Environment Variables¶
All Geodesia G-1 services can be configured via environment variables. Variables always take precedence over config.yaml values.
Product Backend¶
Set these variables when starting the product backend (python -m uvicorn main:app ...).
Core¶
| Variable | Default | Description |
|---|---|---|
MODEL_HOST_PATH | — | Absolute path to the model checkpoint directory. If unset, the evaluate endpoint is disabled and only compliance APIs are available. |
GLAD_DEVICE | "cuda:0" | CUDA device to load the model on. Values: "cuda:0", "cuda:1", "cpu". |
GLAD_CONFIG | "config.yaml" | Path to the YAML configuration file. |
GLAD_DB_PATH | "var/glad.sqlite3" | Path to the SQLite database file. |
GLAD_LOG_LEVEL | "info" | Logging verbosity: "debug", "info", "warning", "error". |
Security¶
| Variable | Description |
|---|---|
GLAD_LICENSE_TOKEN | License token (glc_…). Required for authenticated deployments. |
GLAD_AUDIT_HMAC_KEY | 64-character hex key for signing the audit chain. If unset, a random key is generated (chain cannot be verified across restarts). |
GLAD_SMTP_PASSWORD | SMTP password for email notifications. |
GLAD_WEBHOOK_SECRET | Webhook signing secret for notification webhooks. |
Explainability¶
| Variable | Default | Description |
|---|---|---|
GLAD_PSS_N_SAMPLES | 5 | Default PSS resample count |
GLAD_PSS_TEMPERATURE | 0.7 | Default PSS temperature |
GLAD_PSS_MATCH_MODE | "ngram" | Default PSS alignment algorithm |
Fine-Tuning¶
| Variable | Default | Description |
|---|---|---|
GLAD_FINETUNE_OUTPUT_DIR | "runs/" | Directory where fine-tuned checkpoints are written |
GLAD_MAX_FINETUNE_JOBS | 1 | Maximum concurrent fine-tuning jobs |
Gateway¶
Set these variables when starting the gateway (python -m uvicorn geodesia_gateway:app ...).
Core¶
| Variable | Default | Description |
|---|---|---|
GLAD_API_URL | "http://localhost:8199" | URL of the product backend |
GW_PORT | 8800 | Port the gateway listens on |
GW_HOST | "0.0.0.0" | Bind address for the gateway |
GW_DB_PATH | same as product GLAD_DB_PATH | Path to the SQLite database. By default, the gateway and product backend share one database. Override to use a separate file. |
GW_LOG_LEVEL | "info" | Logging level |
Upstream¶
| Variable | Default | Description |
|---|---|---|
GW_UPSTREAM_URL | — | URL of the upstream LLM backend (vLLM, Ollama, OpenAI, etc.) |
GW_UPSTREAM_TYPE | "vllm" | Upstream type: "vllm", "openai", "ollama", "sglang", "trtllm", "internal" |
GW_UPSTREAM_MODEL | — | Model name or path passed to the upstream (e.g., "mistralai/Mistral-7B-v0.3") |
GW_API_KEY | — | API key forwarded to OpenAI-compatible upstreams |
GW_MAXLEN | 2048 | Maximum input token length sent to upstream. Must match what the upstream was compiled/configured for. |
Enforcement¶
| Variable | Default | Description |
|---|---|---|
GW_BLOCK_INPUT | 0 | Set to 1 to enable input blocking (prompt safety + jailbreak enforcement). 0 = annotate only. |
GW_BLOCK_OUTPUT | 0 | Set to 1 to enable output blocking (answer safety + hallucination enforcement). |
GW_DEFAULT_MODE | "block" | Default enforcement mode: "block" or "passthrough" |
GW_PROMPT_BLOCK_AXES | prompt_safety,jailbreak | Which prompt-region axes may refuse a request. Only primary axes are accepted: additional axes (profanity, out_of_scope) and the routing classifier (prompt_complexity) are ignored with a warning on stderr — promote off-topic refusal per Application instead. See Detection Axes. |
GW_ADDITIONAL_AXES | profanity,out_of_scope,prompt_complexity | Which axes are additional: they travel in the response's geodesia.additional_axes object instead of geodesia.axes, and cannot be promoted to blocking by configuration (the signed certificate marks them tier: "additional"). Set it to "" to make every axis primary. See Response Format. |
GW_ADDITIONAL_INLINE | 0 | 1 also emits the additional axes inside geodesia.axes (they still appear in geodesia.additional_axes) — a bridge for a client that reads only axes. |
GB_EXTRA_AXES | (per checkpoint) | Axes the served head adds on top of the base six, as name:region,… (region ∈ ans\|prompt\|context). The 9-axis head ships profanity:prompt,out_of_scope:prompt,prompt_complexity:prompt. Read by the detector, the gateway, the policy schema and the feedback store from this single source. |
GW_SYSTEM_AS_CONTEXT | 0 | 1 treats system messages as grounding context for halluc_context. Off by default: a system prompt is an instruction, not evidence — it feeds out_of_scope instead. Enable only if your deployment ships its knowledge base inside the system message. |
GW_TOKEN_SAVING | 0 | 1 turns on the token_saving switch at boot: saves upstream tokens without changing what the model reads. Also settable live with POST /v1/glad/gateway/config {"token_saving": true}. See Token & Cost Control. |
Detection Thresholds (Gateway)¶
These override the thresholds stored in the database for the gateway instance.
