Changelog: 2026-05-10
Launched our hosted Model Context Protocol (MCP) server with tiered access, alongside major improvements to search accuracy, voice workflows, and large-document performance.
- Introduced a hosted MCP server with tiered access, plus refinements to key tools used for search and document retrieval.
- Strengthened access controls and made API key management clearer for users across plans.
- Revamped Voice with a thread-based conversation experience and a new document-scoped mode for more focused assistance.
- Improved search reliability and result consistency across keyword and semantic search.
- Boosted stability and usability for red-flag screening and deep links within long documents.
- Delivered a major performance win that unblocks work with very large filings.
- MCP server & tool experience: Added a hosted MCP server and improved tool outputs so search results include richer snippets and more consistent identifiers, and document retrieval accepts additional reference formats.
- Access & API keys: Tightened product access rules, moved to a safer default access posture for API-key usage, and updated the UI so API key controls are easier to find for eligible users.
- Voice UX overhaul: Introduced a thread-style conversation panel with improved controls (including interruption and better session ending), plus better formatting of assistant messages. Added optional context summaries and a new filing-scoped voice mode designed to keep conversations anchored to a specific document.
- Search correctness: Fixed issues affecting array-based filters and improved parity between keyword and vector-based filtering so results are more predictable across search modes. Tuned hybrid search so semantic matching contributes appropriately when needed.
- Red flags & deep links: Improved red-flag visibility and filtering in filings lists, fixed highlight/deep-link behavior for truncated excerpts and long filings, and ensured failed screening runs don’t appear as successfully completed.
- Performance & reliability: Optimized evidence anchor normalization to scale linearly, improving responsiveness on large documents, and improved authenticated content fetching to reduce intermittent loading issues.