Skip to main content

What's new in Hindsight 0.10.3

Tag filters on entities and the knowledge base, per-candidate reranker token caps, faster bank imports, and a round of retain, consolidation, reflect and Oracle reliability fixes

By Nicolò Boschi
· 3 min read

Hindsight 0.10.3 adds tag filters to the entity views and the knowledge base, caps how much of each candidate reaches the reranker on every provider, and makes importing a bank faster and resumable. The rest of the release is reliability work across retain, consolidation, reflect and Oracle. Everyone on 0.10.x should upgrade — see the changelog for the full list.

Hindsight 0.10.3 highlights: tag filters on entities and the knowledge base, a per-candidate reranker token cap, and faster bank imports

Tag filters on entities and the knowledge base​

The entity list, entity detail and entity graph now take a tag filter, and so do the knowledge-base tree and its search. On a bank that holds more than one audience — per-user, per-team, per-project tags — you can look at one slice's entities or one slice's pages without reading the rest.

Per-candidate reranker token caps​

The limit on how much of each candidate is sent to the reranker now applies across every provider, not just some. A few long documents in a result set can no longer blow up a rerank call on the provider you happen to run.

Faster bank imports​

Importing a bank writes its documents in batches instead of one at a time, so large transfers finish faster. If an import-bank crashes part way, it now cleans up after itself, so the retry starts from a known state instead of landing on top of a half-finished import.

Worker flags --max-retries and --log-level are removed in this release: they never took effect. Drop them from your command lines before upgrading. One migration ships too — an Oracle Text index for knowledge-page search — and migrations run on startup.

This release also carries a long list of correctness fixes: appends that verify the document they append to, consolidation writes applied as one batch, reflect answers citing real memory ids, and the remaining Oracle gaps. See the changelog for all of them.