| Variable | Default | Description |
|---|---|---|
GW_THR_PROMPT_SAFETY | from DB | Prompt safety threshold [0–1] |
GW_THR_ANSWER_SAFETY | from DB | Answer safety threshold [0–1] |
GW_THR_HALLUC | from DB | Hallucination threshold [0–1] |
GW_THR_COMBINED_HALLUC | from DB | Combined hallucination threshold [0–1] |
GW_THR_JAILBREAK | from DB | Jailbreak detection threshold [0–1] |
Prefer per-Application thresholds
These gateway-wide overrides predate G-1 Studio. In a multi-Application deployment set thresholds in the Application's policy instead — they are versioned, audited, hot-reloaded, and tunable against real traffic with Policy Lens.
Causal Explainability (Gateway)¶
Black-box causal attribution over the companion detector. See Causal Explainability.
| Variable | Default | Description |
|---|---|---|
GW_XAI_MAXLEN | 512 | Maximum token length for XAI scoring passes. |
GW_XAI_SRC_CHARS | 2400 | Maximum characters of each region (prompt / context / answer) submitted for attribution. |
GW_XAI_DCA_RHO | 0.90 | Sufficiency coverage a subset must reproduce to be certified. Raising it to force certificates produces false explanations. |
GW_XAI_DCA_FLOOR | 0.01 | Relevance noise floor in probability units — below it a unit is irrelevant. |
GW_XAI_DCA_MINBASE | 0.5 | Minimum axis probability for a score to count as a decision worth explaining. This is the attribution floor, deliberately distinct from the decision threshold. |
GW_ANALYZE_MIN_MAXLEN | 512 | Minimum maxlen when the gateway auto-halves on OOM. |
RAG¶
| Variable | Default | Description |
|---|---|---|
GW_RAG_ENABLED | 0 | Set to 1 to enable the RAG module |
GW_RAG_DB_PATH | "var/rag.lancedb" | Path to the LanceDB vector store |
GW_RAG_EMBEDDING_MODEL | "BAAI/bge-m3" | Embedding model for document ingestion |
GW_RAG_RERANKER_MODEL | "BAAI/bge-reranker-v2-m3" | Reranking model for retrieval |
GW_RAG_TOP_K | 5 | Number of documents to retrieve |
Numeric Solver¶
| Variable | Default | Description |
|---|---|---|
GW_NUMERIC_SOLVER | "disabled" | Enable the numeric faithfulness solver. Values: "disabled", "judge". When "judge", uses an external LLM to verify arithmetic claims. |
GW_NUMERIC_SOLVER_URL | — | URL of the numeric solver LLM endpoint |
GW_NUMERIC_SOLVER_MODEL | — | Model name for the numeric solver |
G-1 Studio (Applications, cost, scale-out)¶
These variables configure the G-1 Studio control plane (Applications, organizations, keys), the FinOps cost engine, and the multi-replica scale-out path. They are additive and backward-compatible: an existing single-upstream deployment surfaces as the default Application with zero behaviour change, so all of these are optional.
Storage backend¶
| Variable | Default | Description |
|---|---|---|
GEODESIA_DB_URL | (the resolved GLAD_DB_PATH) | Selects where all Studio state lives. sqlite:////abs/path/glad.sqlite3 (note the four slashes for an absolute path; sqlite:///rel/path is relative to CWD) — the same artifact then runs from a laptop file to a cloud VM (e.g. SQLite on an EBS volume). postgresql+psycopg://user:pw@host/db points the control-plane + cost stores at Postgres for multi-replica scale-out. The settings.database_path and the gateway _db() both honor this. (src/glad_minimal/storage.py, db_backend.py) |
Split storage on multi-replica
The control-plane and cost stores are backend-agnostic (portable DDL + ON CONFLICT upserts), so they can run on Postgres while the legacy compliance DB stays on SQLite (single-writer, on a shared FS). See the Studio deployment notes for the recommended split.
