All Settings

This chapter explains, area by area, the system environment variables managed on the Admin Center → Environment → All Settings screen. This screen is an advanced view that lets you review and edit every environment variable the solution uses from a single page.

Use the dedicated menus first

For day-to-day configuration, the dedicated menus — LLM, Search / Embedding, Guardrail — are easier to use. The All Settings screen is for when you need to review or edit the same values precisely at the key-value level.

Screen Layout

Select Admin Center → Environment → All Settings in the left menu to open the screen. A search box at the top ("Search by variable name, path, value...") lets you filter instantly by key name, dotted path, or value.

Each setting is listed as a card showing the following.

Element	Meaning
Variable name	Upper-case key (e.g., EMBEDDING_PROVIDER)
Path	Dotted configuration path (e.g., embedding.provider)
Type badge	Value format — Str / Num / Bool / Json / Enum (fixed choices) / Array
Set / Default badge	Set means it has been changed from the factory default, Default means it is unchanged
Current value / Default	The currently applied value alongside the factory default

Some items also show a description of what the setting does and the available choices. Sensitive values such as passwords and API keys are masked as ********.

Confirm the blast radius before changing

All Settings exposes keys that directly affect how the solution behaves — LLM, embedding, guardrail, workflow, and more. A wrong value can immediately affect in-progress chats, agent runs, and knowledge search. Review changes with your operations team beforehand, and since the same key can also be edited from a dedicated menu, record which screen you changed it on.

Areas at a Glance

The category tabs may vary by environment, but they generally cover the following areas. Most are easier to configure from a dedicated menu, noted alongside.

Area	What it configures	Dedicated menu / chapter
Embedding / Search	The embedding model and reranker that turn documents and queries into vectors	Embedding & Vector Search
Vector DB	Connection to the vector database that stores and searches embedding vectors	Embedding & Vector Search
LLM providers	The language models used for chat and inference (OpenAI / Anthropic / Gemini / vLLM, etc.)	LLM Settings
Guardrail	Harmful-content, forbidden-word, and personal-information (PII) filters	Guardrail Model, PII Policy
Vision-language	Models that understand images and documents (OCR)	—
Document processing	Extracting text from images inside uploaded documents	—
Audio (STT / TTS)	Speech-to-text and text-to-speech	—
Application	Operating policies: deployment approval, sharing, sessions, security	(see Application below)
Approval integration	Integration with an external e-approval system	—
Knowledge retention (TTL)	Automatic expiry period for knowledge	—
Notification	Slack, email, and system alerts	—
GPU infrastructure	On-demand GPU rental and model serving	—
Training / experiments	Training server and experiment tracking	—

Each area's key settings are explained below. Typing a variable name into the search box jumps straight to that item.

Embedding / Search

This area turns documents and queries into vectors for knowledge search (RAG). For registration steps, see Embedding & Vector Search.

Variable	Meaning
EMBEDDING_PROVIDER	Embedding provider — an external API (e.g., OpenAI) or a self-hosted server (custom).
CUSTOM_EMBEDDING_URL / CUSTOM_EMBEDDING_MODEL_NAME	Address and model identifier of a self-hosted embedding server.
OPENAI_EMBEDDING_MODEL_NAME, VOYAGE_*	Model, key, and URL when using an external provider such as OpenAI or Voyage.
AUTO_DETECT_EMBEDDING_DIM	Whether to auto-detect the vector dimension the model outputs.
EMBEDDING_RERANKER_PROVIDER / EMBEDDING_RERANKER_MODEL	Method and model of the reranker that reorders first-pass search results.

Embedding dimension must match the vector DB dimension

If the vector dimension produced by the embedding model differs from the vector DB's vector_dimension, ingestion and search fail. When changing the embedding model, check the dimension too, and re-embed existing collections if needed.

Vector DB

Connection details for the vector database that stores vectors and performs similarity search.

Variable	Meaning
QDRANT_HOST / QDRANT_PORT	Vector database host and port.
QDRANT_VECTOR_DIMENSION	Dimension of stored vectors. Must match the embedding model's output dimension.
QDRANT_USE_GRPC / QDRANT_GRPC_PORT	Whether to use gRPC, and its port.
QDRANT_API_KEY	Access key when authentication is enabled.

