> ## Documentation Index
> Fetch the complete documentation index at: https://docs.flowx.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# FlowX.AI 5.11.0 Release Notes
> What's new in FlowX.AI 5.11: workflow tools and agentic RAG on Custom Agents, per-Knowledge-Base embedding models, document classification, extended response caching, subworkflow data mapping, structured template data models, and chat and UI Flow renderer improvements.
**FlowX.AI 5.11 is a SaaS release.** New capabilities are available on managed (SaaS) deployments; self-hosted deployments will receive them with the next LTS release family.
**Deployment models:**
* SaaS Managed FlowX cloud deployments
* All Deployments Available in both self-hosted and SaaS
FlowX.AI **5.11** is a SaaS-managed release: FlowX handles deployment-level changes such as infrastructure, Kafka topics, and service configuration. This release turns Custom Agents into orchestrators that call FlowX workflows as tools and search knowledge bases agentically, adds per-Knowledge-Base embedding models with local embedders for personal-information content, and introduces document classification. It also extends response caching and subworkflow data mapping in the Integration Designer, moves document and notification templates onto the structured data model, and adds chat notifications, session ownership, and richer runtime audit to the renderers.
### What changed since 5.10.0
| 5.10.0 | 5.11.0 |
| ------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------- |
| Custom Agents call MCP servers, built-in tools, web search, and Knowledge Bases | + **FlowX workflows as callable tools** on a Custom Agent, with per-tool usage hints and tool-calling caps |
| Knowledge Base retrieval is a plain vector search on the org default embedding model | + **Agentic knowledge base search** and **per-KB embedding models**, including keyless local embedders for personal-information content |
| REST caching covers GET; SQL operations are cacheable | + **POST** REST responses, **file-storage** reads, and **Mongo** reads are cacheable too |
| Subworkflow input/output is raw JSON | + **Input and output mappers** on the Subworkflow node, with per-mapping test and import/export |
| Document and notification templates use a free-form key/value Data model tab | + Templates use the **structured data model** with typed attributes, validators, and Input Parameters |
***
**What's new?**
๐งฐ [**Workflow tools on Custom Agents**](#workflow-tools-on-custom-agents) - Attach FlowX workflows as callable agent tools, with usage hints and tool-calling caps\
๐ [**Agentic knowledge base search and Show Reasoning**](#agentic-knowledge-base-search-and-show-reasoning) - Query decomposition, relevance grading, and a reasoning summary in the run log\
๐ง [**Per-knowledge-base embedding models**](#per-knowledge-base-embedding-models) - Pick an embedding model per KB and keep personal-information content on a local embedder\
๐๏ธ [**Knowledge Base content lifecycle**](#knowledge-base-content-lifecycle) - Delete entries by metadata filter, structured failure handling, per-chunk timestamps\
๐ท๏ธ [**Document classification**](#document-classification) - Train or import a classifier and route document pages in a workflow\
๐ฌ [**Conversational chat reply requirement**](#conversational-chat-reply-requirement) - A conversational run that never replies to chat now fails with a clear error\
โ๏ธ [**Custom Agent instruction limit**](#custom-agent-instruction-limit) - The Instructions field is capped at 10,000 characters
๐ [**Reusable functions in Integration Designer**](#reusable-functions-in-integration-designer) - Call reusable functions from Script and Condition (Fork) nodes\
โก [**Extended response caching**](#extended-response-caching) - Cache POST REST, file-storage reads, and Mongo reads alongside GET and SQL\
๐ [**Subworkflow data mapping**](#subworkflow-data-mapping) - Map data in and out of a subworkflow with input and output mappers\
๐ [**Email attachment filtering by extension**](#email-attachment-filtering-by-extension) - Filter attachments by individual file extension instead of broad categories
๐งฑ [**Structured data model for templates**](#structured-data-model-for-templates) - Document and notification templates use typed attributes and Input Parameters\
โ
[**Data model attribute validators**](#data-model-attribute-validators) - Required, min/max, length, pattern, email, items, and currency rules on attributes\
