Product Updates
Insights that tell you what to do next: introducing Findings
Insights now goes beyond showing what happened. Findings point out where your assistant is falling short, recommend what to change, and keep watching until it is fixed. This release also strengthens security and lays the groundwork for Ragtime in more regions.
Insights has always helped teams understand their conversations. Highlights sit at the top of the page and bring forward what stands out, such as a shift in sentiment, a topic that is growing, or a change in how conversations end, and each one comes with a question you can explore further. Highlights tell you what is happening. Findings, a new tab in Insights, tell you what needs your attention. Every day, Ragtime looks across your conversations and your knowledge base for places where the assistant is struggling, explains each one with the numbers behind it, recommends a next step, and keeps track of it until it is handled. The result is a short, practical to-do list for improving your assistant, grounded in what visitors actually experienced.
- Three things Ragtime now watches for — Missing knowledge: visitors keep asking something your assistant cannot answer, so Ragtime groups those questions by meaning and shows them to you in anonymized form. Confusing source: conversations that rely on a particular document end unresolved or get escalated far more often than usual. Unused source: a document nobody has needed in months, which may be outdated or hard to find. Each finding explains what Ragtime noticed in plain figures, and links to the conversations behind it for the people allowed to see them.
- A recommendation on every finding — Each finding comes with a clear suggestion from Ragtime. That might be adding a source that answers the questions visitors are asking (using those questions as a checklist), reviewing a document for content that is outdated or unclear, or retiring one that no longer earns its place. Where it helps, the recommendation points to the existing document that would be quickest to extend.
- Your team decides, and Ragtime follows up — Mark a finding resolved and Ragtime watches new conversations for two weeks before confirming the fix, or reopens it if the problem returns. Defer it to a later date, or decide not to address it and note why. Deferred and declined findings come back on their own if the issue grows significantly, so nothing important quietly slips away. Every decision is recorded with who made it, and you can choose which team members are allowed to act on findings.
- Findings where you already work — Documents with an open finding are marked in the Knowledge Base, and a finding links straight to its document with that document highlighted. An Open Findings card on the project overview shows at a glance whether anything needs attention. Teams can also have new findings posted to a Slack channel, with a direct link to each one. A Check now button runs the analysis on demand, and test conversations raise their own clearly labelled findings, so you can see the feature at work before going live.
- Stronger security, and ready for more regions — We tightened how customer data is kept apart, so each organization's conversations, insights and knowledge are separated at the database level. We also extended encryption to more of the stored data behind the platform. In parallel, Ragtime is now built to run in more than one geographical region. All customers are still served from the EU, and when we open further regions, organizations will be able to keep their data where they need it.
- Over 50 bug fixes and UX improvements — Across voice, avatars, chat, the dashboard and the knowledge base, for a smoother and more dependable experience overall.
A clearer view across your assistants, with safer conversation learning
Teams can now understand performance across an entire organization, optionally learn from privacy-preserving transcript collections, and turn PowerPoint decks into assistant-ready knowledge.
This release helps teams move from isolated assistant activity to a fuller picture of what is working across their organization. Managers can compare assistants in one place, teams can choose to collect conversations in a form that removes personally identifying information, and familiar presentation files can become useful knowledge without having to rebuild them elsewhere.
- See the whole organization in one set of insights — Organization-wide Insights brings usage, conversation volume, topics, sentiment, and outcomes together across your assistants. Teams can spot patterns between projects, understand what is changing over time, and make decisions from the overall customer experience instead of stitching together separate reports.
- Learn from conversations without collecting personal details (enhanced feature) — Significant updates to PII-stripped transcript collections; now available as an opt-in capability. When enabled, conversations are sanitized before they are stored for analysis, giving teams a way to review what people experienced, investigate quality questions, and improve content while keeping personally identifying information out of the collection.
- Put PowerPoint knowledge to work — PowerPoint presentations can now be added directly to a knowledge base. Slide text, tables, titles, and speaker notes are preserved so onboarding decks, training materials, product briefings, and internal presentations can become sources an assistant can use without a separate conversion step.
Meet Mouthy, plus a more accessible and capable assistant platform
Ragtime's own Mouthy avatar brings an illustrative character to life with speech and emotion, alongside major improvements to project duplication, transcript insights, accessibility, and the player experience.
This release introduces Mouthy, Ragtime's own avatar provider. It starts with an illustrative character design that can talk and emote, giving teams a new way to create experiences where emotional realism matters more than realistic visual fidelity. Mouthy is now a standard part of the platform, with more character-driven possibilities to come. The release also makes the rest of the system easier to operate: project duplication is much more complete, Insights includes a PII-stripped transcript browser, people have more ways to navigate the player, and several player-focused bugs have been fixed.
- Meet Mouthy: an avatar with character — Mouthy is Ragtime's own avatar provider, starting with a pack of illustrative characters that can speak and express emotion. It is designed for experiences where the feeling of a believable interaction matters more than photorealistic visual detail — from learning and guidance to welcoming, coaching, and other human-facing conversations. Mouthy is now a standard platform feature, and this is only the beginning of what we plan to build with it.