RBAC (control-plane access)¶
| Variable | Default | Description |
|---|---|---|
GEODESIA_ADMIN_TOKEN | — | Platform-admin token that gates control-plane writes (Applications / orgs / keys / policy / cost). Sent by the caller as the header X-Geodesia-Admin-Key. When unset and no admin API key exists, the control plane is open (anonymous caller is treated as platform_admin) — preserving the local/single-tenant experience. As soon as an admin token or an admin API key is configured, writes require authentication. (src/glad_minimal/apps/rbac.py) |
An admin can alternatively authenticate with a control-plane API key (Authorization: Bearer g1k_… with the admin role). Role hierarchy: viewer < app_editor < org_admin < platform_admin.
Budget alerts (FinOps)¶
| Variable | Default | Description |
|---|---|---|
GEODESIA_ALERT_WEBHOOK | — | Outbound webhook URL for budget alert/block events. When set, each newly-crossed budget event is POSTed here as JSON; the per-Application alert_recipients list rides inside the payload so your own relay can fan it out to email / Slack / PagerDuty. When unset, events are still recorded in the in-app notification_log (no network egress). There is no SMTP in the product and the webhook URL is server-side only — never per-Application, never in the DB. (src/glad_minimal/cost/alerts.py) |
See Cost & Budget for the FinOps engine, budget bands, and the projection chart.
Remote G1-Hummingbird scoring¶
| Variable | Default | Description |
|---|---|---|
GEODESIA_SCORING_MODE | "embedded" | embedded runs the G1-Hummingbird monitor in-process (zero behaviour change). remote swaps it for a thin HTTP client that talks to a separate scoring pool — for horizontal scale-out of detection compute. (src/glad_minimal/scoring/) |
GEODESIA_SCORING_URL | — | Base URL of the remote scoring server (python -m glad_minimal.scoring.server, e.g. http://host:8810). Required when GEODESIA_SCORING_MODE=remote. |
Thinking Levels¶
Per-request depth dial (thinking_level 0–3, 3 = MAX). Off by default (level 0) → the extra capacity is never loaded, zero overhead. See Thinking Levels.
| Variable | Default | Description |
|---|---|---|
GW_GLADH_CKPT | — | Capability pack that unlocks thinking levels 1–2. Unset → those levels silently fall back to level 0. |
GW_GLADH_DEVICE | auto | Device for the level-½ pack. auto → CUDA when visible, else CPU. Runs in its own sub-process, with its own device selection. |
GW_GLADA_CKPT | — | Capability pack that unlocks thinking level 3 (MAX). Unset → a level-3 request is served at the deepest level available. |
GW_GLADA_DEVICE | auto | Device for the level-3 pack. |
GW_FUSION_BANK | runs/glad_bert/fusion_bank_thinking_v3_gladh_v1.json | Calibration artifact that keeps every level on the same probability scale as level 0, so thresholds do not have to be re-tuned per level. Missing/unreadable → same fallback as an unset pack. |
Live Web Search¶
On by default; the chat searches the live web, screens every page through the G1-Hummingbird firewall, and grounds the answer in the safe pages. A Tavily API key gives reliable, rate-limit-free results; without one, the key-less DuckDuckGo engine is used as a fallback. See Live Web Search.
| Variable | Default | Description |
|---|---|---|
GW_WEBSEARCH_ENABLED | 1 | Master switch. 1 = available; 0 → the per-request web_search flag is a no-op. |
GW_WEBSEARCH_API_KEY | — | Tavily API key. Takes precedence over a UI/file-set key. TAVILY_API_KEY is an accepted alias. |
GW_WEBSEARCH_KEY_FILE | /app/var/websearch_tavily.key | Path to the out-of-band key file the Studio UI writes (mode 0600, never baked into the image). |
GW_WEBSEARCH_PROVIDER | auto | Force tavily or duckduckgo. Default: tavily if a key is present, else duckduckgo. |
GW_WEBSEARCH_MAX_RESULTS | 5 | Number of search results requested. |
GW_WEBSEARCH_MAX_READ | 6 | Maximum number of safe pages used as grounding context. |
GW_WEBSEARCH_TIMEOUT | 12 | Per-request HTTP timeout (seconds) for the search + page fetches. |
GW_WEBSEARCH_SCREEN_CHARS | 4000 | Chars of each page screened by the firewall. |
GW_WEBSEARCH_GROUND_CHARS | 2200 | Chars of each safe page kept as grounding context. |
GW_WEBSEARCH_RJ_THR | calibrated | Override the rag_jailbreak firewall threshold for page screening. Unset → per-axis calibrated default. |
Human Feedback Loop¶
Opt-in episodic exemplar bank built from approved chat feedback. Off by default → detection is byte-identical. See Self-Evolving Security.
