is explicit, and the old citation behaviour is retained for those results until re-indexing reaches them.

What to watch for

Alignment confidence under about 0.7 is shown with the source by default. This happens most on legal terms of art without clean equivalents and on long sentences with nested provisos. It is not a fault; it is Recall declining to let you rely on a rendering it is unsure of.

Right-to-left languages render correctly in the result pane and in the trace view. Side-by-side display with a left-to-right source is functional but not yet comfortable. We know.

Bilingual two-column originals. Where both columns are indexed, Recall now identifies which column was marked prevailing, if the document says, and anchors there. If the document does not say, it anchors to the column matching the matter’s recorded governing language, and the citation notes the assumption.

Known limitations

• Tables and schedules with heavy numeric content align poorly across languages. Citations into them are page-and-cell, not sentence-level.
• Mixed-language documents (an English body with German schedules) are handled per section, but a citation that spans a language boundary shows as two citations.
• OCR’d scans retain their OCR confidence, and a low-confidence OCR span produces a low-confidence citation. Recall does not improve on the scan.
• Alignment maps are not yet available in the DMS connector’s preview pane; open the result in Recall for the full view.

Ethical walls, enforced at query time since 2.3, apply to source and rendering alike; a user who cannot see the source document does not see its translation either.

A citation is a promise that the words are there. From 2.6, the promise is kept in the language the words were written in.