Skip to content

v1.10.0

Latest

Choose a tag to compare

@elamceritech elamceritech released this 29 Sep 10:44
· 56 commits to main since this release
d20d7c9

v1.10.0

This release makes GenAssist faster, more stable and more secure when many people use it at once, and adds a set of everyday improvements to conversations, cost tracking and evaluations.

Heavy screens such as the dashboard, analytics and conversation lists can now read from a separate read-only database copy, so they no longer slow down live chats. The platform also limits how many agent replies run at once, so a traffic spike returns a clear "busy, try again" instead of making everything slow. Background jobs are much more dependable: jobs are no longer lost, duplicated or stuck after a restart.

On the security side, admins can now restrict which websites may call agent endpoints from a browser, and sensitive tenant settings can only be given to the admin role. New features include read receipts ("Seen") in chat, a new Conversations workspace, optional prompt caching to cut LLM costs, the ability to add your own LLM models per provider, exact date ranges on the dashboard and analytics, and more flexible evaluation rules.

API Allowlisting & Security

These changes control who can call the API and from where, and close a few gaps found in security reviews.

  • Added CORS_AGENT_ALLOWED_ORIGIN_REGEX, an optional allowlist of the websites allowed to call agent endpoints from a browser. When it's set, only matching sites can use the agent from a browser with credentials; any other site is blocked by the browser.
  • The match is on the full origin (scheme://host[:port]), so a look-alike domain such as https://app.example.com.attacker.com can't slip through a rule for https://app.example.com.
  • A blocked site's preflight request (OPTIONS) returns 400. When the setting is empty (the default), nothing changes. A mistyped pattern is logged and ignored instead of breaking the app. Example: https://([a-z0-9-]+\.)*example\.com.
  • Action for gateway / WAF owners: add the new routes listed under New Endpoints (mark-read, llm-model-catalog, operators/list, ml-models/profile-csv) to your allowlist. The chat widget on customer websites calls mark-read directly, so it must also be allowed on the public ingress.
  • Permissions for tenant-wide settings (App Settings such as File Manager and Security, and Feature Flags) can now only be given to the admin role. Trying to give them to any other role returns 403 ADMIN_ONLY_PERMISSION.
  • A user who loses the supervisor role immediately loses access to the groups they used to supervise. Supervisors now also see their own group's data.
  • The FAISS knowledge-base cache is now loaded with a safe loader that only accepts plain data, so a tampered cache file can't run code.
  • Updated frontend and plugin libraries (nanoid, postcss, axios, js-yaml, brace-expansion) to fix high-severity vulnerabilities.

Performance & Scalability

GenAssist now holds up better under heavy load and recovers cleanly from traffic spikes.

  • Added optional Aurora read-replica support (DB_READ_HOST). Dashboard, analytics, LLM usage and conversation-list screens read from the replica, so they don't compete with live chats for database connections.
  • Right after you save something, your next reads still go to the main database for a few seconds, so you always see your own change even if the replica is slightly behind. If the replica has problems, all reads automatically go back to the main database until it recovers.
  • Added a limit on how many agent replies one server process handles at the same time (CHAT_TURN_MAX_INFLIGHT, default 24). Extra requests wait briefly in a queue. If one waits longer than CHAT_TURN_QUEUE_TIMEOUT_SECONDS (default 10), the caller gets 503 CHAT_TURN_CAPACITY_EXCEEDED ("The assistant is busy right now") instead of a slow timeout.
  • If a user closes the page while their request is still waiting, it is dropped instead of holding a place in the queue.
  • Added a dedicated thread pool for model calls (LLM_EXECUTOR_THREADS), and a time limit on graceful shutdown (GRACEFUL_SHUTDOWN_TIMEOUT_SECONDS) so deployments and pod restarts finish cleanly.
  • Database saves are now committed or rolled back in one central place, so a failed operation never leaves half-saved data behind.
  • The operators menu is now paginated, and the Agent Performance screen loads faster.

Celery Reliability

Background jobs (scheduled workflows, evaluations, ML training) are now much more dependable.

  • A job is only marked done after it actually finishes, and a job that is delivered twice can't be picked up twice.
  • Jobs whose worker crashed are no longer silently requeued. A run whose worker was lost is marked as failed instead of being run a second time.
  • Only one scheduler runs at a time (leader lock), and a scheduler that freezes is detected and replaced automatically.
  • Redis connections now have timeouts and keepalive, so a dead connection can no longer freeze a worker.
  • Evaluation failures are now saved, so you can see why a run failed.

Prompt Caching & LLM Cost

You can now lower LLM costs and track them more accurately.

  • Added optional prompt caching (promptCaching) on LLM, Agent and Sub-Agent nodes. When it's on, the fixed start of the system prompt is cached for 5 minutes, and repeat calls are billed at a lower rate. Works with Anthropic and Bedrock models that support caching.
  • Cost reports now count cached tokens separately (cache read and cache creation) and price them with their own per-model rates (cache_read_per_1k, cache_creation_per_1k).
  • Added an LLM model catalog: tenants can add models that aren't built in (for example a newly released model) and pick them in provider settings without waiting for a release.

Conversations & Chat

Conversations are easier to follow and manage in real time.

  • Added read receipts: the visitor sees "Seen" when a staff member has read their message, and the console sees when the visitor has read the reply. Updates arrive live over WebSocket.
  • Added a Conversations workspace with split and list views, and conversations update in real time without a page refresh.
  • You can import several conversations at once, add a conversation to a dataset in one click, and use the redesigned Datasets page.

Evaluations

Evaluation rules can now describe real conversation behavior more accurately.

