Enterprise Configuration Specification

Wizard derives its runtime behavior from environment variables, backend/.env, and local credential stores. This reference specifies all available configuration keys, data types, defaults, and architectural implications.

1. Configuration Resolution Precedence

Settings resolve with deterministic precedence (highest to lowest):

System Architecture & Flow

Verified Spec

┌────────────────────────────────────────────────────────┐
│ 1. Process Environment Variables (export KEY=value)    │  (Highest)
├────────────────────────────────────────────────────────┤
│ 2. Environment File (`backend/.env`)                   │
├────────────────────────────────────────────────────────┤
│ 3. Local Secure Credential Store (`credentials.json`)  │
├────────────────────────────────────────────────────────┤
│ 4. Hardware Auto-Sizing Defaults                       │  (Base)
└────────────────────────────────────────────────────────┘

2. Tri-Model Architecture Configuration

Variable	Type	Default	Description
`API_PROVIDER`	`string`	`ollama`	Default inference provider: `ollama`, `lmstudio`, `gemini`, `anthropic`, `openai`, or `custom_gateway`.
`DATA_MODE`	`string`	`local-only`	Session privacy mode: `local-only` (100% air-gapped), `hybrid` (redacted), or `cloud-only`.
`MODEL_NAME`	`string`	`""`	Pins the Manager reasoning model (e.g. `qwen3:8b`, `gemini-2.5-flash`). Empty = auto-select.
`WORKER_MODEL_NAME`	`string`	`""`	Pins the Worker coding model (e.g. `qwen2.5-coder:7b`).
`EMBEDDING_PROVIDER`	`string`	`""`	Pins vector embedding provider: `ollama`, `lmstudio`, `openai`, `gemini`, `custom_gateway`, `none`. Empty follows `API_PROVIDER`.
`EMBEDDING_REMOTE_MODEL`	`string`	`""`	Pins remote embedding model id (e.g. `nomic-embed-text`, `text-embedding-3-small`). Empty = auto-discover.
`EMBEDDINGS_REMOTE_ENABLED`	`bool`	`true`	When true, attempts provider or local sentence-transformers before falling back to Blake2b.
`EMBEDDINGS_FORCE_FALLBACK`	`bool`	`false`	When true, forces deterministic 384-dimensional Blake2b lexical hashing encoder.
`RAG_ENABLED`	`bool`	`false`	Enables semantic document retrieval across uploaded PDF/DOCX attachments.
`VISION_MODEL_NAME`	`string`	`""`	Model used for visual chart inspection and aesthetic critique (e.g. `llama3.2-vision`).
`TEMPERATURE`	`float`	`0.0`	Sampling temperature for analytical determinism.
`MAX_TOKENS`	`int`	`4096`	Upper token generation limit per completion turn.
`LLM_KEEP_ALIVE`	`string`	`30m`	How long local inference engines (Ollama) keep weights loaded in VRAM between manager/worker turns.
`LLM_NUM_THREAD`	`int`	`0 (auto)`	CPU inference threads. Auto-derived from physical core count to eliminate memory bus contention.
`LLM_NUM_CTX`	`int`	`0 (auto)`	KV cache context window requested from Ollama (e.g. `8192` or `16384`).
`LLM_REQUEST_TIMEOUT`	`int`	`300`	HTTP client timeout in seconds for long-context inferences.

3. Cloud Provider & Gateway Authentication

Variable	Type	Default	Description
`GEMINI_API_KEY`	`string`	`""`	Google Gemini API key. Enables Gemini 2.5 Flash / Pro and Google Embeddings.
`ANTHROPIC_API_KEY`	`string`	`""`	Anthropic Claude API key for Claude 3.5 Sonnet / Haiku.
`OPENAI_API_KEY`	`string`	`""`	OpenAI platform API key for GPT-4o and OpenAI text embeddings.
`GATEWAY_API_URL`	`string`	`""`	Base URL for custom OpenAI-compatible gateways (Groq, Together AI, OpenRouter, vLLM).
`GATEWAY_API_KEY`	`string`	`""`	Bearer authorization token for the custom gateway endpoint.
`OLLAMA_BASE_URL`	`string`	`http://127.0.0.1:11434`	Local Ollama daemon REST endpoint.
`LMSTUDIO_BASE_URL`	`string`	`http://127.0.0.1:1234/v1`	Local LM Studio OpenAI-compatible endpoint.

