Agent Platform Release Notes

August, 2026

New Features

• Custom hub agent skills: Super-admins can now author custom agent skills for a hub, end-to-end from the dashboard. A new Skills page under a hub provides an in-app SKILL.md editor — the same file format an author would write locally. Skills are stored per hub, take effect on the hub's next agent turn (no deploy required), and follow the agent skills spec. Built-in skills continue to be available automatically alongside custom ones.

Improvements

• Answer-confidence judge is now advisory: The runtime answer-confidence judge no longer auto-creates QA evaluation tickets on borderline turns. It now records its verdict; opening an evaluation remains a human action. This materially cuts noise on QA review workflows without losing the confidence signal itself.

• Truthful catalog answers on large stores: Product-discovery skills on catalog-v2 stores (100k+ items) now read the entire catalog rather than a silently truncated first page, and follow a grounded resolve → explore → filter tool flow so answers stay faithful to the store's taxonomy. Filter values are matched case-insensitively, so a store that lists items as Copper is no longer reported as "unavailable" for a copper query.

• Grounded citations on catalog lookups: Products the agent surfaces via family or type lookups are now recorded as evidence and become citable in the answer. Closes a class of "the retrieved context was empty" QA flags on hubs whose narrowing runs primarily through family/type lookups.

• Product part-number search matches by base number: Searches like TPS54331 now find variants such as TPS54331DDA; searches like SLES098 find SLES098A. Users no longer need to type the full suffixed part number to locate a document.

• Adobe Commerce product cards show images and "View product" links: Products ingested from an Adobe Commerce (Magento) source now carry image assets and a View product CTA into copilot answers, so recommended-product cards display images and link back to the store page.

Bug Fixes

• Safety filter blocks multi-turn jailbreaks: Content blocked by the conversation safety filter on one turn is no longer replayed to the agent as history on the next turn, closing a two-turn gap where a blocked request could reach the model indirectly.

July, 2026

New Features

• Institutional Knowledge: Mark any source as institutional knowledge so its content — including its images — informs the AI Agent's answers without ever being cited or linked. Ideal for informal or internal material you want the agent to learn from, but not surface to users as an official reference. Applied per source via a simple tag on the source's configuration.

• Catalog V2 — typed-attribute product answers (opt-in per hub): A foundation for typed product answers driven by a customer's taxonomy — attributes with units, vendors, and product relations. Additive to today's chat experience and opt-in per hub. Pairs with the taxonomy browser in the dashboard.

• PDF-creation skill for agents: Agents can now generate PDF artifacts directly in a workflow.

• Direct product lookups on catalog hubs: On catalog-backed hubs, the agent now correctly reaches into the product catalog for direct product questions — specifications, capability checks, attribute searches, "does X support Y?" — not only guided "which product fits me" journeys. Paired with a hard requirement that the agent must run at least one catalog lookup before answering a product question, materially reducing "no evidence" answers on product Qs. Auto-enabled on hubs that declare a product catalog.

• Competitor catalogs are respected in product recommendations: When a catalog is marked as a competitor's, the product discovery agent will never recommend or cite its products; it uses the catalog only to match a specifically named competitor part when a user asks about one.

Improvements

• More reliable visuals in agent responses: A deterministic repair pass now heals malformed markup in agent-generated interactive visualizations at generation time, and also on the read path — so previously broken widgets self-repair when a past conversation is reopened. Interactive-widget alerts now fire when a user actually interacts with a widget, not when the tool runs.

• Pro agent completion emails are now opt-in per hub: Proactive completion emails are off by default and only send when a hub explicitly opts in, reducing email volume during onboarding.

• Web sources describe images by default: URL list and Sitemap sources now describe embedded images at ingestion time by default, so figures, diagrams, and schematics inform the agent's answers without an opt-in. Matches the same default already in effect for PDF sources.

• Safety filter enabled by default for new copilots: Newly-created copilots now ship with the conversation safety filter turned on out of the box; existing copilots are unchanged.

Bug Fixes

• is_answered analytics counts clarifying MCQs and declined-out-of-scope correctly: Success-rate metrics now reflect what the agent actually did — clarifying multiple-choice turns and correctly-declined out-of-scope questions are counted as answered.

June, 2026

New Features

• Catalog typing drives accessory recommendations: Catalogs can now be tagged with a type — Primary, Competitor, Starter or Accessory. Starter-tagged catalogs are automatically picked up by the product selection agent as companion / starter product suggestions during the narrowing phase, replacing the previous hard-coded companion-types configuration. Pairs with the new "Catalog type" selector available in the source configuration screen (see Admin Experiences).