LLM Providers

Settings for the language models used in chat and inference. Each provider repeats the same set of items — API key / default model / base URL / temperature / max tokens / request timeout. For registration, see LLM Settings.

Variable	Meaning
DEFAULT_LLM_PROVIDER	The LLM provider used by default (OpenAI / Anthropic / Gemini / vLLM, etc.).
LLM_AUTO_FALLBACK	Whether to switch automatically to another provider if the default call fails.
LLM_CONNECTION_TIMEOUT / LLM_MAX_RETRIES	Connection wait time (seconds) and retry count.
OPENAI_* / ANTHROPIC_* / GEMINI_* / VLLM_*	Each provider's API key, default model, base URL, temperature, max tokens, and timeout.

temperature controls how varied responses are (closer to 0 = more consistent, closer to 1 = more diverse); max_tokens caps the length of a single generated response.

Guardrail / Safety

This area inspects the risk of LLM inputs and outputs. For details, see Guardrail Model and PII Policy.

Variable	Meaning
GUARDER_PROVIDER	Guard model method (none / external / self-hosted, etc.).
GUARDER_API_BASE_URL / GUARDER_MODEL_NAME	Guard model server address and model name.
IS_AVAILABLE_GUARDER	Whether harmful-content guarding is enabled.
IS_AVAILABLE_PIIS	Whether personal-information (PII) detection/masking is enabled.
IS_AVAILABLE_FORBIDDEN_WORDS	Whether the forbidden-word filter is enabled.
GUARDER_FAIL_OPEN	Behavior when the guard model fails. true keeps serving despite a failure; false blocks.
GUARDER_RIGOROUS	Strict mode — judges more conservatively.

Guard connection and enablement are separate

Even when a guard model server (GUARDER_*) is configured, no checks run if IS_AVAILABLE_GUARDER, IS_AVAILABLE_PIIS, and IS_AVAILABLE_FORBIDDEN_WORDS are all off. To enforce control policies, enable both the model connection and each toggle.

Vision-Language Models

Settings for models that understand images and scanned documents, or extract text via OCR.

Variable	Meaning
VISION_LANGUAGE_MODEL_PROVIDER	The vision provider to use (none / OpenAI / Anthropic / Gemini / vLLM / AWS, etc.).
VISION_LANGUAGE_<provider>_*	Each provider's API key, base URL, model name, temperature, image quality, and batch size.
VISION_LANGUAGE_VLLM_MODEL_TYPE	The type of self-served (vLLM) model; the type auto-applies sampling defaults and OCR prompts.
VISION_LANGUAGE_VLLM_OCR_INSTRUCTION	The OCR instruction prompt sent to the vision model (uses the default when blank).

Document Processing

Settings for extracting text from images inside uploaded documents. Choose the model provider with DOCUMENT_PROCESSOR_IMAGE_TEXT_MODEL_PROVIDER, then set the provider's base URL, key, and model name. Leave it as no_model if unused.

Audio (STT / TTS)

Variable	Meaning
IS_AVAILABLE_STT / STT_PROVIDER	Whether speech-to-text (STT) is enabled, and its provider.
IS_AVAILABLE_TTS / TTS_PROVIDER	Whether text-to-speech (TTS) is enabled, and its provider.

Each provider also offers model-name and device (cpu/gpu) items. Turn the toggles off in environments that do not use audio.

Application

This area sets operating policies — deployment approval, sharing, sessions, security — and directly affects governance and security. Take particular care when changing it.

Variable	Meaning
DEPLOYMENT_MODE	Approval procedure for external agent deployment: immediate / admin one-step / admin + governance two-step. The two-step flow is covered in Agent Operations and AI Governance.
REQUIRE_RE_DEPLOY_ACCEPT	Whether editing/saving a workflow invalidates the existing deployment approval and requires re-approval.
SHARE_POLICY	How sharing relates to deployment: independent, or sharing takes effect only once deployed.
SHARE_PERMISSION_MODE	Scope of share permissions: read/run only, or read-write allowed per role.
SHARE_ON_UNDEPLOY	How existing shares are handled when deployment is removed: suspend / revoke / retain.
AGENT_DEV_PLAN_REQUIRED	Whether selecting an Agent development plan is mandatory at deployment (request) time.
ADMIN_IP_ACL	IP whitelist allowed for super-admin login and requests. No IP check is performed when empty.
ACCESS_TOKEN_EXPIRE_MIN	Maximum session lifetime (minutes) after which the session auto-expires.
INACTIVITY_TIMEOUT_MIN	Inactivity timeout (minutes) that force-expires a session after no activity. 0 disables it.
REFRESH_TOKEN_EXPIRE_DAYS	Period (days) during which auto re-login is possible; once expired, the user must log in again.
WORKFLOW_MAX_WORKERS	Cap on concurrent workflow executions.
SIDEBAR_CONFIG	Sidebar menu composition (JSON), usually managed via Environment → Sidebar.
PATHFINDER_DOWNLOAD_URL	Download link opened from the Pathfinder banner on the Agent design start screen.

