v2026.3.8

Releases of the 2026.3 line now ship from this repository, openviglet/turing, and the Docker image is ghcr.io/openviglet/turing. The repository that carried the community edition took over the openviglet/turing name with the full history; the previous one is archived as openviglet/turing-legacy. This release also carries the changes prepared for 2026.3.7, which was never published.

SEARCH

• An exact title phrase outranks synonym-only matches on Lucene, Solr and Elasticsearch.
• A search with no results retries with the query's main words.
• Portuguese plurals in -ões, -ães and -ãos are stemmed to their singular on Lucene and Elasticsearch.
• Spell-check no longer changes numbers, symbols or the case you typed.
• Similar documents no longer come back empty for documents that only have a title.
• Text search is stemmed by the site's locale on Lucene and on new Elasticsearch cores, so a plural finds its singular.
• Facet filters and custom sorts now apply on Elasticsearch.
• Hybrid sites recall semantic hits the lexical page missed, keep the request's filters on them, and can tune the ranking per site.
• An instant search matches the last word typed as a prefix.
• Applied facets stay in place on drill-down sites and when their field resolves to the OR item type.
• Synonyms on Lucene sites survive a restart, and wildcard queries or half-typed words are no longer mined as synonyms.
• Similar documents honour the search's filters and locale.

SEMANTIC NAVIGATION COPILOT AND CHAT

• A question can pin the document it is about as result 1, in SN chat, the copilot and both SDKs.
• The copilot ranks each kit stage against the whole plan and writes the kit in the request's language.
• /answer honours _setlocale and answers from the site's indexed locale.
• The copilot stages the catalog products a described plan needs, explains each choice, respects availability, and cites the site's guides beside the kit; both SDKs expose the kit call.
• A follow-up question in SN chat and the copilot is retrieved as a standalone question, and Live Preview shows the query it searched.
• Deleting a document also removes its chunks from the RAG collection (Milvus included), and orphaned chunks can be purged.
• The console flags a site that has no vectors, and Default-agent sites can reindex.
• A site's GenAI and facet behaviour can be turned on from the SN manifest, and a site's AI capabilities can be probed without calling a model.

SECURITY AND MULTI-TENANT SITES

• An API key can be bound to a partition of a shared SN site; search, import, autocomplete, spell-check, similar documents and deletes stay inside it.
• Requests authenticated by an API key no longer need a CSRF token, and their audit entries name the key's user.

ANALYTICS, COST AND SDK

• Failed searches, content gaps and click-through rates per term are available as REST metrics, per partition.
• A site's search metrics can be cleared from the API and the console, and a structured search can opt out of them with populateMetrics: false.
• The JS and React SDKs ask for consent before sending search analytics or result clicks, send GA4 search events, report search abandonment, and cancel a search a newer one supersedes.
• SN chat, NL filters and summaries are recorded in the cost ledger against their site; a call with no known price is shown as unpriced, not free.

FIXES

• Importing a bundle keeps its agent ids and imports agents before sites; a seeded site's imported agent becomes the default AI agent.
• Frontend dependencies with high-severity advisories (browserslist, adm-zip, source-map-js, zod, follow-redirects) are patched.
• Streamed answers keep the space before each token.
• Wrong-type query parameters return 400 naming the parameter.
• Relative dates in natural-language filters resolve against today.
• Solr: currency fields exist in every configset, cores are created through the v1 Core Admin API, and exclusive range filters work.

DOCUMENTATION

• The end-user documentation moved into this repository and is published at openviglet.github.io/turing/docs.