0.8.1 — Retrieval for large reference corpora

Released 2026-07-13.

What's new

Retrieval-augmented context (FR-23). Some reference material is too large to attach in full. Such a document can now be indexed as embedded chunks, and only the passages most relevant to the requirement under analysis are retrieved into the prompt. This complements the static Context Library from 0.8.0 rather than replacing it — a document is either included whole or retrieved on demand, chosen per document.
Chunk-level provenance. Where a retrieved passage influenced an analysis, you can see exactly which passage of which document it came from.
Retrieval budget view. A per-organisation view of how much retrieved material is entering analyses.

Attached context was not reaching analyses. A fault introduced in 0.8.0 meant no attached context — neither static nor retrieved — was actually applied when an analysis ran through the engine. Analyses completed normally and gave no indication that the context had been dropped. Fixed. If you attached context documents under 0.8.0 and want their benefit, re-run those analyses.

Vector store adapter with mandatory per-organisation filtering, so retrieval can never cross an organisation boundary.
Asynchronous embedding and indexing, so indexing a large document does not block the interface.
Retrieval relevance threshold recalibrated after live measurement.

Apply migrations/0.8.1/migration.sql before deploying.
Retrieval requires an embedding API key and a reachable vector store. Without them, static context continues to work normally and retrieval is skipped.