4. Sandboxing & Runtime Execution Controls

Variable	Type	Default	Description
`EXECUTION_BACKEND`	`string`	`host`	Where Python code executes: `host` (subprocess), `docker` (container), or `inprocess` (CI only).
`HOST_SANDBOX`	`string`	`best-effort`	OS kernel containment policy: `off`, `best-effort`, or `require` (refuses to start uncontained).
`HOST_SANDBOX_NETWORK`	`string`	`deny`	Outbound network egress for generated code: `deny` or `allow`. Loopback always permitted.
`SANDBOX_TIER`	`string`	`standard`	Pre-installed data science image toolkit: `core` (pandas/duckdb), `standard` (+scikit-learn/scipy), `full` (+geopandas/lifelines).
`SANDBOX_MEM_LIMIT`	`string`	`2g`	Cgroup memory ceiling for sandbox containers.
`SANDBOX_CPU_QUOTA`	`int`	`0`	CPU quota in microseconds (`100000` = 1 core, `0` = unrestricted).
`SANDBOX_PIDS_LIMIT`	`int`	`256`	Maximum simultaneous child processes per sandbox.
`SANDBOX_EXEC_TIMEOUT`	`int`	`180`	Wall-clock execution timeout in seconds before code is forcefully interrupted.

5. Agentic Loop, DAG & Grounding Parameters

Variable	Type	Default	Description
`AGENT_TIER`	`string`	`auto`	Reasoning budget depth: `auto` (derived from model size), `compact` (4 steps), `balanced` (12 steps), `full` (24 steps).
`AGENT_MAX_ITERATIONS`	`int`	`24`	Hard ceiling for agent turn iterations to prevent runaway gateway billing.
`AGENT_REQUIRE_APPROVAL`	`bool`	`false`	When true, pauses execution after the planning stage for explicit user plan sign-off.
`AGENT_PERMISSION_PROFILE`	`string`	`ask-always`	Policy for external actions: `auto-approve`, `ask-always`, or `custom`.
`AGENT_VERIFY`	`bool`	`true`	When true, independently recomputes headline analytical figures via alternative routes.
`AGENT_GROUNDING_CHECK`	`bool`	`true`	Validates narrative claims and Markdown table cells against execution DataFrames.
`AGENT_TURN_TIMEOUT`	`float`	`300.0`	Total turn wall-clock deadline in seconds.

6. SRE, Observability, Scaling & Queues

Variable	Type	Default	Description
`REDIS_URL`	`string`	`redis://localhost:6379/0`	Connection string for distributed job queue, trajectory caching, and pub/sub event bus.
`QUEUE_BACKEND`	`string`	`inprocess`	Job queue engine: `inprocess` or `redis`.
`CACHE_BACKEND`	`string`	`inprocess`	Semantic trajectory cache engine: `inprocess` or `redis`.
`BACKUP_INTERVAL_HOURS`	`int`	`6`	Automated background SQLite online WAL checkpoint and snapshot backup interval.
`BACKUP_KEEP_COUNT`	`int`	`5`	Maximum number of rotating database backups retained in `backups/`.
`IDEMPOTENCY_TTL_SECONDS`	`int`	`300`	Time window for caching completed turn results by client request hash.
`OTEL_EXPORTER_OTLP_ENDPOINT`	`string`	`""`	Target OTLP gRPC/HTTP endpoint for OpenTelemetry distributed traces.
`OTEL_SERVICE_NAME`	`string`	`wizard`	Service identifier reported in distributed trace spans.
`PLOT_FORMAT`	`string`	`html`	Chart output format: `html` (interactive Plotly DOM) or `png` (static matplotlib).
`CORS_ALLOW_ORIGINS`	`string`	`http://localhost:3000`	Comma-separated allowlist for browser cross-origin requests.
`API_KEY`	`string`	`""`	When set, all mutating control plane routes require the `X-API-Key` HTTP header.
`SESSION_MAX_ACTIVE`	`int`	`8`	Maximum concurrent analytical sessions before queueing.

Reviewed for Wizard v1.0.15Edit this page on GitHub

PreviousInstalling a Skill from GitHub NextAPI & Event Protocol