๐งพ [**UI Flow runtime audit**](#ui-flow-runtime-audit) - Audit workflow-instance runs, chat messages, conversation reads, and process renders\
๐ค [**Cross-version export adaptation**](#cross-version-export-adaptation) - Exports adapt UI templates for older LTS targets by removing unsupported handlers\
๐ [**Platform and runtime improvements**](#platform-and-runtime-improvements) - Runtime access-control hardening
๐ [**Chat notifications and conversation search**](#chat-notifications-and-conversation-search) - Unread-message badge on the chat button and search across conversation history\
๐ซ [**Disable chat input on a condition**](#disable-chat-input-on-a-condition) - A business condition can make the chat input read-only\
๐งญ [**Navigate To form validation**](#navigate-to-form-validation) - Require form validation before a Navigate To action, on web, iOS, and Android\
โ [**Expired UI Flow sessions**](#expired-ui-flow-sessions) - Expired sessions show a session-expired screen with a start-fresh option\
๐ค [**UI Flow session owner**](#ui-flow-session-owner) - Restrict a running session to a single user\
๐ [**Process Renderer process-end handling**](#process-renderer-process-end-handling) - Run a configurable action when an embedded process ends\
๐๏ธ [**Renderer improvements**](#renderer-improvements) - Batched enumeration fetch and process-start loader alignment
### **Workflow tools on Custom Agents**
**Before 5.11:** a Custom Agent's tools were MCP servers, built-in tools, web search, and Knowledge Base retrieval. **Now:** you can also attach FlowX workflows as callable tools.
In the Custom Agent node's **Tools** section, add a workflow as a tool and fill the per-tool **When should the agent use this workflow?** hint that tells the agent when to call it. The agent calls the workflow like any other tool, passing arguments that match the workflow's Start-node input parameters. A workflow with no input parameters cannot be attached. Each tool call starts a fresh workflow run whose output returns to the agent, and the child run is linked in the console log.
See the Custom Agent node documentation.
Attach up to ten workflows as tools on one Custom Agent, each with a usage hint the agent reads when deciding to call it.
The agent passes arguments that match the workflow's Start-node input parameters; workflows without input parameters are not eligible.
An agent is limited in how many tool calls it can make per run, and workflow-tool call chains are depth-limited so recursion cannot run away.
Each workflow tool call is a fresh run whose output returns to the agent and is linked in the console log.
SaaS
### **Agentic knowledge base search and Show Reasoning**
The Custom Agent node gains **agentic knowledge base search** and a **Show Reasoning** toggle.
With agentic search on, the agent breaks a question into sub-queries, retrieves and grades results for relevance, and retries with reworded queries if the first pass finds nothing. It is better on complex questions and uses more model calls; a single pass with three sub-queries is the default. Agentic search never returns worse results than plain retrieval. **Show Reasoning** captures the model's reasoning summary in the run and console log; the structured response and the chat reply shown to the user are unchanged.
See the Custom Agent node documentation.
Turn on **Enable agentic knowledge base search**, then set **Max attempts** and **Subqueries per attempt** under Settings.
Retrieved candidates are graded against the original question, with corrective retries until a relevant result is found or the attempt budget is spent.
Surface the model's reasoning in the run log without changing the answer returned to the user.
Agentic mode returns the same chunk metadata as plain retrieval, so citations and telemetry are unchanged.
SaaS
### **Per-knowledge-base embedding models**
Knowledge Bases can now use a **per-KB embedding model**, and a KB that holds personal information can keep its content on a local embedder.
Turn on **Containing Personal Information** on a Knowledge Base to enable a local embedding model, then pick one from the **Embedding model** list. Local embedders (for example a text-embedding-inference or Ollama endpoint, or any OpenAI-compatible endpoint) are configured as a single **CUSTOM** provider in the organization AI settings, with a free-form base URL and no API key. The embedding model is set before the first store and cannot be changed afterward. Non-personal-information Knowledge Bases use the organization's default embedding model.