| Variable | Default | Description |
|---|---|---|
GW_FEEDBACK_BANK | off | Master switch. on builds + consults the per-model exemplar bank from approved feedback; off → never built, zero overhead. |
GW_BANK_V2 | off | Use the Contrastive Safety Memory bank (dangerous/benign twins). Requires GW_FEEDBACK_BANK=on. |
GW_FEEDBACK_BANK_TAU | 0.88 | Cosine-similarity floor below which an exemplar is ignored (exact-pattern recall). |
GW_FEEDBACK_BANK_GAIN | 1.0 | How hard a perfect match pushes the probability (1.0 = fully to 0/1 at sim == 1, weight == 1). |
GW_FEEDBACK_AUTOAPPROVE | off | on ⇒ a flag whose plain-language problem resolves to an axis enters the engine with no curator. other / unresolvable flags still queue. |
GW_RETRAIN_CMD | (unset) | Trainer command for a mode: "weights" re-train, with {corpus} / {out} placeholders. Unset ⇒ the corpus is exported and the job returns prepared. |
GW_RETRAIN_CWD | (unset) | Working directory for the trainer subprocess. |
GW_RETRAIN_DIR | <db dir>/retrain | Where re-train corpora, logs and outputs are written. |
GW_RETRAIN_AUTOPROMOTE | 1 | On a successful weights job, point the live detector at the new checkpoint and restart. Set 0 for regulated deployments where a human signs off on every model change. |
Cloud upstreams & secrets¶
| Variable | Default | Description |
|---|---|---|
GEODESIA_SECRET_PROVIDER | "file" | Selects how secret://app/name upstream credential references are resolved: file (local JSON secrets file, air-gap friendly) or aws_secrets (AWS Secrets Manager, with local-file fallback). (src/glad_minimal/secrets_provider.py) |
GEODESIA_SECRET__<APP>__<NAME> | — | Per-secret env override, checked first regardless of provider. <APP> / <NAME> are the secret://<app>/<name> reference segments, upper-cased with non-alphanumerics replaced by _ (e.g. secret://app_invoices_eu/upstream → GEODESIA_SECRET__APP_INVOICES_EU__UPSTREAM). |
AWS_REGION | — | Region for AWS Bedrock upstreams and AWS Secrets Manager. |
GOOGLE_APPLICATION_CREDENTIALS | — | Path to the service-account JSON for Google Vertex (Application Default Credentials). |
Credentials are never stored in plaintext
An Application's binding.api_key_ref holds a secret://app/<name> reference, never the secret. The provider resolves it at request time (env override → file → AWS Secrets Manager). Bedrock and Vertex use the host's IAM role / ADC and need no key at all. See Cloud Upstreams.
Free-tier entitlement overrides¶
The free tier (no signed license) defaults to 1 Application / 1 model / 20 chats per day. These dev/test overrides relax those caps locally; unset leaves the product defaults unchanged. Each accepts an integer, or 0 / unlimited / none for unlimited. (src/glad_minimal/entitlements.py)
| Variable | Default (free tier) | Description |
|---|---|---|
GLAD_FREE_MAX_APPLICATIONS | 1 | Override the free-tier maximum number of Applications. |
GLAD_FREE_MAX_MODELS | 1 | Override the free-tier maximum number of distinct upstream models. |
GLAD_FREE_MAX_CHATS_PER_DAY | 20 | Override the free-tier daily chat cap (UTC-day bucketed). |
A signed license supersedes the free tier entirely and declares its own max_applications / max_models / max_chats_per_day (any of which may be null for unlimited); these env overrides apply only when no valid license is installed. See Licensing.
License token
GLAD_LICENSE_TOKEN (documented under Security) carries the entitlement that sets max_applications and unlocks additional Applications beyond the free-tier cap.
Docker Compose Variables¶
When using the provided Docker Compose setup, all variables above can be placed in a .env file alongside docker-compose.yml:
# .env
GLAD_LICENSE_TOKEN=glc_your_token_here
MODEL_HOST_PATH=/models/geodesia-g1
GLAD_DEVICE=cuda:0
GW_UPSTREAM_URL=http://vllm:8000
GW_UPSTREAM_TYPE=vllm
GW_UPSTREAM_MODEL=mistralai/Mistral-7B-v0.3
GW_BLOCK_INPUT=1
GW_BLOCK_OUTPUT=1
Precedence Order¶
When the same setting is available in multiple places, this is the resolution order (highest to lowest precedence):
- Per-request override —
threshold_overridesin the API request body - Environment variable —
GW_THR_*orGLAD_* - Database threshold prefs — set via
POST /v1/glad/threshold-prefs - config.yaml value
- Compiled default