  • Route Taken and Action Taken rules can now be checked on every turn, once for the whole conversation, or only on specific turns you choose. For example, "a ticket is created once per conversation" can now be expressed correctly.
  • Each rule result now shows which kind of check produced it (tool_used, route_taken or action_taken).
  • Evaluation workflow lists now show the agent's name, and only include workflows whose agent you're allowed to see.

Integrations & Platform

  • Added Zendesk OAuth client-credentials login (ZENDESK_AUTH_METHOD, ZENDESK_CLIENT_ID, ZENDESK_CLIENT_SECRET). Zendesk removes API tokens on 2027-04-30, so new connections should use OAuth.
  • Gmail and Office365 calls no longer block the server while they wait for a response.
  • Train Data Source can now preview and profile a CSV file, and Train Model supports a time-based train/test split.
  • Audio providers, fallback chains and Ask GenAI can now be turned on or off with feature flags.
  • Compare versions can ignore IDs and node names to show only meaningful changes. Workflow node config fields have a variable picker for inserting variables inline.

New Endpoints

Conversations (/api/conversations)

Method Path Params Example
PATCH /api/conversations/in-progress/mark-read/{conversation_id} conversation_id (path, UUID); Body: last_read_sequence (int ≥0, required). Reader role (customer / supervisor) is derived from the auth principal, never from the body. Returns ConversationReadReceiptState PATCH /api/conversations/in-progress/mark-read/3fa85f64-…

LLM Model Catalog (/api/llm-model-catalog)

Method Path Params Example
GET /api/llm-model-catalog — GET /api/llm-model-catalog
GET /api/llm-model-catalog/providers — (extendable provider types with their builtin_model_keys) GET /api/llm-model-catalog/providers
POST /api/llm-model-catalog Body: provider_key (str 1–64, required), model_key (str 1–512, required), label (str 1–255, required), is_active (int 0/1, default 1). Returns 201 POST /api/llm-model-catalog
PUT /api/llm-model-catalog/{entry_id} entry_id (path, UUID); Body: label (str, optional), is_active (int 0/1, optional). Identity is fixed — delete + recreate to move an entry; 404 if not found PUT /api/llm-model-catalog/3fa85f64-…
DELETE /api/llm-model-catalog/{entry_id} entry_id (path, UUID). Returns 204; 404 if not found DELETE /api/llm-model-catalog/3fa85f64-…

Operators (/api/operators)

Method Path Params Example
GET /api/operators/list skip (query, int ≥0, default 0), limit (query, int 1–100, default 20), search (query, str, optional — first / last name). Paginated, ordered by positive sentiment, then score GET /api/operators/list?skip=0&limit=20&search=ann

ML Models (/api/ml-models)

Method Path Params Example
POST /api/ml-models/profile-csv Body: file_url (str, required — path under DATA_VOLUME or HTTP(S) URL). Returns a downloadable HTML profiling report; 404 if the file is not found POST /api/ml-models/profile-csv

Updated Endpoints

Method Path Change
GET /api/conversations/{conversation_id} Response now includes read_state (customer_last_read_sequence, customer_last_read_at, supervisor_last_read_sequence, supervisor_last_read_at, supervisor_user_id)
GET /api/dashboard/summary days is now optional (defaults to 30). Added from_datetime / to_datetime (exact range; both required, with timezone offset) and all_time (bool); mixing modes returns 422. Response adds conversations; workflow_runs is deprecated (alias of conversations)
GET /api/analytics/agents/summary Added activity_from_datetime / activity_to_datetime (exact activity interval; both required, with timezone offset, need from_date + to_date, not combinable with compare); invalid combinations return 422
GET /api/analytics/agents/export Added activity_from_datetime / activity_to_datetime with the same validation
GET /api/analytics/llm-usage/summary Response adds total_cache_read_tokens and total_cache_creation_tokens
POST /api/llm-cost-rates Body now accepts cache_read_per_1k and cache_creation_per_1k (decimal, optional)
PUT /api/llm-cost-rates/{rate_id} Body now accepts cache_read_per_1k / cache_creation_per_1k. Sent blank, a cache rate resets to the provider default; omitted, it is kept
GET /api/llm-cost-rates/export Added include_cache_rates (query, bool, default false); false keeps the 4-column CSV layout
POST /api/llm-cost-rates/import CSV upload is now capped at 1 MB (400 above the cap)
GET /api/llm-providers LlmProviderRead now includes prompt_caching_mode (explicit, automatic, none)
GET /api/genagent/workflow/minimal Returns only workflows whose agent the caller can see; response adds agent_name (null when the agent is gone or retired)
GET /api/operators/{operator_id} No longer enriches the operator with its latest call
PATCH /api/llm-analyst/{llm_analyst_id} Inactive LLM analysts can now be updated; reading an inactive analyst by id returns 409 LLM_ANALYST_INACTIVE
PUT /api/user/{user_id} role_ids may be empty for console users; other login-capable users, including on a user_type_id change, still need a role (400 USER_ROLES_REQUIRED)
POST /api/role-permissions Returns 403 ADMIN_ONLY_PERMISSION when assigning an App Settings / Feature Flag permission to a non-admin role
PATCH /api/role-permissions/{rp_id} Same admin-only guard, applied to the resulting role / permission pair
POST /api/internal/agents/execute Busy (CHAT_TURN_CAPACITY_EXCEEDED) and abandoned (CHAT_TURN_CLIENT_DISCONNECTED) turns now keep their HTTP status instead of returning { success: false }

Full Changelog: v1.9.5...v1.10.0