See the Knowledge Base documentation.
Each Knowledge Base carries its own embedding-model reference; a new Settings tab shows the current model.
Keep personal-information content on-box using a keyless local embedder fronted by a CUSTOM AI provider.
Project version export preserves the KB's embedding-model reference, so a re-import keeps the same model.
A Knowledge Base's embedding model is chosen before its first store and cannot be changed afterward.
SaaS
### **Knowledge Base content lifecycle**
The **Update Knowledge Base** node gains a **Delete entries** operation and structured failure handling, and retrieval results now carry per-chunk metadata.
**Delete entries** soft-deletes chunks whose entry-level metadata matches a filter, in one store or across the Knowledge Base; an empty or match-all filter is rejected so a misconfigured node cannot wipe a store. When an update operation fails, the node's **FAIL** branch returns a structured output with the operation, store, failure reason, and how many chunks were rolled back, and any partial changes are reversed. Retrieval results now include a per-chunk identifier and an ingestion timestamp.
See the Managing Knowledge Base content documentation.
Delete entries by metadata filter from a workflow, with an empty or match-all filter rejected as a safety guard.
The FAIL branch reports the operation, store, reason, and rolled-back chunk count; partial changes are reversed.
RAG results carry a chunk identifier and an ingestion timestamp (empty for chunks indexed before this shipped).
Retryable failures such as extraction or chunking errors are distinguished from non-retryable ones such as size or timeout.
SaaS
### **Document classification**
FlowX.AI adds **Document Classifiers**: deterministic ML models that sort document pages into categories. Define classes, upload training data, and train or import a model, then route each page in a workflow with the new **Document Classification** node.
The node runs the active model over a document, returns a predicted label and probability per page with a configurable confidence threshold, and flags low-confidence pages. A page-level failure does not fail the node.
See the Integration Designer documentation.
Build classes and training data in the Designer, then train a model or import a pre-trained one.
The Document Classification node returns a label and probability for each page, with a confidence threshold and low-confidence flags.
Combine the node with a condition to route pages by their predicted category.
A page that cannot be classified is reported per page without failing the whole node.
SaaS
### **Conversational chat reply requirement**
A conversational (chat-driven) workflow must send at least one chat reply, either from a Custom Agent node with **Send as Chat Reply** on or from a Navigate in UI Flow node that completes the turn. A run that reaches the end without replying now fails with a clear error instead of ending silently.
Learn more about [chat-driven workflows](/5.9/ai-platform/chat-driven-workflows).
SaaS
### **Custom Agent instruction limit**
The Custom Agent **Instructions** field is now limited to 10,000 characters.
SaaS
### **Reusable functions in Integration Designer**
Reusable functions can now be called from **Script** nodes and **Condition (Fork)** nodes in the Integration Designer, not just from Business Rule actions in processes. Functions are referenced in the node's JavaScript or Python and resolved at runtime from the app version's function copies, including functions that call other functions and cross-app references. A Fork node resolves the references used across all of its branch conditions.
For more information, see the [reusable functions](/5.9/docs/building-blocks/reusable-resources/reusable-functions) documentation.
SaaS
### **Extended response caching**
Response caching in the Integration Designer now covers more operation types. **POST** REST endpoints are cacheable alongside GET, with the request body forming part of the cache key so different bodies cache separately. Read **file-storage** operations (LIST and DOWNLOAD) and read **Mongo** operations (FIND, FIND\_ONE, AGGREGATE) are cacheable alongside SQL. Mutating operations are never cached.
For more information, see the [Integration Designer](/5.9/docs/platform-deep-dive/integrations/integration-designer) documentation.
Cache POST responses, keyed on the request body so distinct requests stay separate.