Approval Integration

Settings to integrate agent deployment approval with an external e-approval system.

Variable	Meaning
APPROVAL_ENABLED	Whether external approval integration is used. When off, only the solution's internal approval works.
APPROVAL_HOST	Approval API base URL.
APPROVAL_CLIENT_ID / APPROVAL_CLIENT_SECRET	Pre-issued integration credentials (the secret is sensitive).
APPROVAL_CALLBACK_BASE_URL	Base URL that receives approval results.
APPROVAL_ADMIN_APPROVERS	Candidate email list for first-step approvers (platform admins).
APPROVAL_GOVERNANCE_APPROVERS	Candidate email list for second-step approvers (governance), used in two-step mode.

Knowledge Retention (TTL)

Automatic-expiry policy for uploaded knowledge.

Variable	Meaning
KNOWLEDGE_TTL_MAX_DAYS	Maximum number of days that can be set as a retention period.
KNOWLEDGE_TTL_DEFAULT_DAYS	Default retention days for new knowledge.
KNOWLEDGE_TTL_GRACE_PERIOD_DAYS	Grace period (days) between expiry and actual deletion.

Notification

Settings to notify external channels of system events and alerts.

Variable	Meaning
NOTIFICATION_ENABLED	Master switch for sending notifications.
SLACK_WEBHOOK_URL	Slack notification destination.
EMAIL_SMTP_HOST / EMAIL_SMTP_PORT / EMAIL_USERNAME	SMTP server details for email sending.
CPU_ALERT_THRESHOLD / MEMORY_ALERT_THRESHOLD	Resource-usage alert thresholds (%).
NOTIFICATION_* (unified messaging)	Integration details for an in-house unified messaging system (host, client, interface identifiers, etc.).

GPU Infrastructure · Model Serving

Automation for renting external GPUs to serve models, plus connections to additional inference servers. These are configured by infrastructure staff during installation and operations.

Variable	Meaning
VAST_*	On-demand GPU rental (VAST) settings — container image to use, instance selection criteria such as price cap, disk, and GPU RAM, plus auto-destroy and timeout.
VLLM_* (vast)	Options for the vLLM server launched on the rented GPU — served model, max context length, GPU memory utilization, parallelism.
SGL_*	SGLang inference server connection.
SESSION_STATION_*	Connection and retention time for the session/conversation-state store.
WORKFLOW_EXECUTION_TIMEOUT	Time limit (seconds) for a single workflow execution.

Training · Experiments · External Integrations

Variable	Meaning
TRAINER_HOST / TRAINER_PORT	Connection details for the model training service.
MLFLOW_*	Experiment/model tracking (MLflow) connection — tracking URL, default experiment, cache location.
HF_TOKEN / IS_AVAILABLE_HUGGINGFACE / HUGGINGFACE_*	HuggingFace model hub integration — token, enablement, API, cache, offline mode.
AWS_*	Credentials, region, and endpoint for calling AWS Bedrock.

Operating Recommendations

• Prefer the dedicated menus — LLM, embedding, guardrail, and others provide input validation and connection tests. Use All Settings for precise review and exceptional edits.
• Back up and record before changing — note the current value before changing it, and record the reason and the screen used. All changes are written to the audit log.
• Handle sensitive values carefully — deliver API keys, secrets, and tokens only through a secure channel, and avoid exposing them during screen sharing or captures.
• Check the enable toggles — guardrail, audio, notification, and others only work once their enablement flags (IS_AVAILABLE_*, *_ENABLED) are turned on, even if the model/connection is set.

Contact

For questions about All Settings, please contact the Xgen Solution Administrator.