• Email → Pro Agent (end-to-end): Hubs that have Pro Agent enabled can now receive customer emails at a dedicated agent email address. The agent picks up the email thread, runs its multi-step / long-horizon work, and replies inline in the same thread — typically over the course of 20–30 minutes per turn. Projects attached to the hub (see Admin Experiences) are auto-attached to the conversation as read-only reference material. Initial rollout is on a single customer agent alias.

• Projects: hub-scoped reference material for the Pro Agent: Admins can upload files (or whole folders) to a hub as a Project, and the Pro Agent will mount the attached projects read-only during each conversation turn. The platform supports both native-stored projects and projects backed by an external store such as Google Drive. The corresponding super-admin management surface is documented in Admin Experiences.

• Discord forum opt-in mode with delayed reply: A new opt-in posting mode for the Discord agent on forum channels — instead of replying immediately to every new post, the agent presents an opt-in greeting card on each post and falls back to an auto-reply after a configurable delay if the user does not respond. Reduces noise on busy forums while still capturing the long-tail of unanswered questions. Per-hub configurable.

• Image and PDF attachment safety screening: The conversation safety filter now downloads and classifies image and PDF attachments alongside the user's text query, before they reach the agent. Previously only text attachments (folded into the query) were checked, so image and PDF attachments bypassed the filter entirely.

Improvements

• Product-discovery filter loop hardened: Fixed a tail case where the product-lookup-by-filters tool could spin in an infinite retry loop on a no-match query. The tool now caps retries, falls back to keyword and semantic search, and verifies recommended products against the user's stated hard constraints before returning them.

• Faster answers on retrieval-heavy questions: The retrieval sub-agent now stops searching as soon as the documents it has saved are sufficient to answer the question (hard multi-hop queries still iterate as needed). On retrieval-heavy queries this cuts search rounds by ~35% with answer quality held — translating into noticeably faster answers on long, multi-step questions.

• Unified conversation origin model: Every conversation is now recorded with a consistent channel (Web, Slack, Discord, Email, Public API) and, for the web channel, a form factor (embedded launcher vs AI Overview). This underpins the new analytics breakdowns in the admin dashboard (see Admin Experiences).

• Per-turn agent telemetry: Latency milestones and token counts are now captured on every AI message across our agent stack, forming the foundation for conversation-level observability (see the admin-only Agent Stats popover in Admin Experiences).

Bug Fixes

• Source citations restored on keyword-classified queries: Closed a long-tail gap where keyword-routed answers were skipping the standard citation and answer-generation path, occasionally producing "I can't find the part" responses even when the source was indexed. Keyword-classified answers now consistently cite their supporting sources, on parity with the other query types.

• reCAPTCHA hard-block on transient network retries: A hard-block bug on embedded flows — where an HTTP/2 network-level retry was being flagged by reCAPTCHA as a duplicate token and rejecting real users — has been fixed. The client now transparently retries with a fresh token.

May, 2026

New Features

• End-to-end agent quality evaluations: We've improved our evaluation framework to handle more nuanced agent verification.

• Modernized faster and better Slack app: Our slack app has been fully revamped and is now better, faster and full featured!

• Hub reference material via file uploads: Admins can now upload reference documents directly to a hub via a new files API. The agent can read and cite these files as ground-truth reference material starting on the next conversation turn, giving operators a fast path for steering agent behavior with first-party docs without a full ingestion cycle.

• Web-search-augmented answers for product selection: The product selection agent can now consult a web-search signal in parallel with the knowledge base, using it to cross-check and augment expert answers from indexed sources. Currently enabled for select customers, with answer quality and citation parity verified before broader rollout.

Improvements

• More flexible model provider routing: New gradual-rollout controls allow percentage-based traffic shifts to alternate model providers. This enables smoother provider migrations without disrupting customer workloads.

• Table-aware retrieval for datasheet-style sources: A new table-aware chunking strategy preserves column-header context on every table cell when chunking documents, so row-level technical meaning survives chunk boundaries. This materially improves retrieval over parameter tables and spec sheets. Available per source via the source's chunking configuration.

• Citations on product selection and technical support answers: Both agent types now cite source documents back inline within their responses, bringing them in line with citation support on our core conversational agent.

• Faster image-heavy PDF extraction: PDF extraction for image-rich datasheets now uses a provider-native code-execution capability under the hood, materially reducing extraction latency and cost while keeping output quality on par. The image-description step used during ingestion is also standardized across the platform.