LIST and DOWNLOAD operations are cacheable; UPLOAD, DELETE, and MOVE are not.
FIND, FIND\_ONE, and AGGREGATE are cacheable alongside SQL; mutating Mongo operations are not.
All cacheable surfaces reuse the same time-based expiry and per-operation clear-cache action.
SaaS
### **Subworkflow data mapping**
The **Subworkflow** node adds data mapping as an alternative to raw JSON. Turn on **Enable data mapping** to define an **Input mapping** per subworkflow start node and an **Output mapping** per end node, mapping individual data slices instead of hand-editing JSON. Output mapping does not apply to asynchronous subworkflows. Each mapping opens a mapper with a test action, and mappings travel with import, export, and clone.
For more information, see the [Integration Designer](/5.9/docs/platform-deep-dive/integrations/integration-designer) documentation.
Map data into the subworkflow per start node and out of it per end node, without editing raw JSON.
Test a mapping with source and target slices and apply values from a past instance.
Mappings are preserved on import, export, and clone.
Output mapping is hidden for asynchronous subworkflows, which do not return data to the caller.
SaaS
### **Email attachment filtering by extension**
Email Trigger and Microsoft Outlook data sources now filter attachments by **individual file extension** instead of the five broad file-type categories. Choose the exact extensions to accept and to forward into a process. Existing category selections are migrated to their individual extensions automatically on upgrade, so current triggers keep working.
For more information, see the [Email Trigger](/5.9/docs/platform-deep-dive/integrations/email-trigger) documentation.
SaaS
### **Structured data model for templates**
**Before 5.11:** document and notification templates declared parameters in a free-form key/value Data model tab. **Now:** templates use the same structured data model as the rest of the platform.
On the **Data Model** tab you define typed attributes with validations, sensitive-data and reporting flags, and references to project or library data types. On the **Input Parameters** tab you select which attributes the template expects. Values sent in the request are substituted for those attributes at generation time.
See the notification templates documentation.
Define string, number, boolean, enum, date, currency, object, and array attributes with validations.
Pick the attributes the template expects on a dedicated Input Parameters tab.
Reference project and library data types like other resources.
The OTP flow contract is unchanged; the reserved `otpValue` and request `data` object work as before.
SaaS
### **Data model attribute validators**
Data model attributes now support validators. Available rules depend on the attribute type and include **Required**, **Min/Max Value**, **Min/Max Length**, **Pattern**, **Email Format**, and **Min/Max Items**, plus currency-specific amount and code rules. Validators are stored on the attribute and apply wherever that data model is used, including project, process, template, and reusable UI template data models.
For more information, see the [project data model](/5.9/docs/building-blocks/project-data-model) documentation.
SaaS
### **UI Flow runtime audit**
UI Flow runtime sessions are now audited in more detail. The audit log records **workflow-instance** runs, chat **message sends**, conversation **list** and **history** reads, and whether the Process Renderer performed a **render** or a start. Entries name the app and the workflow instance.
For more information, see the [audit](/5.9/docs/platform-deep-dive/core-extensions/audit) documentation.
SaaS
### **Cross-version export adaptation**
When you export a project to an older platform version, UI-template features the target version cannot understand are removed during down-adaptation, so the export imports cleanly. This currently covers Process Renderer process-end event handlers and the Navigate To **forms to validate** setting.
For more information, see [Cross-version export and import](/5.9/docs/flowx-designer/managing-a-project-flow/export-import-a-process-definition#cross-version-export-and-import).
All Deployments
### **Platform and runtime improvements**
Runtime API requests are now authorized against the roles defined in the running build across start-process, start-workflow, task views, conversation history, and the public runtime API. Requests without a matching build role are denied, and enforcement is on by default.
All Deployments
Exporting a theme no longer includes duplicate copies of built-in system icons, and importing a theme no longer creates new ones. Duplicate icon copies that accumulated in earlier versions are cleaned up automatically on upgrade โ no action is required.
