> For the complete documentation index, see [llms.txt](https://oswaldai.gitbook.io/docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://oswaldai.gitbook.io/docs/release-notes-3.5.8.md).

# Release notes - 3.5.8

Release date: 06/08/2026

## Knowledge Base Index V3

We've introduced a new version of our search index (V3), now the default for all newly created bots and indexes. It brings smarter search, better multilingual handling, and more control over how your content is prepared and ranked.

* *How it works:* V3 is opt-in for existing chatbots. You can upgrade an existing chatbot using the "Switch index version" button, found in the LLM Hub tab under Settings. Because upgrading rebuilds the index from scratch — a destructive operation on a live index — we strongly recommend testing the upgrade on a separate test bot first. A confirmation dialog warns you before the change is applied, and older index versions remain fully supported, so there's no obligation to switch.

## **Hybrid search**

V3 chatbots now blend three ways of finding results instead of relying on one: exact keyword matching (BM25), meaning-based matching that catches results even when the wording differs (vector similarity), and a final reranking step (Azure's semantic reranker). In practice, keyword-heavy queries — product codes, article numbers, proper names — return noticeably better results.

* *How it works:* Hybrid search turns on automatically once a chatbot is upgraded to V3. Existing V2 chatbots are unaffected, since the upgrade is per-chatbot and must be triggered manually.

## Language-aware search

V3 indexes handle Dutch, French, English, and German with language-specific processing — reducing words to their root form and splitting compound words (stemming, lemmatization, and compound-word splitting) — so searches match more reliably in each language. You also get new control over how the widget's language influences results.

* *How it works:* Each V3 Knowledge Base has a Language mode setting with three options:
  * Disabled — widget language is ignored and all content ranks equally (the default, matching V2 behaviour).
  * Prefer — content in the widget's language ranks higher, while other languages still appear.
  * Strict — search is limited to the widget's language only.

You can also set an optional **Primary language** per Knowledge Base, used as a fallback when a data source or item has no language of its own.

## **Configurable embedding model**

V3 lets you choose which model converts your content into the form used for meaning-based search (the embedding model).

* *How it works:* Two options are available — Local (bge-m3), the default, and Azure OpenAI, which uses your own Azure OpenAI embedding deployment (separate from your LLM or Azure Search) and is handled by Azure Search at the moment a query runs. Switching embedding model rebuilds the entire index, so this is also a destructive operation on a live index; a confirmation dialog warns you before you proceed.

## Improvements

* **Smarter indexing**: We now compare a content fingerprint (hash) of each pushed data object against the existing one, so unchanged content no longer generates unnecessary index traffic.
* **Confirmation dialogs on destructive actions**: The index-switch and embedding-model-swap buttons now show a warning dialog before applying these irreversible changes.
* **Automatic data source cleanup**: Data sources belonging to a deleted chatbot or Knowledge Base are now removed automatically.
* **More thorough index cleanup**: Cleanup now also removes leftover individual pieces of content (chunks) within an existing Knowledge Base and clears duplicates (same URL and title) on existing webhooks.
* **Infrastructure update:** changes have been made to the Oswald backend services, resulting in new ingress and egress IP addresses. If you whitelist Oswald IPs in your environment, you'll need to update your allowlists with the new IP addresses before this release goes live. Please refer to our previous communication for the updated IP list.

## Bug fixes

* **Large data pushes**: Fixed an issue where pushing a large amount of data to the index at once could cause problems.
* **Double indexing**: Fixed an issue where pushing data to the webhook just before the scheduled sync job (cronjob) ran could index the same data twice.
* **Unnecessary re-indexing**: Fixed an issue where syncing a webhook data source (via "Sync now" or a scheduled sync) could force a full re-index of every item even when nothing had changed; regular syncs now skip unchanged content, and a full re-index only happens on an explicit manual sync.
* **Item counter accuracy**: Fixed an issue where the item counter on the Knowledge Base data sources overview could show the wrong count for webhook sources.
* **Export/import index version**: Fixed an issue where exporting and re-importing a chatbot didn't carry over its search index version correctly.
* **Password reset**: Fixed an issue where resetting your Oswald account wasn't sending out a reset e-mail correctly.