• Broader model provider coverage: The platform's internal model invocation API now spans an additional provider backend alongside the existing ones, giving us more headroom to route per-customer and per-request to the best-fit option without disrupting customer workloads.

Bug Fixes

• Disabled sources no longer influence retrieval via duplicate URLs: Resolved a long-tail correctness issue where, if two sources contained the same URL (for example a legacy and a new extractor against the same domain) and one source was disabled, the disabled source's content could still appear in retrieval results. Retrieval now reliably honors source enable/disable state in all cases.

April, 2026

New Features

• Visual generation is now first-class: Agent can generate diagrams and structured visuals (Mermaid, HTML, SVG etc.) via a dedicated harness node for richer agent conversations.

• Discord improvements: Interactive agent in Discord capable of asking clarifying questions as Multiple Choice Questions (MCQs) with buttons, select menus etc. Agent can now auto-reply in forums, and handle sessions in more significant ways.

• Email inbox for AI agents (experimental): Agents can now interact with customers via inbound email address, expanding AI support beyond web and Discord.

Improvements

• Product selection agent model upgrade: All agent responses now use the latest model generation, improving answer quality across the board.

• Product selection agent citations: Improved citations back to references that contribute to answer quality.

• More relevant retrieval results: Search retrieval now preserves top-k diversity after deduplication.

Bug Fixes

• Product selection agent tool dispatch restored: Fixed a regression that dropped tool executions from conversation history.

January, 2026

New Features

• Discord Bot Integration: Full Discord integration allowing AI Agent to respond to questions directly in Discord channels with support for public replies, ephemeral messages, and threaded conversations.

• Discord Forum Auto Reply: The Discord bot now monitors forum channels and automatically responds to new posts using the AI Agent, without requiring users to invoke the /ask command. Requires the Message Content Intent to be enabled in the Discord Developer Portal.

• Inline Citations: AI responses now include inline citations that reference source documents directly within the answer text, providing clearer attribution and easier verification of information sources.

• Promoted Answers: Configure specific responses for frequently asked questions that take precedence over AI-generated answers, ensuring consistent, fast and accurate responses for common queries. Promoted answers will also power our future organizationally visible, trending questions section, newsletters and so on.

• Automatic QA Evaluation: The system now automatically creates QA evaluations for runtime errors and bad feedback, helping teams identify and address quality issues proactively.

• Salesforce Ticket Creation: AI Agent can now create support tickets directly in Salesforce when human handoff is requested or when specific conditions are met.

Improvements

• Enhanced PDF Processing: Significant improvements to PDF extraction, with support for persisting and describing the semantic content of images for use in later question answering and presentation.

• Improved Citation Handling: Citations now properly display URLs and support PDF references with better formatting.

November, 2025

New Features

• Web Search Capability: AI Agent can now search the web for real-time information when the indexed knowledge base doesn't contain the answer, providing more comprehensive responses.

• Reasoning Transparency: AI now streams its reasoning process ("thinking aloud") to users, showing the steps taken to formulate an answer for better transparency.

Improvements

• State-of-the-Art Model Support: Added support for the latest frontier models with enhanced reasoning capabilities, enabling more accurate and reliable responses for complex, multi-step queries.

September, 2025

New Features

• Text Database Upgrades - We've moved on to a significantly more capable, performant and efficient chunk database. Allowing us to store even greater content, without sacrificing retrieval performance.

Improvements

• Upgraded Default Model: Default AI model upgraded to the latest high-performance model for improved response quality and faster generation.

• Enhanced Reranking: Improved AI-based reranking for document relevance, resulting in more accurate search results and better answers.

• Abuse Prevention: Enhanced prompt engineering to prevent misuse and ensure AI Agent stays on-topic.

July, 2025

New Features

• Rapid Sales Agent: Our new AI agent - a guided product discovery experience that takes you through a series of clarifying questions to find the right products, iteratively exploring your product catalogs.

• Search Improvements: Significant improvements to search for improved retrieval accuracy, especially for technical terms, part numbers, and product codes.

Improvements

• Query Processing Improvements: Enhanced query rephrasing with better handling of conversation history for more accurate multi-turn conversations.

• Self Service Structured Reasoning Guidelines: More granular control over AI behavior with component level guidelines.

May, 2025

New Features

• Product Selection: New AI capability for helping users select products based on their requirements, with support for:

  • Use case-based product lookup
  • Filter-based product search
  • Product comparison capabilities
  • Streaming responses for real-time feedback

• Multi-Agent Support: Organizations can now have multiple AI Agents with different configurations, knowledge bases, and behaviors tailored to different use cases.