All Deployments
### **Chat notifications and conversation search**
The chat floating action button gains an **unread-message notification badge**, with a themeable **Notification Dot** you can position in the chat style panel. The chat panel also adds **search across conversation history**, which filters conversations by a text query, highlights matches, and shows an empty state when nothing matches.
For more information, see the [Chat component](/5.9/docs/building-blocks/ui-designer/ui-component-types/chat) documentation.
The chat button shows a badge with the unread agent-message count; it is hidden while the chat is open in full screen.
Position and size the notification dot from the chat style panel, with theme support.
Search conversation history from the chat panel, with highlighted matches and an empty state.
SaaS
### **Disable chat input on a condition**
The chat **Input** component can be disabled by a business-condition expression. When the condition evaluates true, the input becomes read-only with an explanatory tooltip, and any pending attachments and voice input are cleared.
For more information, see the [Chat component](/5.9/docs/building-blocks/ui-designer/ui-component-types/chat) documentation.
SaaS
### **Navigate To form validation**
A **Navigate To** action can now require selected forms to pass validation before navigating, matching how Start Process and Start Workflow actions gate on validation. This is available on web, iOS, and Android.
For more information, see the [UI Flows](/5.9/docs/building-blocks/ui-flows) documentation.
SaaS
### **Expired UI Flow sessions**
Expired UI Flow sessions now show a **session-expired** screen with a start-fresh option instead of leaving the previous data visible in read-only form. Expired and unauthorized access is written to the audit log.
SaaS
### **UI Flow session owner**
A UI Flow can restrict a running session to a single owner. When **Allow multiple editors** is off, other users who try to open the session are shown an access-denied screen, and unauthorized access is audited. Newly created flows default to single-owner; imported and existing flows keep the previous shared behavior. A session-owner value type is available for permission-based hide and disable rules.
For more information, see the [UI Flows](/5.9/docs/building-blocks/ui-flows) documentation.
SaaS
### **Process Renderer process-end handling**
The Process Renderer can now run a configurable action when its embedded process ends, using a new process-end trigger on event handlers. Configure what happens at that point instead of leaving the renderer idle when the process completes.
For more information, see [Process end handling](/5.9/docs/building-blocks/ui-flows#process-end-handling).
SaaS
### **Renderer improvements**
* **Batched enumeration fetch:** renderers batch enumeration lookups into a single bulk request, cutting the number of calls when rendering enum-backed fields. Applies to web and mobile renderers.
* **Process-start loader alignment:** the web renderer keeps the process-start loader visible until the first user task, matching the mobile SDKs.
SaaS
***
## Shipped by
This release is the work of the FlowX engineering team across backend, frontend, AI, infra, QA, and docs. It also belongs to the customers who told us what was broken last quarter. Thank you.
***
## Deployment guidelines
FlowX.AI 5.11 is a SaaS-managed release: FlowX handles deployment-level changes such as infrastructure, Kafka topics, and service configuration. For component versions and infrastructure notes, see the [5.11 deployment guidelines](./deployment-guidelines-v5.11).
## Related topics
- [FlowX.AI 5.12.0 Release Notes](/release-notes/v5.x/v5.12.0-september-2026/v5.12.0-september-2026.md)
- [FlowX.AI 4.7.11 Release Notes](/release-notes/v4.x/v4.7.x-lts/v4.7.11-december-2025/v4.7.11-december-2025.md)
- [FlowX.AI 5.1.11 Release Notes (LTS)](/release-notes/v5.x/v5.1.x-lts/v5.1.11-july-2026.md)
- [FlowX.AI 5.1.12 Release Notes (LTS)](/release-notes/v5.x/v5.1.x-lts/v5.1.12-august-2026.md)
- [FlowX.AI 5.2.0 Release Notes](/release-notes/v5.x/v5.2.0-november-2025/v5.2.0-november-2025.md)