1.0.38

v1.0.38

🧹 Prune v0.11.0

---

Files removed from a knowledge base could keep answering questions: when their chunks were left behind in the knowledge base's vector collection (a deleted folder does this in current Open WebUI releases), retrieval kept finding them and citations still opened them, and Prune did not catch it because the file records themselves were still referenced by chats. Reported in #82.

Added

• Stale knowledge base embeddings are reconciled on every sweep: each kept knowledge base's collection is compared against the files it actually lists, and the chunks of files it no longer lists are deleted. Works through the unified vector client, so ChromaDB, PGVector, Milvus and Qdrant (including multitenancy) are all covered. ChromaDB and PGVector read only chunk metadata; Milvus and Qdrant load each knowledge base's chunks during the check.
• delete_stale_knowledge_base_chunks valve, on by default, with a matching checkbox on the manual page and a "Files with stale knowledge base embeddings" row in the preview.

Changed

• orphan_file_grace_hours now also protects knowledge base embeddings of files updated inside the window, because a file is embedded before it is linked to its knowledge base. Keep the window above 0 while files get added.

👁️ Vision Bridge v1.1.0

---

Images the model creates in Open Terminal are terminal files, not chat attachments, so analyze_image could not find them and the generate-then-check workflow stalled on a missing-image error.

Added

• analyze_terminal_image(path, query): reads an image straight from the Open Terminal attached to the chat, through the terminal's own read_file, and hands it to the vision model. The text-only model can now generate an image, look at the result, and iterate. Contributed by @d3cker in #84.

Changed

• The filter hides analyze_terminal_image from the model when the chat has no terminal, so the model never sees a tool it cannot use. Needs Open WebUI 0.11.2 or newer (the tool still works on 0.11.0 and 0.11.1, it is just always visible there). Native function calling only.
• Both tools share one vision call and one status helper.

📊 Inline Visualizer v2.2.3

---

Fixed

• Bold and other markdown in the text after a visualization rendered as literal asterisks whenever the model glued its follow-up prose to @@@VIZ-END without an empty line, because Open WebUI's markdown parser then treats the whole tail as one raw HTML block. The skill and the tool's instructions now require one empty line after the end marker.
• A marker word the model mentioned while thinking (for example "the block must close with @@@VIZ-END") blanked the first sentence of the answer and deleted the words from the thought block. Reasoning is now left alone by both the hide pass and the finalize cleanup, unless the visualization itself lives inside reasoning, which still renders as before.
• A page that cannot run injected code (a content security policy without unsafe-eval) left the tool call stuck on "Executing visualize" forever, because the UI language probe never returned. The probe now times out after five seconds and falls back to English.

Changed

• The build marker in the iframe (data-iv-build) now tracks the plugin version.
• README and skill corrected where they contradicted the code: the chime is muted through the chime valve only, CDN allowlisting covers scripts only (inline the CSS a library needs), color ramp classes reach direct children only, the skill is named visualize, and the obsolete pre-0.9.5 reload note is gone.

📚 Repository

Changed

• The main README lists the maintained plugins in one table and moves Inline Visualizer v1 and Keep reasoning_content into a legacy table marked as no longer maintained.