h2oGPTe v1.7.4​

New features​

OpenAI-compatible API​

• Full tool/function-calling support: The OpenAI-compatible /v1/chat/completions and /v1/responses endpoints now support the complete tool-calling round trip, with both streamed and non-streamed responses returning tool calls.
  • Requests can pass tools, tool_choice, and parallel_tool_calls, or the legacy functions/function_call form, which is automatically translated to the modern form.
  • /v1/responses also emits per-call streaming lifecycle events for tool calls.
• Additional passthrough parameters: response_format, logit_bias, stop, reasoning_effort, frequency_penalty, presence_penalty, and user are now forwarded to the upstream model on both endpoints.

Bug fixes​

OpenAI-compatible API​

• Fixed chat completion choices returning the index field under the key id instead of index, which could break clients that correlate parallel or streamed choices.
• Fixed non-streamed responses being sent without a Content-Type header, which crashed the openai-node SDK and integrations built on it, such as n8n.

Authentication​

• Fixed users who signed in with only a bearer token showing a blank email address, for example when sharing a collection. The email is now backfilled from the JWT claim.

Python client​

• Fixed the client crashing on construction when a proxy, ingress, or WAF redirected its cookie-session capability check to an HTML page instead of a JSON response.

h2oGPTe v1.7.3-post1​

New features​

Agents and tools​

• Auto-generate system prompts for General Code tools: When creating a General Code tool, click the sparkle button next to the System prompt field to auto-generate a tailored system prompt from your uploaded code.
• Agent Tool Server: Agent tools can now run in a dedicated code-execution mode that skips LLM reasoning between steps, giving faster, more deterministic execution for scripted workflows.
• Download custom tool code: You can now download the Python code behind your local tools and General Code tools directly from the UI.
• Custom agent integration: h2oGPTe now supports integrating custom agents into your workflows.

RAG and AI models​

• Graph RAG: Introducing Graph RAG, a graph-based retrieval approach that improves answer quality for questions that span relationships across multiple documents.
• New models added: Support for Kimi-K2.6, DeepSeek-V4-Pro, Qwen3.6-35B-A3B, and MiniMax-M3, all available via OpenRouter.

Security and guardrails​

• Guardrail Insights: A new Guardrail Insights page shows violation metrics, helping administrators track and analyze guardrail activity across the platform.
• Multimodal guardrails: Guardrails can now evaluate multimodal content in agent conversations, extending protection beyond plain text.

Chat and collaboration​

• Share Chat Replay: Shared chats can now be replayed by recipients, turning a static share into a fully replayable conversation.
• Showcase: Administrators can curate a set of featured chats on a new Showcase page to highlight best-practice examples across the organization.

Documents and collections​

• .mov file ingestion: h2oGPTe now supports ingesting .mov video files into your collections.
• Extractors without a collection: Extractors can now run directly on uploaded files, even before they are attached to a collection.
• SharePoint drive discovery: The SharePoint Online connector can now automatically discover available drives, or you can configure a drive manually if automatic discovery doesn't find it.

Scheduled tasks​

• Scheduled-task notifications: Configure email notifications for scheduled tasks using customizable HTML templates.

Authentication and access control​

• Keycloak session token forwarding: Custom and MCP agent tools can now receive the calling user's Keycloak session token, enabling authenticated tool calls without separate credential configuration.
• Azure AD transparent authentication: Users signing in through an Azure AD identity provider are now authenticated transparently, with Keycloak automatically refreshing the underlying token in the background.
• Admin session management: Administrators can now view all active user sessions and revoke them individually from a new dashboard.
• Admin User Limits dashboard: A new dashboard lets administrators browse and export (CSV) per-user usage limits, with infinite scroll for large user lists.
• Per-user concurrent limits: Administrators can cap the number of concurrent ingestion jobs and chat sessions per user, improving fairness in multi-tenant deployments.
• HAIC audit trail integration: h2oGPTe can now stream a comprehensive audit trail to the HAIC Audit Trail service for compliance and governance tracking.

Improvements​

OpenAI-compatible API​

• Response shaping: Improved response shape and added several related fixes across the OpenAI-compatible chat completions and responses endpoints.

Authentication and security​

• Stronger agent key protection: Agent keys are now stored more securely.
• Admin bypass for agent key cleanup: Administrators can now delete any user's agent keys and their associations, not just their own.
• Document download restrictions: Document downloads are now restricted based on your permissions.
• PII gating for backend logs: Administrators can now gate personally identifiable information out of API gateway logs, extending the existing log-redaction control to cover all services.

Documents and collections​

• Type-aware document preview: Document previews now adapt to the file type for more reliable in-app viewing.
• Improved OCR: Upgraded the underlying OCR engine for improved text recognition accuracy and format support.
• Source column for connectors: Connector job views and the collection connectors panel now show a Source column, making it easier to trace where each document came from.
• Faster connector ingestion: Ingestion throughput from Azure Blob Storage and Google Cloud Storage connectors has been improved through network tuning.

