What changed
Section titled “What changed”New 6
- Knowledge Knowledge bases: files, web pages and notes searchable by meaning and keyword, stored and searched in the browser. Knowledge bases
- Knowledge Save to a knowledge base from a page, a selection or a chat. Save from anywhere
- Knowledge Portable .awkb files to export a base and import it on another device. Export & import
- Aria & agents A knowledge picker in chat with citations you can open, and a Knowledge tab for agents. Knowledge in chats & agents
- Aria & agents A work timeline shows each model and tool step as it happens, with a reply receipt. Following Aria’s work
- Aria & agents Turn traces (Developer mode) show where the time and tokens of each reply went. Turn traces
Improved 7
- Knowledge Hybrid keyword and vector search, with optional on-device reranking. RAG concepts
- Nodes Workflow knowledge nodes gained hybrid search, MMR, metadata filters and a richer base picker. Local Knowledge
- Aria & agents Project files now use the same knowledge engine. Projects
- Workflows Node parameters Aria generates are validated against each node’s schema before a workflow is applied. Building workflows in chat
- Local AI Gemini Nano can call tools, and local models stream replies token by token even with tools bound. Chrome AI
- Aria & agents Workflow and MCP tools are locked by what they do, not by their name. Action approval
- Aria & agents Stopping a reply is clean, and you are told about step limits, Auto fallbacks and skills that are off.
Fixed 6
- Aria & agents Agent approvals are queued, so approving one never declines another. Action approval
- Aria & agents Aria keeps her skills inside team threads, and team-scoped memories are shared between members. Teams
- Aria & agents A team thread cut off by closing the app is shown and can be retried.
- Aria & agents Cloud tokens are metered per model call, and the limit is enforced during a turn.
- Aria & agents A clarifying question asked during a build survives a page reload.
- App Local database connections share one lock, which fixes hangs when several features open it at once.
In detail
Section titled “In detail”Knowledge bases
Section titled “Knowledge bases”A knowledge base holds sources you add (files, web pages, notes) and makes them searchable by meaning and by keyword. Everything is stored and searched in the browser.
- A Knowledge bases home where you create a base with its sources, and an Add sources dialog for adding more later.
- A detail page per base: test a search, open a source to see its content, see what a delete will affect before you confirm, and re-embed after changing the embedding model.
- Save to a knowledge base from anywhere — from a page, a selection, or a chat.
- Hybrid search combines keyword (BM25) and vector search, with optional on-device reranking for better ordering. Search settings can be tuned per base and per tool.
- Portable
.awkbfiles let you export a base and import it on another device. Importing a workflow lets you map its knowledge-base references to your own bases. - PDF sources keep their page numbers, so citations point to the right page.
- One vocabulary and one privacy control for knowledge across the app.
- There is no plan limit on knowledge bases: they live on your device.
Encrypted account backups of knowledge bases were built but are switched off
for now. Use .awkb export to keep a copy.
Relevant docs:
Knowledge in chats, agents and workflows
Section titled “Knowledge in chats, agents and workflows”- A knowledge picker in chat lets you choose which bases Aria can search. Answers show citations you can open, and the search steps explain themselves.
- Agents get a Knowledge tab and a
search_knowledgetool, and agent knowledge works on every surface where the agent runs. - Project files now use the same knowledge engine.
- Workflow knowledge nodes gained a device-only guard, fenced context, source details, MMR and metadata filters, a search tool, hybrid search, and a richer base picker. Workflows remember which base they came from.
Relevant docs:
The AI harness
Section titled “The AI harness”The engine behind every assistant reply was reworked end to end, covering how it runs, the context it sends, model providers, observability, and the chat UI:
- A work timeline shows each model and tool step as it happens, and a reply receipt sums up what a reply used.
- Stopping a reply is clean, and you are told when a step limit is reached, when Auto falls back to another model, or when a skill is off.
- Turn traces (Developer mode) show where the time and tokens of each reply went, with JSON export. Traces stay on your device.
- The chat UI is built from Svelte AI Elements components (shimmer, chain of thought, context usage, confirmations, plans).
- Generated node parameters are validated against each node’s schema before a workflow is applied.
- Gemini Nano (Chrome AI) can call tools through the prompt, and local models stream replies token by token even when tools are bound.
- A clarifying question asked during a build survives a page reload.
Relevant docs:
Agents and teams fixes
Section titled “Agents and teams fixes”- Workflow and MCP tools are locked by what they do, not by their name.
- Agent approvals are queued, so approving one never declines another.
- Aria keeps her skills inside team threads.
- Cloud tokens are metered per model call, and the limit is enforced during a turn rather than after it.
- A team thread cut off by closing the app is shown and can be retried.
delegate_toalways has its schema, and team-scoped memories are shared between team members.- Approval activity titles are translated again.
Relevant docs:
Fixes and localization
Section titled “Fixes and localization”- All local database connections share one lock, which fixes hangs when several features open the database at once.
- Knowledge base strings are translated into French, German and Chinese, with translation-safe plurals.