A clearer, more capable research workspace

This roundup brings together the work since our last published update, through September 20, 2026. These improvements arrived across multiple updates, rather than in a single release.

AI models: more choice, clearer access

• Updated model picker. Auto uses DeepSeek V4 Flash. Pro choices in the current app include GPT-5.6 Terra, Claude Opus 5, and Gemini 3.8 Flash.
• Model-aware allowances. Usage reflects the selected model rather than treating every answer as the same cost. The plan and usage displays make allowances easier to follow.
• Consistent plan access. Model access is checked when a request is submitted, not only in the menu. If a saved model is no longer included in a user's plan, the picker returns to Auto.

Model availability and allowances depend on the current plan and service configuration.

Thinking steps: follow the work behind an answer

• A clearer activity timeline. Expand the steps to follow document retrieval, web searches, and the evidence being used, with progress and completion states.
• Thinking summaries where supported. Models that provide readable reasoning summaries can show them alongside the activity timeline. This is a progress view, not a promise to expose every internal model operation.
• More useful step details. Retrieval steps can show the angles being explored and source pages; web-search steps can show result links. Repeated activity is grouped to reduce clutter.
• Optional answer verification. The chat supports verification progress and outcomes, including an insufficient-evidence state instead of presenting an unsupported answer as verified. Verification is optional, not applied to every answer by default, and is not a guarantee of factual accuracy.
• Web search controls. Web search can be enabled from chat when outside sources are useful, with its activity shown in the same timeline.

Retrieval & citations: find the right evidence

• Keyword and semantic retrieval. The updated index combines keyword matching with semantic search, then reranks candidate passages for relevance. Older indexes retain a compatible fallback path.
• Search plans for broader questions. Broad questions can be split into several focused searches; narrow factual questions can take a shorter path. Small documents can use their full text within the available context budget.
• Better context around passages. Retrieval carries document and section context so a matching fragment is less likely to lose its surrounding meaning.
• An evidence-quality check. Weak results can trigger a refined search. When the available evidence is insufficient, the system can surface that limitation rather than force an answer.
• More useful source previews. Citation previews show the passage, not just its section title. Hover prefetch reduces repeated waiting, and citation handling has been refined for answers written in a different language from the document.
• Citation validation. Server-side checks filter invalid citation identifiers. A citation still points you back to the evidence; it does not replace checking the original passage.

PDF region tools: work with what is on the page

• On-demand page recognition. The reader detects figures, tables, and formulas on visible pages, rather than running visual extraction across the whole document during upload.
• Explain a region. Use a detected region or a manual area selection to request an explanation in chat. Region explanations do not require the document's full text index to be ready.
• Extract and copy. Request a table as Markdown or a formula as LaTeX, and copy supported region content from its toolbar. AI extraction is performed when requested and can require correction.
• Reuse page detections. Cached region coordinates help when returning to a page. Embedded PDF images also have a local detection path; automatic detection is not guaranteed to find every scanned or complex region.

The latest reader replaces the earlier full-document picture/table ingestion pipeline. Existing saved citation data and crops remain readable. Automatic region recognition depends on the layout service being available; manual area selection remains a separate path.

Document processing: clearer progress and recovery

• Direct uploads. Signed upload URLs send files directly to storage rather than routing the entire file through the application server.
• Background text indexing. Text extraction, chunk storage, and indexing continue in the background, with progress and processing counts in the reader. Whole-document chat waits for the text index to be ready.
• Retry and upgrade flows. Failed processing can be retried, and older document indexes have a version-aware reprocessing flow. Previously processed documents are not silently upgraded just by publishing new code.
• Clearer errors and limits. Upload failures, conversion failures, and plan limits have more specific feedback. CSV upload support has been removed.
• Less repeated work. Retrieval caching, startup warm-up, and safer cleanup when deleting documents improve the supporting workflow without changing the documents you choose to keep.

Reading & notes: a more connected workspace

• Searchable document picker. Find and open documents from the reading workspace, with improved management of open reading tabs.
• Refined reader controls. Sidebar, toolbar, table-of-contents, loading, and citation interactions have been updated around the reading experience.
• Rich-text notes. The note editor moved to TipTap, with block editing and slash commands for structured writing. Note and export workflows have received follow-up refinements.
• Read-aloud improvements. Sentence highlighting follows playback, and preparing the next page's audio helps the transition between pages.

Accounts & billing: clearer entry and plan management

• Refreshed sign-in and sign-up. Responsive authentication pages use garden backgrounds with light and dark variants.
• Clearer subscription state. Plan cards, usage indicators, checkout, and billing-cycle handling have been refined to better reflect the active subscription.
• Access stays in sync. Subscription updates and entitlement checks align model access and usage limits with the account's current plan.

Previously retired: Public Library

The old Public Library and its public book-reader routes were removed on June 17, 2025. They are not part of the current product. This does not refer to the personal documents you upload into your workspace.

The original announcement is retained as history and is now marked as describing a retired feature. See the Public Library retirement note.