Memory Blocks​

• Agent tool memory mode: Memory Blocks now support an agent_tool injection mode. Instead of always being pre-injected, memory content can be read and updated by the agent on demand, through explicit tool calls.
• New system prompt controls: New settings let you control system prompt addenda for LLM and agent chats independently.

Models​

• Reasoning on by default: gpt-oss-120b agents now use high reasoning mode by default for more thorough responses.
• Model update: Replaced gemini-3-pro-preview with gemini-3.1-pro-preview.

Scheduled task reliability​

• Clearer task status: Scheduled task status is now split into separate lifecycle and execution states, making it easier to understand what a task is doing.
• Automatic expiration: Scheduled tasks now automatically expire once they pass their configured expiration date.

System settings and administration​

• Expanded runtime settings: Many more settings can now be changed at runtime by administrators without a redeploy.

User interface​

• Faster dropdowns: Dropdown menus across the app now load on demand, making them faster and more scalable for large lists.
• Clearer empty states: Empty states throughout the app now include filter-aware messaging and consistent phrasing.
• Tidier job UI: The job log link is now hidden when live logs are disabled, removing a link that would otherwise lead nowhere.

Chat and collaboration​

• Chat response timing breakdown: The Usage Stats dialog under a chat response now shows a visual timing breakdown — queueing, file loading, retrieval, first token, and generation — alongside cost and token throughput, instead of raw JSON.

Vector database (VEX)​

• pgvector by default: New deployments now default to pgvector as the VEX vector storage backend.
• Improved migration tooling: Collections can now be migrated to pgvector incrementally, and a new admin UI shows live migration progress.
• Per-collection vector encryption: Choose whether to encrypt vector data on a per-collection basis; full-text search encryption now follows your vector encryption setting automatically.
• Postgres Transparent Data Encryption: On-prem deployments can now enable Transparent Data Encryption for Postgres, encrypting vector data at rest.

Python client​

• Adds full CRUD and sharing support for Memory Blocks.
• Adds CRUD support for scheduled tasks, including collection filtering and per-field configuration patching.
• Adds an endpoint to list recently added documents with summaries, sortable and filterable by type.
• Adds a tag deletion method for metadata cleanup.

Bug fixes​

Agents and tools​

• Fixed code execution occasionally failing due to a missing tool directive, which is now added automatically.

RAG and GraphRAG​

• Fixed GraphRAG defaulting to the the most affordable model currently available for graph building; it now automatically selects a structured-output-capable LLM, improving graph quality and reliability.
• Fixed an empty-graph issue in GraphRAG caused by a NumPy embedding crash; collections no longer report as ready when graph construction fails.

Chat and UI​

• Fixed chats not auto-scrolling to the latest message when opened.
• Fixed elapsed time and the thinking indicator not stopping correctly at the end of a response.
• Fixed a bug that allowed sharing a chat while its response was still streaming; sharing is now blocked until the response completes.
• Fixed the Share action being visible on chats you don't own.
• Fixed shared chat file and thumbnail links, and restored share access for logged-out viewers in production.
• Fixed various UI polish issues, including scrollbar, alignment, wording, dark mode, sticky header, and icon placement inconsistencies across multiple pages.
• Fixed queue-status (cs) frames being sent to non-browser clients; API and other programmatic consumers no longer receive them.

Memory Blocks​

• Fixed memory injection incorrectly running for search agents; memory amendments are now placed in the system prompt instead.

Authentication and permissions​

• Fixed cross-user access to Chat Insights message details, which now correctly returns a 403.
• Fixed get_global_configuration exposing non-public keys.
• Fixed the New API key button not working on the Using the API settings page, and restored correct permission gating on that page.
• Fixed API requests preferring an ambient session cookie over an explicit API key.
• Fixed authenticated user information missing from REST API request logs.

Scheduled tasks​

• Fixed recurring scheduled tasks not honoring their configured timezone when updated.
• Fixed monthly day-of-month scheduling not clamping correctly to the end of shorter months.
• Fixed scheduled task dates not converting with an explicit timezone.
• Fixed scheduled tasks leaking database errors when their collection had been deleted.

Documents and connectors​

• Fixed strict OOXML .xlsx files failing to import.
• Fixed collection deletion failing when the same storage object had already been removed.
• Fixed website imports not recovering from broken redirect downgrades.
• Fixed document crawling occasionally stopping after a service interruption.
• Fixed connector actions not surfacing actionable errors, so failed actions now show a clear message instead of doing nothing.
• Fixed file serving for "sidecar original" files, and cleaned up associated storage on delete.

Vector database (VEX)​

• Fixed pgvector dimension errors during migration and improved related diagnostics.
• Fixed collection migrations to pgvector occasionally stalling after a restart.
• Fixed new collections not always recording their VEX backend.
• Fixed per-collection encryption settings not being honored in SQLite/HNSW paths.
• Fixed index deletion failing when no vector index existed.
• Fixed the Vex full-text search schema migration running without the per-collection write lock, preventing concurrent-access issues during upgrades.

Models​

• Fixed intermittent "Failed to list models" errors.
• Fixed agent RAG code generation failing to properly escape user queries.
• Fixed regressions affecting model and tool flows.