- Emotional realism at scale — Mouthy gives teams a more expressive option when realistic visual fidelity is not the goal. Its character-led approach can be a strong fit for high-concurrency and cost-sensitive experiences, while still making an assistant feel present, responsive, and emotionally engaged.
- Project duplication that carries more of the project with it — Duplicating a project now offers more options what to duplicate. More of the configuration and connected project structure can be customized during this process.
- Browse conversations in Insights without exposing personal data — The Insights dashboard now includes a searchable transcript browser built around PII-stripped conversation data. Teams can inspect the conversations behind aggregate findings, understand what visitors actually experienced, and investigate quality or content questions while keeping personally identifying information out of the view.
- More ways to navigate the player — Accessibility improvements make the player easier to use for people who rely on alternative navigation methods.
- A smoother player experience — We also fixed several player-focused issues affecting controls, layout, and interaction behavior. The result is a more dependable experience across devices, whether someone is starting a conversation, changing a setting, or returning to an active interaction.
Better answers, faster workflows, and guided forms built with AI
Assistants can find stronger answers across varied content, while teams can turn a plain-language goal into an editable workflow or guided form in minutes.
This update shortens the distance between a visitor's question and a useful outcome. Assistants are now better at finding both conceptually relevant information and precise details such as product names, reference numbers, or policy terms. Once an answer needs to become an action, teams can describe the process they want and use AI to create a practical first draft of the workflow or guided form. The result is less time spent configuring from scratch, without giving up review and control before anything goes live.
- More reliable answers, however people phrase the question — Knowledge-base search now looks for both meaning and exact wording, then brings the strongest evidence together. This helps when a visitor uses different language from the source material, while still catching details that should match precisely. It is especially useful for support libraries, policy collections, technical documentation, and product catalogues where both context and exact terms matter.
- Relevant to the person and situation — Retrieval can focus on the topics, audiences, and locations that apply to the current interaction. Visitors get answers drawn from the right slice of the knowledge base, helping organizations serve different roles, cohorts, branches, or physical locations from one assistant without mixing their content.
- Describe an outcome and get a workflow draft — Teams can explain what should happen in plain language and receive an editable workflow as a starting point. AI can lay out the steps, decisions, approvals, waits, and connected actions needed for processes such as following up on a request, routing an escalation, or handing information to another system. You can inspect and refine the draft before publishing it.
- Create guided interactions without mapping every step by hand — The same approach is available for guided forms. Describe an intake, troubleshooting path, product selector, or onboarding journey and AI creates the questions, choices, forms, and branches. Existing actions and other guided forms can be included, making it easier to turn a conversation into a clear next step while keeping the experience consistent for every visitor.
- AI accelerates setup; your team stays in control — Generated workflows and guided forms remain drafts in the visual editor. Teams can change wording, reorder steps, adjust conditions, and confirm connected actions before enabling them. This makes AI useful for the repetitive setup work while preserving the review process that operational and customer-facing experiences need.
Topic cloud, a calmer chat layout, and sharper insights
Visitors can now see what an assistant actually knows about — scoped to their permissions and location — while the chat surface, dashboard, and insights engine all got easier to work with.
One of the hardest parts of any assistant is the blank page: people don't know what they're allowed to ask. This release tackles that from both ends. On the visitor side, a new topic cloud shows the subjects an assistant can genuinely talk about, filtered live by audience, cohort, and even physical location. On the operator side, library topics are easier to curate, the dashboard is easier to navigate, and the insights engine now tells you which conversational topics your data can actually support. We also cleaned up the chat layout so overlays stop fighting each other, and expanded Google-based voice options.
- Topic cloud: show people what they can ask — Assistants can now surface a live cloud of the topics they can talk about, drawn from the knowledge sources behind them. Every topic is validated server-side against the visitor's audience, cohort membership, and geographic scope, so a visitor standing in one location sees a different set than someone elsewhere — and nobody can replay a stale topic list to reach content they shouldn't. Managers control how many topics appear per library, whether the cloud starts expanded, and where the launcher sits.
- A calmer, smarter chat layout — The chat screen now keeps itself tidy on any device. Controls, captions and buttons stay out of each other's way instead of overlapping, so nothing important gets covered up. Typing on a phone no longer pushes the message box behind the keyboard, captions can sit wherever they read best, and everything opens and closes with a bit less jolt. Setting it up is simpler too — fewer knobs, sensible defaults, and the layout adapts on its own.
- Broader Google voice support — The Gemini-based voice pipeline now covers both speech recognition and speech synthesis, with 30+ prebuilt multilingual voices and no voice cloning required. Organizations can connect via API key or service account depending on their infrastructure, and the low-latency text-to-speech tier is available for real-time conversations. Spoken language is detected automatically, or pinned to a specific language when you need deterministic transcription.