• Conversation Analytics Enhancements: New analytics including breakdown by source URL, related documents by type, and improved trend analysis.

Improvements

• AI-Powered Reranker: New document reranking that uses AI to improve the relevance of search results before generating answers, leading to higher quality responses.

• Improved Error Handling: Better error handling throughout the system, with automatic quality evaluation creation for issues to help identify and resolve problems faster.

March, 2025

New Features

• Explainability Events: AI Agent now emits detailed events explaining its reasoning process, which can be displayed to users or logged for analysis.

  • Configurable per organization
  • Shows query classification, retrieval steps, and answer generation process

• Flagged Conversation Email Alerts: Automatically receive email notifications when conversations are flagged or receive negative feedback.

• Human Agent Handoff: Support for transferring conversations to human support agents with configurable conditions and email notifications.

Improvements

• Multi-Provider AI Support: New unified AI layer supporting multiple leading AI providers, allowing automatic selection of the best model for each task type.

• Faster Response Times: Improved caching for agent configurations and catalog data, resulting in noticeably faster response times.

• Enhanced Security: Added a comprehensive security layer to all API responses.

Dec 6, 2024

Improvements

• Numerous QoR (Quality of Results) improvements to the query type classification based question answering AI workflow.
• Better resiliency and retry mechanisms to allow for AI to recover from model communication issues.
• AI generated titles for conversations before completing them
• Query classification now covers a broader range of types and are stored as part of our ongoing analytics
  • Coming Soon (Analytics dashboard access to this and more!)
• Improvements to accuracy of related content suggestions presented as part of the conversation
• Better history interpretation in a multi-turn conversation

Oct 25, 2024

We are catching up on our August to October changes via this release note.

Improvements

• New query type classification based, multi-step question answering flow
• New - Product Catalog and Product Glossary based query types for:
  • Looking up technical specifications for one or more products
  • Looking up products that match certain technical specifications
  • Comparing two or more specific products
  • Search by keywords
  • General queries
• AI Explainability - now you can see clearly what steps our AI workflows took in order to present a result.
• Artifacts!
  • The specialized query types above will now present artifacts such as comparison tables, data sheets allowing you to get an even more dynamic answer, plus you can download those artifacts as PDFs, share it with others etc.

Talk to us if you are interested in a product catalog powered expert assistant!

Aug 9, 2024

We are catching up on our July release notes in this early August release notes update.

Improvements

• We've improved our automated evaluation metrics for our proprietary RAG system.
• We've added initial query enrichment to improve answers for brief or terse queries (select customers)
• We have made significant improvements in automated ingestion systems, though we are still a few more weeks away from productizing it and providing self service control.
• We have improved our conversation history document schema, and now can add more useful insights by analyzing conversations after they are complete, and to provide richer, more useful analytics.

June 21, 2024

New Features

Improvements

• Query type specific additional guidelines and prompts help us surgically handle edge cases for that type of question.

Bug Fixes

• More robust error handling when conversation streams encounter errors. There's now feedback in the AI Agent UX about the error and a rendering of the partial answer in case the error happened mid-flight in a conversation.

June 7, 2024

New Features

Improvements

• For customers with product catalogs, we now also handle use case type questions. Given a use case, we can find matching products and answer specific attributes of the matching products.
• We've improved our internal tracing systems allowing us to track all steps involved in the agent's AI workflow to find issues and improve accuracy
• In addition to keyword and semantic search, we now support fuzzy search of product catalog terms
• Latency improvements for our advanced model. 90% of questions are now answered ~5-10 seconds faster.

Bug Fixes

May 24, 2024

New Features

• Multi-step question answering with query classification, query rewriting, product catalog based retrieval

Improvements

• Retriever improvements when retrieved information overflows context

Bug Fixes

May 24, 2024

New Features

• AI Agent now retrieves related videos, learning modules for customers who provide those as sources.

• More advanced agentic retrieval and question answering system

• AI Agent now supports Advanced AI models.

  • Our Advanced models follow instructions better, are more robust at answering questions and in general more powerful for a variety of tasks, with greater world knowledge.
  • Contact us at support@rapidflare.ai if you would like to set your enterprise to this more advanced model.

  Note: These models will take longer to generate responses.

• Responses from the AI Agent will now be always streamed as they are being generated. Since advanced models take longer to complete this will enable humans to start processing the AI Agent's responses as they are being generated.

Improvements

Bug Fixes