Database​

• Fixed unstable collection list ordering.
• Fixed an error that could occur when a system setting held an unexpected value type.
• Fixed mux not failing fast when required database roles are absent, so misconfiguration now surfaces immediately instead of continuing silently.

Budgets​

• Fixed inconsistent handling of "unlimited" budgets, standardizing on -1 and resolving related edge cases, including NULL budgets.

Security​

• Addressed several security vulnerabilities identified through dependency scanning.

h2oGPTe v1.7.1​

New features​

Models​

• New model added: Support for Google Gemma-4-31B-IT (via OpenRouter), a multimodal model with a 256 K context window and optional reasoning mode.

Chat​

• Chat name column: The chat list within a collection now displays the chat name as a dedicated column for easier browsing.
• Sample questions on welcome screen: Default sample questions appear below the chat input on the welcome screen to help you get started.
• Improved references panel: References in chat responses are now displayed inline and in a dedicated side panel for easier review.

Documents and collections​

• Document download: Download individual documents directly from the document list. The actions column is now pinned and stays visible as you scroll.
• Multilingual PII detection: PII detection now identifies personally identifiable information across multiple languages, extending privacy protection beyond English-only content.

Memory Blocks​

• Auto memory blocks: Memory blocks are now created automatically for chats, so context persists across sessions without any manual setup.

Rate limiting​

• Rolling 30-day quota: Usage limits now reset on a rolling 30-day window rather than at the start of each calendar month, ensuring consistent enforcement throughout the month.
• Rate-limit notifications in chat: When a chat session is rate-limited, the notification is now saved in the chat history so you can see exactly where throttling occurred.

Scheduled tasks​

• Run Now on paused tasks: You can now manually trigger a paused scheduled task at any time. Pausing only stops the automatic schedule; manual runs are always available.
• Collection sync across schedule dialog and sidebar: Changing the collection in the schedule dialog or the chat sidebar now keeps both views in sync automatically.

Integrations​

• Excel Add-in: A new Microsoft Excel add-in brings AI-powered spreadsheet functionality directly into Excel. Chat with your data, generate formulas, and run h2oGPTe queries without leaving your spreadsheet.
• Google Sheets Add-on: A new Google Sheets add-on adds AI-powered chat capabilities to Google Sheets. Ask questions about your data and get AI-generated insights inline, directly within your existing Google Workspace workflow.

Authentication and access control​

• Role-based access allow list: Administrators can restrict platform access to users in a configured set of roles. Users outside the allow list are denied access at sign-in.
• Role-based access revocation: Administrators can revoke access for entire role groups, complementing the existing per-user revocation controls.
• Public chat sharing: Shared chat links can now be opened by recipients without requiring them to sign in.
• Share link in dialog: The share link for a chat is now shown in the same dialog, without navigating away.
• Live Logs enhancements: The Live Logs admin view now supports filtering by log level and downloading log output directly from the UI, making it easier to triage and share diagnostic information.

Logging​

• Sensitive data redaction in logs: Administrators can now configure runtime logs to redact PII and user-supplied content across all services. When redaction is enabled, the following fields are redacted: usernames, email addresses, OIDC subject claims, group and role memberships, document names, file paths, bucket and object names, and source URLs. Non-sensitive identifiers (user ID, guest status, and admin status) are always logged regardless of this setting. Redaction is disabled by default to preserve historical logging behavior.
• IP address logging control: Administrators can now enable or disable client IP address logging independently of the broader log redaction setting. IP logging is enabled by default.
• Custom log substring blocking: Operators can define a comma-separated list of substrings to block from log output. Any log message containing a listed substring is replaced with [redacted]. This applies to all Python services.

Improvements​

Scheduled tasks​

• Prompt required: Scheduled tasks now require a prompt to be set before saving, so tasks cannot be created without input.
• Timezone and collection labels: Timezone and collection fields in the schedule dialog are now fully localized.

User interface​

• Cancel reason guidance: The cancel reason field now shows a character limit and placeholder text to help you provide a useful reason.

Python client​

• Adds cancel_reason support when cancelling jobs.
• Improves forward compatibility when connecting to a newer server version.

Bug fixes​

Scheduled tasks​

• Fixed paused scheduled tasks accumulating stale execution state across multiple run cycles.
• Fixed the /schedule slash command not working after a recent update.
• Fixed scheduled task timestamps not updating correctly when a task was paused mid-execution.
• Fixed completed and expired scheduled tasks incorrectly accepting manual trigger requests.
• Fixed the "Next run" column showing a timestamp for paused and completed tasks.

Chat and UI​

• Fixed the chat side panel remaining open and empty after navigating between collection chats.
• Fixed the prompt template not reflecting the collection's setting when the session had no override set.
• Fixed the memory block selection not saving when cleared at the collection level.
• Fixed the collection edit dialog failing when the selected collection had been deleted.

Authentication​

• Fixed file downloads failing for certain authentication methods.

AI and processing​

• Fixed agent mode failing to activate correctly in certain configurations.