- Dashboard: fewer clicks to the thing you need — Analytics and insights now live behind bookmarkable, shareable tabs so you can link a colleague straight to the view you're looking at. Navigation adapts to the active organization, management pages gained inline search for finding organizations by name or slug, and related settings have been regrouped so configuration lives where you'd expect it.
- Smarter library topic management — Libraries now separate topics extracted from your ingested sources from the topics your team curates by hand — the curated set extends the automatic one instead of overwriting it. Topics are deduplicated by meaning, so casing and near-duplicates collapse into a single entry, and the editor suggests topics based on what's actually in your sources. Per-library limits keep the resulting topic cloud focused rather than overwhelming.
- Insights that point at real conversations — The insights engine now communicates far more clearly which conversational topics your data can support: top topics are reported with conversation counts and share of total, you can compare a single device or location against the rest of your network, and plain-language questions return a grounded narrative alongside the metrics behind it. Saved questions can be re-run over new date ranges for recurring reporting, and data-quality warnings flag conversations analysed with older models. All of it stays aggregate — never individual transcripts.
Direct RAG APIs, clearer scopes, and smarter topic detection
You can now query the project knowledge base directly via the KB API and benefit from more consistent topic detection across ingested sources.
This update focuses on making the RAG layer easier to use in real integrations. Instead of going through an assistant for every query, you can now access the knowledge base directly, list available sources and topics, and read document content with explicit permission boundaries. We also tightened the topic-detection pipeline so newly ingested sources are labelled more consistently and feel more useful when you later filter or search by topic.
- Direct knowledge-base access — The public API now exposes direct project-level access to the knowledge base via search, source listing, and topic listing. Retrieval audiences are assigned when an API key is created, so each key only sees unrestricted libraries and audience-restricted libraries it was explicitly granted. This is the easiest way to query the RAG database without building a custom assistant flow just to run retrieval.
- Topic detection improvements — We improved the topic extraction (now tied to each individual sources rather than the library itself) and normalization flow so complex or repeated concepts are tagged more consistently. This helps filtering, browsing, and search behave more predictably when you work with large libraries or mixed content sets.
- Bug fixes and reliability — This release also closes a few rough edges: scope labels were cleaned up, missing documentation gaps were filled, and some retrieval and metadata issues were corrected so the updates are easier to understand and easier to trust in production.
Encrypted database volumes and a rebuilt billing engine
The Ragtime database now runs on encrypted volumes, billing has been rewritten, and BYOK support is being prepared for partners and direct customers.
August focused on the layers underneath the product. Moving to volume-level encryption removes a constraint that had been blocking work on retrieval, and the billing rewrite clears the way for bring-your-own-key deployments.
- Encrypted database volumes — The Ragtime database now runs on encrypted volumes, so all data is encrypted at rest. This replaces the previous column-level encryption, which limited the retrieval features we could build on top of stored content. No migration is required on your side.
- Rebuilt billing engine — Billing has been rewritten from the ground up for clearer usage accounting and more flexible plan structures. Existing plans and invoices are unchanged.
- Groundwork for BYOK — The new billing foundation prepares support for bring-your-own-key setups, where channel partners and direct customers supply their own AI provider credentials. More details to follow as this rolls out.
- Ongoing security updates — Continued dependency updates and hardening across the platform.
July platform release: security, API validation, and a rebuilt chat UI
A large refactoring round: 30 pull requests and more than 150 fixes covering security hardening, API validation, a unified chat renderer, spreadsheet ingestion, and chat theming.
July was a consolidation month. Over roughly six days of focused refactoring we merged 30 pull requests and resolved more than 150 issues on staging, all of which are now live. The work concentrated on closing abuse vectors, making the public API predictable and documented, and removing duplication that had built up in the chat client.
- Security hardening — Closed four abuse gaps covering webhook signature verification, analytics authentication, tool name collisions, and unbounded input sizes. Draft assistants can no longer be reached before publication, and several localization bugs were fixed.
- API validation and expansion — Thirteen v1 handlers moved to schema-based request validation, which surfaced and fixed five bugs. We published an OpenAPI 3.1 contract for the public API and exposed eight new analytics endpoints.
- Chat UI refactor — Three duplicated chat renderers were unified into one. The theme system was rebuilt on OkLab colour math with eight presets, microphone pickers were consolidated into a single component, and rendering issues on iOS Safari were fixed.
- Chat theming — Eight presets with device-aware light and dark schemes, laying the groundwork for broader visual customisation of embedded assistants.
- Spreadsheet ingestion — Ingestion now accepts .xlsx and .xlsm files up to 20,000 rows per sheet, with clearer upload validation errors.
- Voice and performance — Inline performance cues for text-to-speech are now enabled, and the speech pipeline recovers more gracefully from transient failures.
- Client stability — Six widget bugs fixed across idle timeouts, sentence splitting, and health monitoring, backed by 138 new tests.
- Session continuation during outages — Kiosk deployments now continue a conversation after a connectivity interruption. The assistant acknowledges the gap and resumes without losing the session.
- Test coverage — 336 new tests added across eleven chat modules.